Building Brains to Understand the World's Data

构建大脑以理解世界数据 杰夫·霍金斯（Jeff Hawkins）概述了皮质计算原理——特别是结构可塑性、突触持久性和稀疏分布式表征——如何让大脑以极高的生物容错性和二进制信号效率学习连续的时间模式。他展示了 Numenta 如何将这些原理工程化应用于 Grok 中；Grok 是一款托管在 AWS 上的自主流式计算引擎，无需超参数调优即可在电网、云服务器和风力涡轮机等工业环境中执行实时异常检测与预测建模。霍金斯认为，硅基机器智能的扩展速度将远快于生物大脑，同时他驳斥了人形机器人、意识上传以及失控 AI 叛乱等流行科幻桥段，认为这是对原始生物生存驱动力的误解。在最后与雷·库兹韦尔（Ray Kurzweil）的互动问答中，霍金斯指出，尽管 Grok 成功解决了实际的序列分析问题，但实现完全的通用智能仍需要分层扩展、选择性注意力和感官运动探索。 Numenta 将其皮质学习算法产品化为 Grok，这是一款商业流式分析引擎，可将原始数字、文本和时间戳转换为稀疏分布式表征，无需超参数调优即可自主建模高速机器数据。 (36:40) 生物突触学习可以通过二进制信号传输来实现工程化，采用用于结构性突触棘生长的标量“持久性”指标（0.0 至 1.0），从而摒弃依赖高精度数值突触权重的人工神经网络设计。 (31:33) 通过消除 5 毫秒的生物神经元吞吐量瓶颈，并整合宏观、微观以及地理分布的感官输入，硅基智能的运行速度有望比生物组织快多达一百万倍。 (45:41) 真正的通用机器智能需要融合层次结构、注意力机制和感官运动整合；仅凭对数据的被动观察不足以构建完整的世界模型。 (60:31)

Jeff Hawkins argues that genuine machine intelligence requires replicating the core operating principles of the biological neocortex—principally sparse distributed representations, sequence memory, and continuous online learning—rather than relying on brittle, task-specific rules or abstracted artificial neural networks. Biological Constraints Versus Classical Artificial Intelligence (4:07) Hawkins critiques classical artificial intelligence for adopting a 'no neuroscience' approach stemming from Alan Turing's 1950 behavioral framing. While systems like chess engines, Jeopardy! winners, and autonomous vehicles solve complex dedicated tasks, Hawkins considers them evolutionary dead ends because they are brittle, confined to narrow domains, and incapable of continuous adaptive learning. Conversely, he rejects brute-force whole-brain simulations, such as Europe's Human Brain Project, for cataloging exhaustive cellular details without a theoretical computational model. Instead, Hawkins treats the neocortex—which constitutes roughly 75% of human brain volume and stores all learned models of the world—as an essential set of physical constraints. Because it provides the only proof that general intelligence is physically possible, Hawkins argues that any valid theory must reflect its uniform two-to-three-millimeter architecture, using sensory-motor feedback to continuously learn an internal predictive model of the environment. Sparse Distributed Representations and Columnar Sequence Memory (17:14) In contrast to dense digital computing words where individual bit positions lack inherent meaning, the neocortex encodes information through sparse distributed representations (SDRs). In Numenta's baseline model, an array of 2,000 bits has only 2% (40 bits) active at any time, with each bit carrying learned semantic meaning. This structure grants vital mathematical attributes: overlapping bits indicate genuine conceptual similarity, patterns can be reliably recognized by subsampling as few as five to ten active bits, and set unions allow the network to evaluate multiple simultaneous hypotheses with negligible false-positive rates. Because roughly 90% of cortical memory processes temporal transitions, Hawkins models sequence memory through dendritic computation. Pyramidal neurons utilize distal dendritic coincidence detectors to enter depolarized, predictive states when familiar temporal contexts occur. By organizing neurons into vertical cortical columns—such as 10 to 30 cells per column—identical sensory inputs evoke distinct cellular subsets depending on prior history. Furthermore, Hawkins models learning through 'synaptic permanence' scalars between 0 and 1 that physically form or sever binary connections based on repeated coincidence, rejecting high-precision floating-point weights as biologically unrealistic. Commercial Streaming Analytics Through Continuous Sequence Learning (36:40) Hawkins asserts that conventional big data workflows suffer from storing massive volumes of idle data in databases, relying on manual feature engineering and static models that discard temporal sequence information. To operationalize cortical principles, Numenta developed Grok, a commercial engine built to process continuous streaming data in real time. Sensory-inspired software encoders automatically convert continuous numbers, categories, and timestamps into SDRs, feeding them directly into sequence memory without requiring manual hyperparameter tuning or retraining phases. Hawkins cites operational case studies to validate the approach. For a Paris manufacturing facility, Grok generated rolling 24-hour power demand forecasts; when an unmodeled midweek holiday occurred, the system flagged the anomaly and dynamically shifted to weekend-like consumption patterns. In another deployment on an offshore North Sea windmill, Grok detected anomalous oscillations in gearbox oil temperature prior to mechanical breakdown, distinguishing acceptable probabilistic variations from true, unpredicted operational failures. Engineering Trade-offs, Biological Limits, and General Intelligence (52:55) Addressing audience inquiries, including questions from Ray Kurzweil, Hawkins defends several deliberate engineering simplifications. While living neurons display graded firing rates and scalar synaptic variations, Grok relies on binary activations and binary connections. Hawkins argues that distributed representations tolerate high biological noise, eliminating the need for scalar precision and allowing Grok to execute learning and inference cycles within 10 milliseconds. Similarly, he treats complex subcortical neuromodulatory and emotional systems simply as binary gating switches that signal whether the model should learn or ignore specific transitions. Hawkins dismisses common science fiction tropes such as mind uploading and hostile, sentient robot uprisings, arguing that cortical algorithms lack biological survival instincts, emotional desires, and reproductive imperatives. However, he emphasizes that silicon-based cortical intelligence can operate far faster than the 5-millisecond throughput limit of biological neurons and can process non-human sensory feeds directly. Hawkins acknowledges that Numenta's commercial engine is not artificial general intelligence; its 60,000-neuron model represents roughly one-thousandth of a mouse neocortex and one-millionth of a human neocortex. Fully realizing general intelligence remains contingent on expanding beyond passive sensory time-series prediction to incorporate hierarchical scaling, attention mechanisms, and closed-loop motor exploration.

Jeff Hawkins 提出，构建真正的机器智能不能仅依靠传统规则引擎或脱离生物实际的人工神经网络，而必须立足于大脑新皮质的核心组织原则，通过稀疏分布式表征和序列记忆模型来实现对高维流式数据的实时预测与异常检测。 新皮质约束与传统AI的局限 (4:07) Hawkins 指出，占人脑体积约 75% 的大脑新皮质是现实世界中通用智能存在的唯一物理证明，其解剖与生理特征为机器智能研究确立了根本的现实约束。图灵在 1950 年提出的测试将智能定义为模仿人类外在行为，这一理念促成了国际象棋引擎、IBM Jeopardy! 问答系统以及自动驾驶汽车等“无神经科学”经典 AI 的诞生。然而，此类系统本质上是脆弱的规则执行器，高度局限于专用领域，缺乏跨领域的持续自适应学习能力。 与此同时，发端于 McCulloch 与 Pitts 的人工神经网络虽吸收了“极简神经科学”，但其加权求和并经过激活函数的点神经元模型严重偏离了活体神经元的运行本质；即便是现代深度信念网络，也依旧脱离大脑的实际工作方式。在另一极端，欧洲人脑计划等“极大神经科学”路线试图通过暴力计算仿真还原大脑全部物理细节，却缺乏提炼通用计算原理的全局理论。Hawkins 提倡将新皮质视作记忆系统，以生物结构为物理约束提炼一般原理，并在专用硅芯片中加以工程实现。 稀疏分布式表征与树突预测机制 (17:14) 传统计算机采用 8 至 128 位的稠密编码，单个比特毫无内在语义；大脑新皮质则依托稀疏分布式表征（SDR）。在典型的 2,000 比特神经元空间中，通常仅有约 2%（40 个比特）处于激活状态，每个比特均承载通过学习获得的明确语义。SDR 具备极高的鲁棒性与语义泛化能力：两个模式在相同位置上的重叠比特直接反映了概念层面的相关性；系统通过随机下采样（如仅比对 5 至 10 个比特）即可完成准确识别，即使其余未检测比特存在差异，也只会表现为语义层面的合理泛化而非系统崩溃。 在微观神经机制上，单个锥体神经元拥有数千至数万个突触，其中约 90% 的远端树突突触作为局部的阈值巧合检测器工作。当一段树突上的突触在数毫秒内协同激活时，会触发非线性树突脉冲，使细胞体进入预去极化的“预测状态”，从而在预期输入到来时抢先放电。通过将细胞组织为垂直皮质柱（例如每柱包含 10 个细胞），系统能根据先验时序上下文在激活柱内指派特定细胞，使得同一感觉输入可在不同上下文中形成多达 10 的 40 次方种不同表征，彻底解决了连续语音同音词等高阶时序歧义。 持续在线学习与突触持久度理论 (31:33) 新皮质是一套严格运行在实时流式环境下的在线学习系统，无法像传统机器学习流水线那样暂停运转并对历史数据执行全局统计批处理。真实神经元的突触具有高度随机性与不稳定性，Hawkins 明确反驳了计算模型中要求突触权重具备高精度浮点数的传统教条，认为高精度的数值权重在生物学上无法成立。 Numenta 提出使用介于 0.0 至 1.0 之间的标量“持久度”（Permanence）来建模学习机制。在接收持续数据流的过程中，未重复出现的模式会被系统迅速遗忘，而持续共现的连接其持久度会逐步增长。一旦持久度跨越预设阈值，突触即建立二进制连通（权重为 1），低于阈值则视为未连通（权重为 0）。这种基于结构可塑性的设计具备极强的硬件容错能力，任意丢弃突触、树突、神经元甚至皮质柱，系统功能均能保持平稳，为低功耗硅基神经形态芯片提供了实现路径。 Grok系统与流式工业时序预测 (36:40) 针对物联网时代数以百亿计的设备数据流，Hawkins 认为未来的计算范式不会将所有原始字节存入静态数据库后再调用人工建立的模型，而是通过海量自主更新的模型直接消费数据流。Numenta 据此开发了商业化引擎 Grok，将皮质学习算法工程化。系统利用模拟耳蜗等感觉器官的通用软件编码器，将数值、分类标签、文本和时间戳等输入直接转化为 SDR，随后送入序列记忆系统提取时空统计关联，输出未来状态的连续概率分布并实时捕捉异常。 该技术已应用于多个工业实际场景：在法国巴黎的一家制造工厂，Grok 在未预先学习法国公共假期历史的情况下，于节假日首日及时将未激增的用电量识别为异常，并即刻调整内部序列状态模拟周末模式，精准输出了未来 24 小时逐时电力负荷预测；在云视频转码业务中，系统动态调配计算实例以降低服务器空转成本；在北海海上风力发电机组监测中，Grok 依据齿轮箱油温的异常微幅振荡，在设备指标尚未突破物理告警阈值前，成功预测了随后发生的机械故障与停机检修。 工程妥协、理论局限与未来演进 (60:31) 在与 Ray Kurzweil 等人的研讨中，Hawkins 阐明了系统在生物真实性与工程落地之间的取舍：尽管生物神经元具备连续的发放率标量，但 Grok 采用二值化激活和二值化连接，将单次学习与推理循环压缩至 10 毫秒以内；针对驱动生物进化的下丘脑等下皮质情绪与神经调节递质，系统将其简化为控制“学与不学”的二元门控信号，以剥离非必要的生物开销。对于多维数据流，Grok 则利用自动化进化算法搜寻最优编码与模型组合。 Hawkins 坦承，当前运行约 60,000 个神经元的 Grok 仅相当于小鼠新皮质的千分之一、人脑新皮质的百万分之一，只是用于处理单一感觉序列的局部皮质切片，远非通用智能。真正的通用智能需要进一步集成皮质层级架构、注意力机制以及关键的感觉-运动闭环（Sensory-Motor Integration）。机器必须通过主动探索环境而非被动接收数据流来构建世界模型。Hawkins 预测硅基智能在虚拟环境中的思考速度可达人脑的百万倍，但他同时驳斥了科幻迷思：由于缺乏生物繁衍与生存驱动，智能机器不会演化出反叛意识；而意识上传因记忆与生物湿件的深度绑定，在技术和哲学层面皆不可行。 Hawkins 提出了一套以大脑新皮质为蓝本、以稀疏分布式表征和序列记忆为基石的机器智能架构，并在流式工业时序预测与异常检测中验证了其有效性。尽管现阶段工程实现为满足实时性对生物标量做出了二值化妥协，且在物理规模与感觉-运动闭环机制上距离完整通用智能仍有巨大差距，但该框架展示了一条摒弃暴力参数堆叠与外在行为模仿、回归生物物理原理的机器智能探索路径。

[0:23] ALAN: So [INAUDIBLE] [0:25] actually invited me to dinner with Jeff here a [0:29] couple weeks ago. [0:30] And I think we probably talked for three hours in that. [0:34] And we probably could have talked for 10 hours. [0:36] Because the area that he's interested is one that Larry's [0:40] also interested in, which is machine intelligence. [0:42] And how do we push the next generation? [0:44] And it's obviously really important to our company. [0:47] But he has an entire career where he's basically built [0:51] from graduate school on a mechanism to basically allow [0:56] him to study what is I think one of the most fascinating [0:59] fields that any of us will ever be involved in. [1:02] And it's biologically inspired machine intelligence. [1:06] And for me, it combines two really, [1:09] really interesting things. [1:10] One is obviously computing and how we commute and things like [1:13] that, but more importantly how we think, how humans think. [1:17] And I think what was really exciting to me is the advances [1:20] made in biology to understand what the mind does and how it [1:24] processes things and the different mechanisms it uses. [1:28] And what really came to me was it's an entirely different [1:31] computing model. [1:32] And I think that's what's so exciting about it. [1:35] So I won't go into the long and illustrious history. [1:38] I'll let Jeff do that. [1:40] But I am just so pleased and happy, both that he could come [1:45] and all the interest in the room here. [1:46] So thank you very much. [1:47] JEFF HAWKINS: Thanks, Alan. [1:49] AUDIENCE: [APPLAUSE] [1:51] JEFF HAWKINS: I should be good with this mic. [1:54] Hopefully, you all can see me. [1:55] Well, thanks for coming out here and just keep piling in [1:57] and sit wherever you want, I guess. [1:59] It's a pleasure to be here. [2:01] I'm not going to tell you about my long career. [2:04] I'll skip all that. [2:05] I've had two careers-- one in mobile computing and now [2:08] another one in brains and neuroscience. [2:09] But really, it's the brains and neuroscience that's been [2:11] the one that's been going on for a long time. [2:13] So I'm going to talk about machine intelligence. [2:15] I'm going to talk about what it is. [2:17] I'm going to talk about how I believe we can get there and [2:19] what we're going to do when we get there. [2:22] And some of the parts of the talk will be pretty technical. [2:24] Some will be less so. [2:24] I'm going to start off with little sort of an intro [2:28] history here. [2:29] My approach to this-- hopefully, this is going to [2:31] work now-- let's try this. [2:32] Oh, I had that down. [2:32] That's it. [2:33] My approach to this has been the following. [2:36] It's a very biological approach. [2:37] It's first to discover the operating [2:38] principles of the neocortex. [2:40] And once we understand those principles, we can then build [2:42] machines that work on those principles. [2:45] And the neocortex, just in case you've forgotten, it's [2:47] the big, wrinkly thing on the top of your brain. [2:49] It's about 75% of the volume of your brain. [2:51] It's where all high level thought, [2:53] language, vision occurs. [2:56] Anything you can tell me about the world is [2:57] stored in your neocortex. [2:58] Mine's speaking. [2:59] Yours is listening. [3:00] That's an organ of intelligence. [3:03] Our approach is like the follow-- oops. [3:04] Oh, I'm going backwards here. [3:05] I got to get used to this. [3:06] Our approach is the following. [3:08] We start with very detailed anatomy and physiology. [3:10] I'm more of a neuroscientist than I am a machine learning [3:12] person in some ways. [3:14] I view the brain as a set of constraints. [3:17] This is the one proof case we know that this can be done. [3:19] And we know a lot about the anatomy and physiology of the [3:23] brain and all kinds of stuff. [3:24] And if we're going to have a theory about what intelligence [3:26] is, it has to be consistent with that. [3:27] We don't have to emulate all of that. [3:29] But we at least have to realize that that is a set of [3:31] constraints. [3:32] From that, we develop theoretical principles. [3:34] We can test those empirically back into neuroscience. [3:36] But we can also test them by building the stuff in software [3:40] and experimenting with it. [3:41] This is what we're doing today. [3:43] Ultimately, this is going to go into silicon. [3:45] We've actually got one project starting this year with [3:47] another major computing company to do some silicon of [3:50] our algorithms. [3:51] But today, it's all in software. [3:53] Now, I thought-- [3:54] I added a few slides here to begin in my talk, which I [3:56] don't normally do. [3:57] I wanted to give you a little sort of historical perspective [3:59] because I know there's a lot of people here in the machine [4:01] learning community and some really experts. [4:03] And I just wanted to give you a little bit of my view of [4:05] where we've come from and how we're getting there. [4:07] So I'm going to go back in time. [4:08] We're going to start a history of AI. [4:10] This guy Alan Turing is where I began in AI. [4:14] Starting in the 1930s, he started talking about [4:16] computers as universal machines-- the universal [4:18] Turing machine as we call them now. [4:20] And he was very much interested in computer [4:22] intelligence. [4:23] And he thought we can make computers intelligent. [4:25] He didn't want to get in arguments about it. [4:27] He wrote about this. [4:28] He said, I didn't want to argue with people whether it's [4:29] possible, whether machines can be conscious. [4:31] So he said, I came up the idea of the Turing test, [4:34] specifically. [4:34] So he just said, look, if a machine could pass the Turing [4:37] test, we'll just have to agree it's intelligent. [4:39] And so that's what he did. [4:40] And he wrote that paper in 1950, the one that's shown [4:43] right here. [4:44] And this is sort of the beginning of the AI movement. [4:46] Unfortunately, what this did is this set the idea that the [4:48] goal of AI is to do things that humans do. [4:51] And I don't think that's the goal of machine intelligence. [4:55] We don't have to replicate what humans do. [4:56] We can do things that are super or less. [4:58] It's not that specifically. [5:00] So the field of AI went on for many years. [5:02] It's still going on. [5:04] I call it the no neuroscience approach. [5:05] There's been many projects and techniques developed over the [5:08] years, some quite successful, some less so. [5:11] I've shown some pictures of some of the great achievements [5:14] here-- playing chess, winning at "Jeopardy." I put the [5:16] Google car there. [5:17] I want one-- [5:19] major initiatives. [5:20] My summary of this is the good news about it, you can come up [5:23] with good solutions to problems we care about. [5:26] The bad thing is that they're very task specific. [5:28] They're kind of brittle. [5:29] The Google car doesn't know how to play chess. [5:31] It's not going to do my laundry. [5:32] And it's not going to think about physics. [5:34] And so they're very task specific. [5:36] And they have limited or no learning. [5:39] And what makes intelligence for humans, of course, is we [5:42] learn about everything. [5:43] And I can do all those things. [5:44] And so can you. [5:46] And so I never viewed this as a really great way of getting [5:48] to true machine intelligence. [5:50] Now I'll just talk about the history of artificial neural [5:52] networks-- so the brain approach, if you will. [5:54] I start this back with these two guys, Warren McCulloch and [5:58] Walter Pitts. [5:59] They wrote a paper in 1943 where they talked about, well, [6:04] neurons could be like logic gates. [6:06] And if we put them together, we can do computing with them. [6:09] So that's an interesting idea. [6:11] Well, it's really not the way brains work at all. [6:13] But what they did is they introduced the idea of an [6:16] artificial neuron that can be used in an [6:18] artificial neural network. [6:20] Now, their approach was kind of odd. [6:22] On the left, you see a real neuron. [6:23] In the middle, you see what is a classic artificial neuron, [6:27] which is a sort of sum to weights [6:29] and activation function. [6:30] That's nothing at all like a real neuron. [6:32] And on the right is what McCulloch and Pitts show, [6:34] which is these little, really, really ridiculous neurons were [6:37] doing ANDs and ORs and NOTs. [6:38] But that was beginning of the artificial [6:41] neural network error. [6:43] Over the years there's been many, many different things in [6:45] this category. [6:45] You really can't put them under one category. [6:48] But I call it minimal neuroscience. [6:50] You'll see in a moment. [6:51] So starting in the '80s, we had a real resurgence and [6:55] interest in artificial neural networks, back propagation, [6:57] Boltzmann machines, and so on. [6:59] This has continued on today into the [7:00] field of machine learning. [7:02] And the current hot topic there is Deep Belief Networks. [7:06] And I would say the good news about this is, well, these are [7:09] learning systems. [7:10] That's good. [7:11] They're distributed. [7:11] That's good. [7:13] They're really good classifiers. that's what [7:14] mostly they're used for, so sort of [7:16] classification problems. [7:17] The downside is they're very limited. [7:19] They don't do much more. [7:21] There's a few exceptions to that. [7:22] And they're not brain-like at all. [7:24] And it didn't feel like this is-- [7:26] when these became hot again in the mid-'80s, and I was [7:28] working on this in the mid-'80s, I was very [7:29] disappointed. [7:30] Because people weren't really paying attention to the [7:32] neuroscience. [7:32] They were just saying, hey, we have these neuron-like things. [7:34] And we're doing stuff with them. [7:36] And we know a lot from the brains. [7:37] We should be going back to them. [7:38] And lately, there's been another thrust here, which you [7:41] might call the Whole Brain Simulator. [7:43] This is maximal neuroscience. [7:44] You might have heard about this. [7:45] There's a project in Europe under Henry Markram called The [7:47] Human Brain Project. [7:50] They just got huge amounts of funding. [7:52] Their goal is to simulate an entire brain from the ion [7:55] channels to the spikes to the neurons to everything, up to [7:58] psychology. [7:59] That chart in the middle is the principal researchers on [8:02] this project. [8:03] It tells you how many people are working on this. [8:06] They do these great simulations. [8:07] The problem with this is there's no theory. [8:10] They'll admit this. [8:11] There's no theory. [8:12] They have no idea what this is supposed to do. [8:14] They hook it all up and see what happens. [8:16] And it's not going to lead to machine intelligence. [8:19] We need theory. [8:20] We need principal ways of going about this. [8:22] So that takes me back to my approach, which is essentially [8:25] like, look, you've got to pay attention to the brains. [8:27] We don't have to copy everything that [8:28] goes on in a brain. [8:29] But you better understand what it's doing before you decide [8:31] what else to do. [8:33] And that's the approach I've been taking. [8:34] So I'm going to walk you through the progress we've [8:36] been making recently. [8:37] And I think it's really cool. [8:40] First of all, we have to talk a little [8:41] bit about the neocortex. [8:42] It's a memory system. [8:43] So if you think about it as a computer, forget it. [8:45] It's not. [8:45] It's a memory system. [8:47] When you're born, it's a blank memory system. [8:49] And as you grow up, you fill it up and you learn things. [8:52] Now, it's a memory system that's attached to some of [8:55] your senses. [8:55] And you learn everything through your senses. [8:57] You have from your eyes, the optic nerve is [8:59] about a million fibers. [9:00] Your somatosensory nerve is about a million fibers. [9:03] Your auditory nerve is about 30,000. [9:04] So you got a couple million inputs going into your brain. [9:07] They're fairly high velocity, not in computer terms. [9:09] But they're changed on the order of tens of milliseconds [9:12] or hundreds of milliseconds. [9:13] Your eyes are moving every few hundred milliseconds-- a [9:15] completely new innovation going on. [9:17] And my speech is changed on the orders of tens of [9:19] milliseconds. [9:20] So you've got this fast data stream coming into the brain. [9:22] The brain has to build a model of the world. [9:25] When you're born, you don't know about things like Google. [9:28] You don't know about buildings and chairs and cars and [9:30] computers and operating systems and bananas-- [9:33] none of that stuff. [9:34] You have to learn everything. [9:36] It's an amazing amount of stuff. [9:37] And you have to learn it through this fast stream of [9:39] data coming in. [9:40] And so the neocortex builds a model of the world. [9:43] And from that model, it does three basic things. [9:45] It makes predictions about future events. [9:47] It says, hey, given my current state and what's coming in, [9:49] what's going to happen next? [9:51] It can detect anomalies, which are just predictions that [9:53] don't come true. [9:53] And it can take actions. [9:55] And it turns out actions are actually the same as [9:56] predictions. [9:58] And so when we take actions, we [10:00] back-interact with the world. [10:01] And so actually, it turns out that most of the things that [10:04] you experience in the world, most of the changes on your [10:06] sensory organs, are caused by your own actions. [10:09] So as I move around the room and I turn my eyes and so on, [10:12] I'm controlling the patterns that are changing here. [10:15] And so there's this real big type feedback loop that goes [10:19] on between a sensory and motor action. [10:22] And if I were to sum this all up, and say, what is this [10:25] system doing, this is what it's doing. [10:27] The neocortex learns a sensory motor model of the world. [10:31] It learns, given a sequence of sensory patterns and a [10:33] sequence of actions, what's going to happen next-- what do [10:36] I have-- my expectations of what's going to happen next. [10:38] Can I predict the future? [10:40] And can I predict what I should do to get the future [10:42] that I want? [10:43] And this is basically what it's all about. [10:45] It's building a sensory motor model of the world. [10:47] And we want to know how it does that. [10:49] Well, we know a lot about it. [10:51] We've learned a lot over the years. [10:52] And I'm going to give you the top six principles right now. [10:56] And this is in no particular order. [10:58] And I made these up. [11:00] I mean, I didn't make them up. [11:01] I made the list up. [11:02] But we're just going through them here. [11:05] First, number one-- it's an online learning system. [11:08] So it's a memory system. [11:09] But it's got to work in an online fashion, meaning it's [11:12] got to work in a streaming mode. [11:14] There's no batch processing. [11:15] The brain doesn't get to look at the [11:17] statistics of a database. [11:18] It says it's coming in. [11:19] I got to act on it immediately. [11:20] I have to incorporate it into my model immediately. [11:22] It's an online learning system. [11:24] And this is essential. [11:26] In a world where patterns are always changing and where the [11:28] underlying structure in the world is changing, you have to [11:30] have an online learning system. [11:32] You can't do this in a batch mode. [11:33] Brains don't work in a batch mode. [11:35] They don't know statistics in that sense. [11:37] The second thing is we know that the [11:39] neocortex is a hierarchy. [11:42] It looks like a sheet of cells. [11:43] It's just two or three millimeters thick. [11:45] It's about the size of a dinner napkin for a human. [11:49] And all of it looks very, very similar. [11:51] But we know from anatomy and other reasons that it's [11:54] actually different regions that are connected together [11:57] with these nerve bundles in a hierarchical fashion. [11:59] There's an estimated-- [12:00] anywhere between dozens and maybe a couple of hundred [12:03] regions in the neocortex-- [12:04] a human neocortex, depending how you're counting. [12:07] We know that these regions are in this hierarchy. [12:09] We know the information flows up the hierarchy [12:11] and down the hierarchy. [12:12] We also know that the regions, amazingly, seem to be doing [12:15] the same thing. [12:16] There are slight variations to this. [12:18] But the basic idea is that every region is [12:20] doing the same thing. [12:21] They're all doing the same type of memory function. [12:23] And so it doesn't matter if you say, well, this is a [12:26] visual area of the neocortex. [12:27] It's visual because it's getting input from the optic [12:30] nerve or other visual regions and auditory areas because of [12:34] where it's getting input from. [12:35] And we can rewire the brains. [12:36] And these regions take on new meanings and new roles. [12:39] So we have this hierarchy of memory regions that are all [12:43] doing the same thing, which tells us that we understand [12:44] what one region is doing, and we understand how the [12:47] hierarchy works, we're a long way towards our goal. [12:51] The third thing is what is the kind of [12:53] memory that it's storing. [12:54] 90%-- [12:55] not all of it-- but 90% of what the neocortex is storing [12:58] is sequence memory. [13:00] It's patterns over time. [13:02] This may not be obvious to you. [13:04] Let me walk you through it. [13:06] Almost all the inference in motor [13:07] behavior is sequence memory. [13:09] So you're listening to my speech. [13:10] And hopefully, I'm not talking too quickly. [13:12] But the patterns are coming in over time. [13:15] The order matters. [13:16] You have in your brain stored what words sound like and what [13:19] certain phrases sound like. [13:21] And your brain is matching them up in time. [13:22] If we move the order and pattern of those patterns in a [13:25] different order, it would be garbage-- totally garbage. [13:28] And people tend to think like, well, vision's not [13:30] like that, is it. [13:31] Vision is like that. [13:33] Real vision is you are moving constantly through the world. [13:35] Your eyes are moving every two, three, [13:37] four times a second. [13:39] And every time it's this constant change. [13:41] These are not images that are just randomly [13:43] represented to you. [13:44] Your brain is directing this. [13:45] And it's figuring out what patterns. [13:47] And so when I look to the right, I know what to expect. [13:49] If I look to the left, I should see [13:50] Alan over here again. [13:52] There he is. [13:52] And so I have these expectations about the world. [13:55] So even vision is this-- all inference. [13:56] France You think about language, music, vision, even [13:59] when I touch things. [14:00] It's all temporal patterns. [14:03] The second thing is we generate motor behavior. [14:06] And motor behavior is, of course, another sequence. [14:08] My speech right now is involving dozens of muscles [14:12] being exercised in very precise patterns-- [14:15] extremely precise patterns-- over time to generate these [14:18] speech patterns. [14:19] And that's true for all my behavior. [14:22] And so we're playing these back. [14:23] These are stored patterns. [14:25] They're stored. [14:25] I could repeat them. [14:26] They're stored. [14:27] I can repeat them because I've got these things in my head. [14:30] I know these. [14:30] I've learned how to do these things. [14:31] I could give this talk blind and probably in my sleep. [14:34] Not true-- [14:35] just for you guys. [14:37] So it's all about sequence memory. [14:39] And that's the key. [14:40] And that's an area that has not been explored enough. [14:43] The fourth item here is that when you look in the brain, we [14:46] find the same type of representations. [14:48] It's called sparse distributed representations. [14:50] Everywhere you look, you see lots of neurons. [14:52] It doesn't matter where in the brain. [14:53] Very few are active. [14:54] Most are relatively inactive. [14:56] This is true on the sensor streams, as well. [14:58] We now understand a lot about this. [15:00] And it's a critical component of how the whole system works. [15:03] I'm going to go into detail in this talk. [15:05] The fifth one here is that we used to think that the regions [15:08] of the brain-- that some were motor regions and some were [15:10] center regions. [15:11] Like this is the primary visual area. [15:12] And this is the primary motor area and so on. [15:13] We now know this is not true. [15:15] Every region in the neocortex has cells that are both [15:20] inference or sensory and motor. [15:22] And there's differentiated-- [15:24] the layer 5 cells or the motor cells. [15:25] But everywhere you go, even primary visual cortex has [15:28] cells that project to something that's motor-- the [15:31] muscles that make your eyes move. [15:32] So there is no pure sensory or pure motor. [15:35] In fact, this is another beautiful discovery. [15:37] Essentially, what every part of the cortex is doing is [15:40] sensory motor learning. [15:42] And so we want to figure out how it does that. [15:45] Finally, the last of my six elements here is attention. [15:50] There's different types of attention. [15:51] I'm not going to go into detail here today. [15:52] But essentially, I need to be able to attend the various [15:55] parts my input stream, various parts of what I'm doing, at [15:57] different points in time. [15:58] Now, I claim these are the primary six things that are [16:01] going on in the neocortex. [16:03] And I'll make a further claim. [16:04] I think they're both necessary and sufficient for biological [16:08] and machine intelligence. [16:10] I don't think you're going to get there [16:12] without these elements. [16:13] It doesn't mean there aren't other things we can do. [16:16] But these in my mind are necessary and sufficient. [16:19] They're necessary to learn a sensory motor [16:21] model of the world. [16:23] All mammals have a neocortex. [16:25] And they all have these principles from [16:27] a mouse to a human. [16:28] Dolphins, monkeys, cats, dogs-- they all have these [16:31] principles operating in their brains. [16:33] So you don't have to be human level to need these things. [16:35] But they all have this going on. [16:37] There's things I didn't put in this list. [16:38] For example, I didn't put language. [16:40] You don't need to have language to be smart. [16:43] You need it to be human smart, yes. [16:45] But to build a model of the world, no. [16:47] Dolphins are really smart. [16:48] And they have very limited language. [16:49] I didn't put other things-- episodic memory, dreaming. [16:52] You don't even have to have a body. [16:53] You have to have motor output. [16:55] But that could be all virtual in some cyberspace someplace. [16:58] So there's a lot of things you don't have to have to build [17:01] machine intelligence, but I believe these six elements. [17:03] And if you're going to forget everything else I talk about [17:05] today, just remember there's six things. [17:07] And the most important one is sparse distributed [17:08] representations, which I'm going to talk about next. [17:12] I'm going to talk about three of these, actually. [17:14] I'm talking about sparse distributed representations. [17:15] But I'm going to talk about sequence memory. [17:17] And I'm going to talk about the online learning. [17:19] We know a lot about some of these. [17:22] But there's a lot of things we don't know. [17:23] There's a ton of stuff I still don't know and don't [17:25] understand. [17:25] But we know enough that we're actually starting to build [17:27] this stuff. [17:28] And it works pretty well. [17:30] We're going to jump into sparse distributed [17:31] representations. [17:33] The best way to understand sparse distributed [17:34] representations is to think compare it to the way we do [17:36] things in computers, which I'll call dense [17:39] representations. [17:40] So, what do we do in computers? [17:41] We want to represent something, we take a word-- [17:43] 8 to 128 bits. [17:44] I don't know what we're up to these days. [17:46] We consider all combinations of 1 and 0s, [17:48] from 0000 to 1111. [17:50] An example, of course, is the ASCII code. [17:52] And that's the representation for the letter M. [17:55] Now, we can say things like, well, what do the bits [17:57] mean in this code? [17:58] They don't mean anything. [17:59] If I say, what's the third bit in an ASCII code mean? [18:02] It doesn't mean anything. [18:03] The whole number means something. [18:04] Somebody could say, well, it means an 8 bit. [18:06] That's not what it means. [18:07] It doesn't tell me anything about the letter M. [18:09] And these representations are arbitrary. [18:13] We could've assigned a different representation of [18:14] the letter M. And it would have been just fine, as long [18:16] as we all use the same convention. [18:18] In brains, it works very differently. [18:21] First of all, you always have lots of things. [18:23] When I talk about bits, you can think of it as a neuron. [18:26] And when I say the bit's a 1, the neuron's active. [18:27] And the bit's a 0, the neuron's not active. [18:30] So we have many, many bits-- [18:32] thousands of them. [18:33] And they're mostly inactive. [18:35] So mostly zeroes and a few ones. [18:37] In our work very often and for this talk, I'm going to stick [18:40] to an example where we're using 2,000 bits, of which 2% [18:43] are active. [18:44] So I have 40 1s and 1,960 0s. [18:46] That's how I'm going to represent everything. [18:48] There's many, many ways I can pick from those. [18:50] But that's how I'm going to represent everything. [18:52] Now, there's several key things about sparse [18:54] distributed representations. [18:56] But this one is that each bit has semantic meaning. [18:59] You can actually say what it means if you knew. [19:02] These are learned. [19:03] No one's going to assign it in a brain or [19:05] an intelligent machine. [19:06] It's learned. [19:06] But it's relatively stable. [19:08] And they mean something. [19:10] And so, when we want to pick a representation, it's sort of a [19:15] competitive process. [19:16] We take the top 40 semantic attributes. [19:19] And those are the ones that are going to be in our [19:20] representation. [19:22] If I want to represent the letters of the alphabet using [19:25] sparse distributed representations-- [19:26] and I would not do this-- this is just purely for example-- [19:30] and I wanted to engineer this, I could say, OK, I would have [19:34] bits that represent is this a constant or a vowel? [19:36] What does it sound like? [19:37] Is it an A, E, I, O sound. [19:39] Is it a fricative sound? [19:40] Is a hard sound, a soft sound? [19:42] How do I draw it? [19:42] Do I have ascenders or descenders? [19:43] Is it a closed shape? [19:45] Where is it in the alphabet? [19:46] What's it next to? [19:47] What other meanings does this thing have? [19:48] And I could come up with all these attributes. [19:50] And then I'd pick the top 40 to represent [19:52] any particular letter. [19:54] That's the basic idea. [19:55] Now there's some properties with sparse distributed [19:58] representations. [19:59] I'm going to go through a few of them here that are really, [20:00] really important. [20:02] The first is semantic similarity. [20:04] Basically, if I took two sparse distributed [20:06] representations, and I compare them bit for bit, if they [20:09] share a bit in the same location, then they're sharing [20:11] semantic meaning. [20:12] This is not arbitrary. [20:14] This doesn't happen by chance. [20:15] It's meaningful. [20:16] And even just a few bits of overlap between two [20:18] representations is statistically very [20:20] significant. [20:21] But it's also semantically significant. [20:24] Now, what if I asked you to store one of these patterns? [20:27] I want to remember this. [20:28] And I'm going to ask you, here's a new [20:29] pattern coming in. [20:30] I want you to tell me if you've seen it before. [20:33] So you might say, well, I'll store 2,000 bits. [20:35] And then when the new 2,000 bits comes in, I'll check and [20:36] see if it's the same thing. [20:37] That's not the way we're going to do it. [20:38] We're going to store the indexes of the 1 bits. [20:41] So we're going to say, OK, all you need to remember are where [20:43] the 1 bits are. [20:44] And so I have 40 1 bits in my representation. [20:46] So I'll have 40 indices to the 1 bits. [20:48] And if I see a new representation-- [20:49] I looked in those locations, and if I see 1s, I know I got [20:51] the same representation. [20:52] Because every representation has 40 1s. [20:54] What if I told you couldn't do that? [20:56] I said, you can only sub-sample. [20:57] You can only sample a few of the 1 bits. [21:00] You can't store the locations of all 40. [21:02] So I'll say 10. [21:04] You can only pick 10. [21:04] We'll randomly pick a 10. [21:05] Now you have indices to the 10 of the 1 bits. [21:08] Now I'm seeing a new pattern. [21:09] You say, is this the same pattern? [21:10] And you look and say, yes. [21:11] The 1s are all in the same location-- [21:13] the 10 1s I know about. [21:14] Is it the same pattern or not? [21:16] Well, you say, well, I don't know. [21:18] What about the other 30 bits? [21:19] They could be different. [21:20] Well, it turns out that it's very [21:21] unlikely they'd be different. [21:22] But even if they were different, I'd be making a [21:25] mistake but a mistake with something that's semantically [21:27] very similar to the thing I stored. [21:30] And this is the key to generalization in the brain, [21:32] is that you don't need to store the location of all the [21:36] bits and know what everything is. [21:37] Basically, when you make errors, you're making errors [21:38] of semantic generalization. [21:40] In fact, I could say, you know what, it's good enough to just [21:43] match five or eight of these things. [21:44] And still I'd have a semantic generalization. [21:46] So I have a way of scaling it back, up and down. [21:49] I'm going to tell you the last property here. [21:50] And you'll see why all this works in a moment-- [21:53] how we're going to use this-- is one of union membership. [21:55] Let's say I took 10 sparse distributed representations. [21:58] And I [? ord ?] them together. [22:00] So now you have 2,000 bits. [22:01] But instead of 2% of the bits being active, it's about 20% [22:04] of the bits being active. [22:05] And that's a one-way street. [22:07] I can't undo that. [22:08] I can't say, oh, what were the 10 bit that were in there? [22:10] Can't do it. [22:11] But I can do something almost as good. [22:13] I can show you a new sparse distributed representation and [22:16] ask, is this one of the original 10 by [22:19] looking at the union. [22:20] And I can do that. [22:21] I can say, well, just look for the ones that are in the new [22:24] one and see if they're in the union. [22:25] And if they are, I'm going to claim it's one of [22:28] the original 10. [22:29] Now, you might say, hey, it could make a mistake. [22:31] It could be picking some 1s from the first one and some 1s [22:34] from the second one and so on. [22:35] Statistically, extremely [22:37] astronomically unlikely to happen. [22:39] But even if it did make a mistake, it wouldn't matter. [22:41] Because I'm going to be making a mistake for something that's [22:44] semantically very similar to the thing I stored earlier. [22:47] And that's good enough. [22:48] In fact, it's what we want. [22:49] It's not even good enough. [22:50] It's actually desirable. [22:52] Again, if you want forget everything else and zone out [22:54] the rest of the talk, remember the future of intelligent [22:58] machines is sparse distributed representations. [23:00] I'm telling you, there's no other way around it. [23:02] And I was just talking to John down here earlier about this, [23:07] I believe, is the future of understanding language, as [23:09] well, and text. [23:10] Because this is how the brain does it. [23:12] We're now going to skip to the next thing-- sequence memory. [23:15] This is 90% of what's being stored in your brain. [23:17] The various types of sequence memory, it's not so simple, [23:19] but the various types. [23:20] And we spent years trying to figure out how this works. [23:23] And we think we got it. [23:24] Let me just give you some of the neuroscience. [23:26] Zoom in on any section of the neocortex. [23:27] It doesn't matter where. [23:29] You'll see there's these layers of cells-- typically [23:30] five layers of cells. [23:32] I'm arguing that they all are a type of sequence memory. [23:35] They have similar attributes. [23:36] It doesn't really matter what layer you look at. [23:38] What I'm talking about here would apply to [23:39] any layer of cells. [23:40] They're all different types of sequence memory. [23:42] So if we zoom in on one of those layers, what you'll see [23:45] is you'll see the cells packed in there really tight. [23:47] There's about 10,000 per cubic millimeter. [23:49] But they have two organizations [23:51] which are worth noting. [23:52] One's shown by the green arrow, which is that cells [23:55] that are in a very skinny vertical column have similar [23:59] response properties, especially a feed-forward [24:01] response property. [24:01] They all seem to respond to the same thing in the world in [24:04] a feed-forward basis. [24:06] However, 90% of the connections are horizontal [24:09] connections, across the columns in different [24:11] areas of the brain. [24:12] So 90% of the connections are elsewhere where they have this [24:14] very strong vertical orientation. [24:15] If we zoom in further and look at one of those cells, we see [24:19] that the cells are dominated by this dendritic arbor, which [24:21] is this tree-shape structure around it. [24:23] All the connections-- [24:24] the positive connections-- to the cell are on the dendrites. [24:26] They're not on the cell body. [24:28] And so on a typical neuron, there's anywhere from several [24:30] thousand to a few tens of thousands of connections on [24:33] those dendrites. [24:34] If we zoom in further, and now you're looking at one little [24:37] section of a dendrite in this picture here, you can actually [24:39] see the synapses in this electromicrograph. [24:42] There's those little spines coming off to about 1 micron [24:44] apart ranged along the dendrite there. [24:47] We now know-- and we didn't know this 15 years ago-- [24:50] but we now know that there's a very nonlinear effect that's [24:53] happening here. [24:54] If a number of these synapses become active at the same [24:57] time-- relatively the same time-- with a few [24:58] milliseconds-- [25:00] and a short distance from each other-- within about 40 [25:02] microns of each other-- [25:03] and enough of them happen at that time, then you get a very [25:06] nonlinear event. [25:07] You generated what's called the dendritic spike. [25:09] And it goes to the cell body. [25:10] And it depolarizes the cell body. [25:11] The cell body goes in a hyperactive mode. [25:14] It's ready to fire. [25:15] It's anticipating. [25:16] It's predicting. [25:17] So every little section of the dendritic tree is like a [25:21] coincidence detector. [25:22] It's a threshold of coincidence detector. [25:23] It says, if I see a bunch of inputs at [25:25] the same time, bingo. [25:26] If I see the same number of inputs spread out over time or [25:28] spread over the dendritic arbor, nothing happens. [25:32] This is such an important feature. [25:34] There's hardly anybody who's modeling this today. [25:37] But this is a key to understanding how the whole [25:38] thing works. [25:40] We didn't know about it too many years ago. [25:42] We model all this. [25:43] And here's a picture of one of our simulations. [25:46] On the left, there's a layer of cells with [25:49] four cells per column. [25:50] I'm going to show more detailed pictures of this. [25:52] Our neurons-- our artificial neurons-- capture a fair [25:55] amount of depth of which is going on in the [25:58] brain in real neurons. [25:59] These colored dots represent the synapses. [26:02] And I'm only showing some of them in this picture. [26:04] The green ones are ones that are close to the cell body. [26:06] I'm not going to talk to them further. [26:07] Those are how we form the sparse distributed [26:08] representations. [26:09] The blue ones are on the distal dendrites. [26:11] These are 90% of the connections. [26:12] This is how we're going to learn sequences. [26:14] And we can model these as a cell-- [26:16] each cell is a set of coincidence detectors. [26:18] And when the pattern comes in, if it detects it, it's going [26:20] to make the cell into a predictive state. [26:22] How does this all work? [26:24] How does this learn sequences? [26:26] Let's start with a picture. [26:27] Here's a picture of our sparse distributor representation. [26:29] But now we're showing it as a sheet of cells. [26:32] And the red cells are the ones that are active. [26:34] And the light ones are the ones that are inactive. [26:35] This is just a part of our 2,000 bits. [26:42] This is is a reminder. [26:42] I'm not going to tell you how we formed this. [26:44] But it's through a local inhibitory competition. [26:49] At any point in time, I have some pattern on here. [26:52] And at another point in time, I have a different pattern. [26:56] And so as I'm talking, and as you move around the world, [26:59] this is what's going on in your brain. [27:00] Everywhere you look, you see these [27:01] cells sparsely activated. [27:03] Oops, I went too far there. [27:05] And you've got these patterns that are changing over [27:06] in time like this. [27:07] And we want to learn the sequence here. [27:09] We want to learn how do I learn the sequence of these [27:10] distributed patterns? [27:12] And the answer is that the brain does [27:14] it a cell at a time. [27:15] Each cell learns to predict its own activity. [27:18] So when a cell becomes active, what it does is it says, let [27:21] me look around for guys who were previously active just a [27:23] moment ago. [27:24] Let me see if I can find a bunch of them. [27:26] And I'm going to form connections to them. [27:28] And I'm going to form those connections, as you see on the [27:30] bottom right here, on one of my dendritic segments. [27:32] So if I see that pattern again, I will [27:35] predict my own activity. [27:37] And this is the beginning of sequence memory. [27:40] If we did this, and let's say I showed it a pattern. [27:43] Here's a situation where I've showed it a pattern. [27:45] The red cells are the cells that are getting a [27:48] feed-forward input. [27:48] They're active. [27:49] The yellow ones are hyperpolar or depolarized. [27:52] They are predictive states. [27:53] Now, why are there more yellows than reds? [27:55] What if I train the system-- an A followed by B and then A [27:58] followed by C and A followed by D. And I show it A? [28:01] I'm going to predict B, C and D. This is a union of [28:04] predictions. [28:05] This goes back to our union property earlier. [28:07] I'll have a union of predictions. [28:09] And I can tell if what happens next was one of things I [28:12] predicted or not, even though I'm predicting multiple things [28:15] at the same time, which is really what [28:16] we're always doing. [28:17] Now, this memory I've just showed you, this is a [28:19] transition memory. [28:21] It's a first order transition memory, meaning I can only [28:24] make a prediction based on my current or this [28:27] previous time step. [28:28] What's happening now, I can make a prediction [INAUDIBLE]. [28:30] I can't use a history of time. [28:32] But we need a high order memory. [28:34] The reason we need a high order memory, because that's [28:35] the way the world is structured. [28:36] The high order memory says I'm may need to go back a long way [28:39] to make the correct prediction. [28:40] So, imagine I'm listening to a melody. [28:42] I can't predict the next note by just listening to the [28:44] previous note. [28:44] I may need to hear five notes or six notes or 10 notes. [28:47] The same with speech. [28:49] To predict what I'm going to say or understand my speech, [28:52] you have to understand a long context. [28:54] And the same if you're walking down the hallway. [28:56] This is a high order temporal pattern. [28:59] You know, it's just like, oh, you have door, door, door, [29:00] then the third door on the left type of thing. [29:02] This is the way the world is. [29:03] So we need to make a high order memory. [29:05] And this is a first order memory. [29:07] So the way we're going to do that is we're going to use [29:08] columns of cells. [29:10] And let me explain how we think this works. [29:13] So, imagine I gave you a sparse distributed [29:15] representation. [29:16] But instead of each bit being a cell, I'm going to make each [29:19] bit be a column of cells. [29:21] So in this case, I show 10 cells per bit. [29:24] And I'm going to randomly choose one of the cells to be [29:27] active in that column. [29:28] So if it's a 1 bit, and I pick a column, I pick [29:31] one of those cells. [29:34] I have a much sparser representation now. [29:35] Instead of 2,000 cells, I have 20,000 cells. [29:40] Now, I could pick the same sparse distributed [29:42] representation a moment later and use a [29:45] different set of cells. [29:46] I just randomly pick a different set of [29:48] cells in the columns. [29:49] And so I have a different representation. [29:50] It's the same sparse distributed representation, [29:52] the same columns, but different cells. [29:55] If you just think about this, we have 40 active columns. [29:57] There's 10 cells per column. [29:58] So there's 10 to the 40th different ways that represent [30:01] the same input in different context. [30:04] So here's an example. [30:05] I'm going to say a sentence that has a sound [30:07] repeated four times. [30:09] There are too many tutus to count. [30:12] Now, the sound "to" was used four times in that sentence. [30:15] You didn't get confused. [30:16] If it was a first order memory, [30:18] you would get confused. [30:19] But you didn't. [30:20] And you heard them, even though it was the same sound, [30:23] that you had different meanings at different points. [30:25] So at one point in the brain, you had to have the same [30:27] representation because the same sound's [30:28] coming in on your cochlea. [30:29] At another point in the brain, you had to have a different [30:31] representation because you didn't hear them as the same. [30:34] And that's what's going on here. [30:35] We have these columnar representations. [30:37] And it allows us to create very long, high [30:39] order sequence memories. [30:40] I'm not going to walk you through all the details of how [30:43] this works. [30:43] You'll just have to take my word. [30:44] You can read it on our website if you want. [30:46] But in the end, if you do this, and you use the same [30:50] type of learning rules I was just talking [30:51] about a moment ago-- [30:52] you form a sequence memory of sparse distributed [30:54] representations. [30:55] It's variable order. [30:56] It can be as high order as the statistics allow. [30:59] So it's not like fixed order. [31:01] It's distributed. [31:02] It can do multiple, simultaneous predictions about [31:04] what's going to happen next. [31:05] It's very high capacity. [31:06] This little memory here can learn millions of transitions. [31:09] And it allows for a semantic generalization. [31:11] Imagine if I train this on a series of patterns. [31:14] And now I give it a new series of patterns that are not [31:18] exactly the same representations. [31:19] But they have some bits overlapping with them. [31:22] So they're semantically similar. [31:23] I can apply my sequence memory to what was a previous [31:27] learning to a new input that is semantically similar. [31:30] And that's what the brain does. [31:33] My last thing I'm going to talk about is how [31:35] we do online learning. [31:37] This is the third of my six elements I'm going to talk [31:39] about here. [31:39] And what's this all about? [31:41] Well, basically, it means because we're doing online [31:43] learning, we have to train on every moment in time-- every [31:46] new input to the system. [31:47] And essentially, you don't know if it's noise or [31:49] something valuable, so you have to train on it. [31:51] And basically, if it doesn't repeat, you forget it. [31:54] And if it repeats, you want to remember it. [31:56] So that's pretty much what it is. [31:57] Now, here's a little bit of neuroscience you probably [31:59] don't know. [32:01] Maybe you do. [32:01] But probably, you don't. [32:03] We used to think-- just pay attention, just listen to me [32:05] for a second. [32:06] We used to think that all memory and formation in the [32:09] brain was the strengthening and weakening of synapses. [32:12] And clearly, that happens to some extent. [32:14] But we've now learned that something much more important [32:16] happens in memory, which is we can form new [32:19] synapses very rapidly. [32:20] And we can forget them very rapidly. [32:22] So instead of just strengthening a connection, we [32:24] can form new ones. [32:25] And that's a much bigger pool of potential things you can [32:27] connect to. [32:28] We can do this on the order of just a few tens of a seconds [32:31] that you can form a completely new connection. [32:32] So even one exposure is enough to do this often [32:35] to form a new neuron. [32:36] So these guys on the dendritic trees or those spines, if you [32:39] look on real neurons, some of them are there for very long [32:41] periods of time. [32:42] Some come and go every day as you're learning things. [32:45] So there's much higher information capacity in [32:49] forming new synapses than strengthening old ones. [32:51] In fact, in the real brains, synapses are highly [32:54] stochastic. [32:54] They're very unreliable. [32:55] So if anyone shows you a neural model that has [32:57] precision of two digits or one digit of precision that's [33:00] required in a synaptic weight, forget it. [33:02] It doesn't work like that. [33:04] So the way we do this is we model this growth. [33:07] We say that we have something called the permanence, which [33:10] is a scalar. [33:10] It goes from between 0 and 1, which is essentially modeling [33:13] the growth of a synapse. [33:15] So I can start growing a synapse and not make a [33:16] connection. [33:17] If the permanence gets above a threshold, we now say the [33:21] synapse is connected. [33:22] And we just give it away to one. [33:24] Binary weights-- [33:25] we don't try to be more fancier than that. [33:28] But we have this idea of a permanence. [33:29] And once I'm over a threshold, if I keep reinforcing this, my [33:34] permanence can go up. [33:36] The connection doesn't get stronger. [33:37] But the permanence goes up. [33:38] And it makes it harder to forget. [33:40] So you want to do that because if things repeat over many [33:42] times or you want to make it so it's very hard to forget. [33:44] So that's how we do that. [33:45] If you put this all together, and you say, hey, I want to [33:48] simulate one of these things, which is what [33:49] we do all day long. [33:50] You have 2,000 columns. [33:52] We do 30 cells per column, 128 segments per cell, 40 [33:56] connections per segment. [33:57] These are all very realistic numbers in neuroscience. [33:59] You put this all together. [34:00] You basically have about 300 million in this little model-- [34:03] about 300 million synapses. [34:04] Each one has a connection index and a connection [34:06] permanence. [34:07] The connections are very, very sparse. [34:08] There's a lot of tricks we can do to make this run fast. [34:11] And noticeably, there's no single points of failure here. [34:14] You can drop out synapses, drop out dendrites, drop out [34:16] cells, drop out columns. [34:17] The system keeps behaving very nicely. [34:19] This has a lot of appeal for hardware guys, who-- [34:22] as we talked about building this stuff in silicon. [34:26] I'm now going to switch gears. [34:27] I'm going to talk about-- and I think it's very important to [34:29] build this stuff and make it work and prove it and make [34:31] commercial value out of it. [34:33] So we're going to do that. [34:34] We have been doing that. [34:36] We're applying this in the space of data. [34:38] I'm going to give you more on my take on data. [34:39] This is the data company. [34:40] I'm nervous as hell talking to anybody about data here. [34:43] Because you guys are the data. [34:44] You own the data. [34:46] You be the data. [34:48] But today, this is my view of the world-- my simplistic view [34:50] of the world. [34:51] Today, we're getting huge numbers of sources of data. [34:53] It's growing exponentially. [34:54] We stick them in data bases. [34:57] The vast majority of the data in the world is [34:59] never looked at ever. [35:00] It just sits there. [35:02] We have two ways of getting value out of it. [35:04] One is the visualization tools. [35:06] And the other is creating models. [35:07] And then if we [INAUDIBLE] those models, [35:09] we can act on them. [35:10] There are challenges here. [35:11] One of the biggest challenges is that this whole system is [35:14] not very automated. [35:16] And it takes data scientists-- people like you-- [35:18] to do this stuff. [35:19] And we want to get to a world where there's not just [35:22] hundreds or thousands or millions of models. [35:24] We want to get to one where there's billions of models-- [35:26] the internet of things-- [35:27] everything in the world's going to be creating data. [35:28] And we need to be able to model all this stuff. [35:31] The other problem is-- [35:32] and so today, it takes lots of people. [35:34] It's not automated. [35:34] The other problem is the models can get obsolete. [35:36] If you're not doing online learning, and most techniques [35:38] today are not online learning, you have to rebuild your [35:40] models all the time because the patterns [35:42] in the world change. [35:43] And people just aren't really looking at [35:46] temporal data very much. [35:47] Much of the patterns in high velocity data, especially high [35:51] velocity data, is temporal patterns. [35:53] And that's almost very rarely do people take [35:56] advantage of that. [35:57] They actually try to get rid of it. [36:00] So my view of the world tomorrow, it's not like the [36:02] current world's going to go away. [36:03] But this is where I think the growth is going to be, is [36:05] we're going to go to a world where there's literally-- [36:07] I'm not joking-- billions of machine [36:10] learning models out there. [36:11] The data's going to stream right into the models. [36:13] There's no storage required. [36:14] You're not to save this stuff. [36:15] The models are going to build and continually update [36:17] themselves. [36:18] And you're going to [36:18] immediately take it to actions. [36:20] And so if you look at it, it looks just like what brains [36:23] do-- go over back to what I said earlier. [36:24] I said, whoa, look at that. [36:25] So let's try to apply our techniques to this. [36:28] And so that's what we've been doing. [36:30] The key criteria here is you need to have automated model [36:33] creation for billions of models. [36:34] You need to have continuous learning. [36:36] And you need to be able to find the temporal, as well as [36:38] the spatial patterns in the data. [36:40] So we built a product called Grok. [36:43] It's an engine for acting on data streams. [36:46] It's essentially a productized version of the thing I was [36:48] just telling you about. [36:49] On the left, you can see we have streams-- the records of [36:52] data coming in through time. [36:53] It can be one or more fields. [36:55] We run those fields through encoders, which are just like [36:57] your sensory organs. [36:58] Literally, they're modeled after sensory organs-- [37:00] modeled after cochlea. [37:02] And we can turn them into sparse distributed [37:04] representations. [37:05] Those encoders can be fairly generic. [37:07] We do not have to do new ones to do every [37:08] different problem we solve. [37:10] We use a set of generic encoders. [37:12] We then get sparse distributed representations. [37:15] And by the way, we can put in field-- any kind of numbers, [37:18] categories, text, dates, times. [37:20] You can do custom things, as well, if you want to-- [37:22] semi-structured data. [37:23] We run this through the sequence memory I was just [37:25] telling you about. [37:26] And it basically looks for the spatial and the temporal [37:29] statistics of that data. [37:30] It can make predictions. [37:30] It can detect anomalies. [37:32] And then from that, we take actions. [37:34] What the used has to do in this case, they have to define [37:36] the problem. [37:37] This is actually tricky. [37:37] They have to do a good job at this. [37:39] Then they have to stream the data. [37:40] Grok creates the models. [37:42] It learns the spatial temporal patterns in the data. [37:44] It outputs predictions with percentages. [37:47] Or you know, it's a probability distribution. [37:49] And it can detect anomalies. [37:50] And we're finding lots of applications for this. [37:52] We're finding in energy and I'll tell you some more-- [37:55] product forecasting, anomaly detection, server loads. [37:56] I'm just going to walk you through a few simple examples. [37:59] And then I'm going to speculate a [38:00] little bit for you. [38:01] Today, this is all running on an Amazon cloud. [38:03] It doesn't have to be. [38:04] But that's just the way we implemented it to begin with. [38:06] There's our simple REST API to use it and some web apps to [38:08] help you get started, although it's only in a [38:10] private mode still. [38:11] We're still doing this with customers, hand holding them. [38:15] We see a lot of applications [INAUDIBLE] any space. [38:18] I'm going to show you this one because it's [38:19] very simple to see. [38:20] And that's the main reason I'm going to start with this. [38:22] It's very simple. [38:23] You may not be realizing this. [38:24] There's a thing called demand response, which is large [38:28] consumers of electricity actually bid on price of power [38:32] throughout the day. [38:33] And the utilities will say, if you can use x amount of power [38:35] at 3:00, I'll give you this price. [38:37] If you use y amount of power, I'll give you this price. [38:38] And they're trying to figure out how to do all this. [38:41] And if you could predict both your demand or the supply, you [38:44] can save energy and you can save money. [38:46] So there's a lot to be saved here. [38:48] Here's a factory in France-- [38:51] in Paris, actually. [38:52] And this is a very simple one. [38:54] It's just showing the electrical usage [38:55] throughout a week. [38:56] You can see the five days. [38:57] They're not working on the weekend, apparently. [38:59] Because it's kind of low there at the last two days. [39:02] And the problem that the customer wanted to do is they [39:03] said, at midnight, they have to sort of make their bids. [39:07] And at midnight, they want to make predictions every hour [39:10] for the next 24 hours. [39:10] We had to put a little wrap around this learning algorithm [39:13] to get this to work. [39:14] But we can do that. [39:16] And what we do is you get something like this. [39:18] Now here's the actual and predicted. [39:20] Now, this looks great. [39:21] Pay [INAUDIBLE] attention. [39:22] You really can't tell if it's good or not because you have [39:24] to really look at statistics and [39:25] look at the data carefully. [39:26] But it turns out this was very good. [39:28] The customer was happy with this. [39:30] The red is predicted. [39:31] The blue is actual. [39:31] We can follow this pretty well. [39:33] Here's a situation where the system was just trained on a [39:37] few months of data. [39:38] And here we are. [39:39] On Wednesday and Thursday morning, it starts picking up. [39:41] And it says, hey, [INAUDIBLE] pick up. [39:42] And it didn't happen. [39:43] And the reason was because it was a holiday. [39:44] And we didn't know about this holiday. [39:46] And the system was never trained on holidays. [39:47] So it says, oop. [39:48] Then it says, that wasn't right and starts to say this [39:51] looks like a weekend. [39:51] So it started acting like a weekend. [39:54] Here's another example. [39:55] Same idea, but a little more complex. [39:57] You get the sense that it's not so obvious all the time. [39:59] This is a company that does video encoding. [40:03] They have a service level [40:04] agreement with their customers. [40:05] They have to guarantee a quick turnaround. [40:07] So they have to leave extra servers running on the cloud [40:10] all the time in case the peak in demand. [40:12] So they're always try to manage how many servers do I [40:14] leave running, which are wasting energy and electricity [40:18] and power and money. [40:19] But I have to meet my service level agreement. [40:21] If they could predict customer demand better, then they could [40:24] leave fewer of these extra servers running around. [40:26] So here you can see the data. [40:27] It's quite spiky. [40:29] There's no obvious patterns in it. [40:30] We can't predict all the spikes. [40:32] It's impossible. [40:32] But the question is, can we discern some patterns in this [40:35] data-- do better than any other technique they've had? [40:37] And can we do this in an automated way? [40:39] And the answer is yes to both those. [40:41] We can do it better than they can do it. [40:43] And we can do it in an automated way. [40:44] So in a sense, basically, we just feed the data in and grok [40:46] this thing. [40:47] I won't go through all the details of it. [40:48] But it was a successful application. [40:50] I want to give you a little sense of what it's like if you [40:52] looked inside of this system. [40:53] So just pay attention to the right side of these images. [40:57] Here we're looking down on the-- [40:59] actually 2,048-- [41:00] but 2,000 columns. [41:02] And the green dot means that this was predicted. [41:05] And it actually occurred. [41:06] So we're sort of probing inside of this cortical [41:09] learning algorithm right now. [41:11] And this is a case where what occurred was [41:13] exactly what was predicted. [41:14] This occurs quite a lot. [41:15] It's not unusual. [41:16] Here's a situation where we predicted multiple things. [41:19] But one of the ones that we predicted occurred. [41:21] So we have more of these little blue circles of things [41:24] that were predicted but didn't occur. [41:26] But that's not a mistake, as long as the things that did [41:28] occur were predicted. [41:29] Those are the green dots. [41:30] So I have 40 green dots and maybe 80 blue dots-- [41:32] something like that. [41:33] Here's a situation where things didn't work too well. [41:35] And I'm sorry you probably can't see this too well in the [41:36] back of the room. [41:38] There's a bunch of red circles in here, as well. [41:41] And those red circles are things that weren't predicted [41:43] but did occur. [41:44] Those are true anomalies. [41:45] I didn't expect this to happen. [41:47] And it did happen. [41:48] But this is so just so typically what happens here. [41:50] It's not an all or nothing affair. [41:52] An anomaly is not one thing. [41:54] It's like, well, some of the things I predicted did occur. [41:56] Some of the things I predicted didn't occur. [41:58] And some of the things that occurred weren't predicted. [42:00] So if I wanted to, I could go in here and say, well, [42:03] semantically, what was wrong, and what was right? [42:07] We don't do that per se. [42:08] But that's how the system works on its own. [42:11] So there's a lot of subtlety here. [42:12] Here's a case where we used it with a offshore windmill. [42:18] And this is looking at the oil temperature and the gear case [42:21] and a large offshore windmill in the North Sea. [42:23] The blue line is the temperature. [42:25] And it's going up and down throughout the day as the [42:26] windmill's speeding up and slowing down. [42:28] And the question is, could we detect anomalous behavior? [42:31] And the red line at the bottom is the anomaly score, which [42:36] Grok is putting out. [42:37] And here you can see an interesting event. [42:39] We had two peaks here. [42:41] The earlier stuff is when the system was first [42:43] started being trained. [42:44] But here we have two peaks. [42:45] And the important thing is the first peak is when the system [42:47] start acting a little bit unusual. [42:48] It wasn't out of range. [42:50] It wasn't out of spec in terms of it was too high or too [42:53] low-- the temperature. [42:53] But it was sort of oscillating in a way [42:55] that hadn't seen before. [42:56] So Grok says, that's unusual. [42:58] I'm not able to predict that as well. [43:00] And so you have a peak. [43:01] And then the second peak is when the system actually went [43:03] down for failure. [43:04] And they worked on it. [43:05] We think there's a ton of applications [43:07] here detecting anomalies. [43:09] Here's my dangerous slide. [43:11] My dangerous slide says, hey, if I was Google, [43:13] how would I use this? [43:14] I'd love to have you guys as a customer. [43:15] That's not why I'm here. [43:16] But I'd love to have you as a customer. [43:17] I said, well, we've had a lot of interesting advertising [43:19] space-- online advertising. [43:21] And we've shown that we can do for people who basically have [43:26] real estate to sell, they're always trying to pick what ad [43:28] network to pick. [43:29] And so we've shown that we can predict the expected return on [43:35] a particular ad network, which changes throughout [43:37] the day all the time. [43:38] You want to do this almost on a 15-minute basis. [43:40] We can do that on a per network, per app, per [43:42] demographic basis. [43:43] And essentially, the user can decide, how do I prioritize [43:46] where I serve my ads from. [43:48] We don't really care what the patterns mean. [43:49] It just [INAUDIBLE] patterns there. [43:50] And they change. [43:51] And we find them. [43:52] We can't get it right all the time. [43:54] Again, we're just trying to do it better than [43:55] they're doing today. [43:56] We have a lot of interest in the finance world. [43:59] You can't predict stock prices. [44:00] Forget it. [44:00] It's not going to happen. [44:02] But you can, and we've shown, that we can successfully [44:05] predict volume and volatility better than the industry [44:08] standards today. [44:09] And this is valuable for various reasons. [44:11] We also believe-- we haven't done this yet-- but we have a [44:13] lot of interest in detecting anomalous trading. [44:15] Both this is internal to companies because they're [44:18] trying to figure rogue traders. [44:20] We also think we can do it across a huge number of [44:24] obtuse, weird trading combinations. [44:26] So when something becomes all the sudden more predictable, [44:28] it represents a trading opportunity, which for some [44:32] reason, not many people in trading are saying, bingo. [44:34] You could do that. [44:35] It's not my favorite application. [44:36] But I think there's going to be a lot of [44:37] applications there. [44:38] I mention this because Google-- because you guys, I [44:41] would have to admit, you're not first in everything. [44:43] Yahoo's still ahead of you in finance, I think. [44:45] And you could add some really cool things to your financial [44:47] stuff if you had these kind of capabilities. [44:49] You guys do a lot of computers. [44:51] You have the biggest server farms in the world. [44:53] We found some really interesting applications in [44:55] managing computer resources like I mentioned with the [44:57] earlier example-- [44:59] predicting demand-- [45:00] resource balancing. [45:02] I've even heard a crazy idea-- a great idea-- where someone [45:04] says, you know, different servers have different [45:06] efficiencies running different applications. [45:08] And if you can predict the efficiency, you could switch [45:11] down some servers and bring up others and save energy. [45:13] I think that's really cool. [45:16] A lot of work in the energy space for us. [45:18] We're getting hit on this all the time. [45:20] You guys are involved in smart grids, solar, wind, demand [45:23] response. this could be something great for you guys. [45:25] And finally, I love your cars. [45:27] And maybe something [INAUDIBLE]. [45:29] If you'd predict where parking spots are going to be or [45:31] routing people-- things like that. [45:34] If you want more details on Grok or you want more details [45:36] on these algorithms, there's a white paper on our website. [45:39] And you can read about this stuff on our website. [45:41] I'm now going to switch to my last part of this [45:42] presentation, which I'm going to just speculate a little bit [45:46] about the future of machine intelligence. [45:48] Because I'm really passionate about this stuff, as you [45:50] probably can tell. [45:51] So, what's the future? [45:52] Is it like this? [45:53] Skynet, "The Matrix," "The Terminator"-- ah! [45:57] I'm not a big science fiction fan. [45:58] But I'm told these are bad things. [46:02] Or is the future something nice, like this, like you [46:05] know, little robot butlers like C-3PO? [46:06] Or we're going to play games of [INAUDIBLE]. [46:08] Or maybe we'll come up with new ways of [46:09] entertaining ourselves. [46:11] Is that the future? [46:13] Or is the future ambiguous? [46:15] Maybe it's good. but it turns bad. [46:18] So I don't know. [46:19] But I'm going to tell you some things. [46:21] I have some prognostications here. [46:22] I'll just tell you where I think this is going. [46:24] Here's some things I think is definitely going to happen. [46:26] And I'm not talking 100 years from now. [46:28] This is going to happen. [46:29] And the reason I'm here is because I'm [46:30] trying to happen sooner. [46:33] We can make machine intelligence that's faster and [46:35] bigger than biological intelligence. [46:38] So we can definitely make machine intelligence a million [46:40] times faster than biological brains. [46:42] Neurons can't do anything less than five milliseconds. [46:44] That's their maximum throughput. [46:47] We can do a lot better than that. [46:48] Now, I can't just speed up a brain if don't speed up its [46:51] sensory organs and I don't speed up its data [46:52] streams and so on. [46:53] But I don't see any reason why we can't do that. [46:55] In virtual worlds, we should be able to make machines that [46:56] are a million times faster and think a million times faster [46:59] than humans think. [47:00] We can make them bigger. [47:02] That's not the only goal. [47:03] But there's no reason at all we can't make bigger [47:06] neocortexes. [47:08] And you can't make it smarter just by making it bigger. [47:12] That's a mistake. [47:13] You can't just say, make it bigger, and it'll be smarter. [47:15] You still have to learn. [47:16] You still have to be exposed to things. [47:17] It takes 20 years to train a human. [47:20] We have to come up with-- these are [47:21] our training systems. [47:22] And they have to be exposed to environments. [47:24] But there's no question in my mind that we can make deeper [47:27] thinking machines than humans. [47:30] This is the area I get very excited about. [47:32] We can do super senses. [47:33] We should not be thinking about the senses of machine [47:35] intelligent machines as hearing and vision and touch. [47:39] I was just showing-- we have sensors that are looking at [47:41] oil temperatures. [47:42] We can have sensors that look at anything. [47:43] We could have distributed sensors. [47:45] We could have microscopic sensors that [47:46] work inside of cells. [47:48] All kinds of things where humans, we have an impedance [47:50] mismatch because of our own senses. [47:53] We spend a tremendous amount of time trying to come up with [47:55] ways of looking and thinking and experiencing stuff that we [47:57] can't normally experience. [47:59] But we could build artificial brains that [48:01] experience it naturally. [48:03] We can do fluid robotics. [48:05] We're nowhere close to this today. [48:06] But I think we can get there. [48:08] And finally, this is another idea. [48:10] The neocortex is a hierarchy. [48:12] And in the brain, they're all co-located because we have to [48:14] run these wires between them and the neurons between them. [48:17] But in a machine intelligent world, we [48:20] don't have to do that. [48:20] We can have parts of the hierarchy all over the place. [48:22] And we can have a distributed hierarchy. [48:24] We can have hierarchies on top of hierarchies. [48:26] I don't even know where that's going to go yet. [48:27] But the idea that it doesn't have to be co-located. [48:30] And as long as we get the communications right, that [48:32] would be very interesting stuff. [48:33] This is all going to happen. [48:35] Here's some things that might happen. [48:36] I don't know. [48:37] Maybe they will. [48:37] Maybe they won't. [48:39] Humanoid robots. [48:41] Maybe. [48:41] Maybe not. [48:42] Will we have something like C-3PO? [48:43] I think it's probably technical possible. [48:46] It's going to be very, very difficult because if you want [48:48] it to be human like, they have to all kinds of other stuff [48:51] that makes them human-like. [48:52] They have to have the rest of the brain. [48:53] And they have to have all these emotional [48:54] things and so on. [48:56] And I'm not sure that's really where the [48:58] business is going to be. [49:00] You know, I know a lot of people want to do this. [49:01] But to me, it's like this is sort of a sideshow. [49:03] And when new technologies come along, we always [49:05] imagine these things. [49:06] When the steam engine came along, they imagined steam [49:08] engine robots, right? [49:09] That's where the term robot came from. [49:11] But it didn't happen. [49:11] So I don't know. [49:12] It may happen. [49:13] Will we have computer brain interfaces for [49:16] all, like "The Matrix"? [49:17] You plug it in the back and go, whoa, you know. [49:19] I don't know. [49:20] Maybe. [49:20] Maybe not. [49:21] There's a lot of technical problems there. [49:22] I'm not sure we really want to have that. [49:24] Who knows? [49:26] Here's some things I don't think are going to happen. [49:29] I don't think you're going to upload your brain to anything. [49:32] Sorry to say. [49:34] There's two reasons for this, really. [49:36] One is forget about the incredible [49:39] difficult technical problems. [49:41] Just forget about that. [49:43] But the memory in your brain is intimately tied to the [49:46] wetware of your brain and the wetware of your body. [49:49] And you would have to recreate the entire thing in some sense [49:52] to get those connections to be meaningful in another form. [49:56] I also think it'd be quite unsatisfactory. [49:59] Imagine if I went up to you right today and said, you [50:00] know, you can upload your brain to this computer. [50:01] Do you want to do it? [50:02] And you say, yeah, sure. [50:03] And you say, yeah, I want to live forever. [50:05] And then we say, oh, yeah we did it. [50:06] [FARTING NOISE]. [50:06] Then the computer comes up. [50:07] Hey, that's great. [50:08] I'm awake. [50:08] And then we say, we're done with you. [50:09] We can get rid of you. [50:10] And you say, whoa. [50:11] Wait a second. [50:11] I'm still here. [50:12] I mean, it's like-- [50:14] AUDIENCE: [LAUGHING] [50:16] JEFF HAWKINS: You're not going to feel so good. [50:18] And in the end, then those two things will diverge. [50:21] You might as well just have kids. [50:22] It's the same thing. [50:24] AUDIENCE: [LAUGHING] [50:28] JEFF HAWKINS: Finally, I don't think we're going [50:30] to have evil robots. [50:31] These things aren't going to one day become sentient and [50:34] say, I don't want to be controlled anymore. [50:36] You know-- you are dead. [50:37] It's not going to happen. [50:39] These are not replicating things. [50:41] These are not emotional things. [50:42] These are not humans. [50:43] They don't want to have sex. [50:44] They're not hungry. [50:46] We're just trying to use the principles by which the brain [50:48] works to build really, really useful things for society. [50:51] And finally, we can be certain that it's not going to be only [50:53] for friendly uses as well. [50:54] People will do bad things with this. [50:56] But that's true of every technology. [50:57] All right, my last slide. [50:59] Why do this? [51:00] Why do we care? [51:00] Why am I so passionate about this? [51:01] Why do I try to get other people to be [51:03] passionate about it? [51:04] Well, there's two reasons. [51:05] The first is to live better. [51:07] There's no question in my mind, just like computers have [51:10] improved our lives tremendously. [51:12] And the products that you guys build have improved our lives [51:14] tremendously. [51:16] And my life has benefited from that. [51:18] I think having intelligent machines is a way of [51:20] improving our lives. [51:20] We can make the world safer. [51:21] We can make it more energy efficient. [51:23] We can make the world-- [51:25] better health-- [51:26] all the things we want to do. [51:28] There's no question at all that this is the ability to [51:31] move that needle significantly. [51:33] But there's another reason, too, which is to learn more. [51:37] If I sit back and say, what's the purpose in life? [51:39] Why should anyone care that I live here and you [51:43] live here and so on? [51:44] In the end, long after many drinks and so on, I come to [51:47] the conclusion that the goal in life is to acquire [51:51] knowledge and to make sure that knowledge is preserved. [51:54] And this is what we do as scientists. [51:55] This is what we do because we're inquisitive species. [51:56] We want to understand how the world works. [51:58] We want to understand the universe. we want to [51:59] understand when did it begin and when did it end. [52:01] I want to know those answers, too. [52:03] And we could use tools to help us do this. [52:06] Imagine we could have physicists that are a million [52:08] times smarter and faster than us and never get tired. [52:12] And they think about this stuff. [52:13] What if we want to explore the universe? [52:16] We're finding human, Earth-like planets only 13 [52:19] light years away. [52:19] Isn't that great? [52:20] That was in the news this morning. [52:21] I think that's wonderful. [52:22] How long would it take to get a human there? [52:24] And will they survive? [52:25] Probably not. [52:26] If we want to explore the universe, I think we have to [52:28] do it with machines that don't breathe oxygen and are not [52:32] sensitive to the things we are sensitive to. [52:34] So, to me, in the end here, it's all about accelerating [52:36] knowledge accretion. [52:38] And I think there's a way of amazingly accelerating that. [52:41] So that's the end of my talk. [52:42] Thank you. [52:44] AUDIENCE: [APPLAUSE] [52:50] JEFF HAWKINS: That was really good. [52:55] ALAN: So with that, we have a few minutes for questions. [52:58] If you have questions, please come to the mic. [52:59] Ray. [53:00] JEFF HAWKINS: Sure. [53:00] Hi, Ray. [53:01] RAY: Hi, Jeff. [53:03] Great presentation. [53:05] On your model, I like it. [53:08] And I agree with the thrust of it. [53:10] I wanted to focus on one aspect of it, which is the use [53:14] of scalars. [53:15] You mentioned scalars in the context of the completion of a [53:19] connection with a view towards its permanence. [53:24] But the properties are basically represented by [53:26] binary values. [53:29] Either there is a fricative. [53:30] Or the loop is closed. or it isn't. [53:32] JEFF HAWKINS: Yes. [53:33] RAY: So in building systems that have at least many of the [53:37] attributes of your model and trying both binary properties [53:42] and then probabilistic properties-- [53:43] there's an 82% chance there's a fricative-- [53:46] and then using [INAUDIBLE] and [INAUDIBLE] [53:48] to combine them appropriately, I've gotten better results [53:51] with the probabilistic-- [53:52] JEFF HAWKINS: Yes. [53:53] RAY: --properties. [53:54] So I was wondering-- [53:55] JEFF HAWKINS: Yeah. so I'll rephrase the question just to [53:57] make sure everyone understands. [53:58] it's a great question. [54:00] And I'll expand it. [54:01] Because we've done two things. [54:03] One is our neuron activations are binary, as well. [54:06] And our synapses are binary. [54:08] The synaptic weights are binary. [54:09] And the question is, can you get better results or you get [54:11] better results using probabilistic or [54:13] scalar values for those? [54:14] And we actually know for certain that in the brain that [54:17] neurons actually have scalar outputs. [54:19] They have firing rates. [54:21] And we know that synapses are also scalars. [54:24] Now, I mentioned earlier also that synapses are very [54:28] stochastic. [54:29] So a large percent of the time, they don't work at all. [54:32] So that's an argument for not relying on scalar properties. [54:36] But the thrust of the answer here is [54:39] we've taken a shortcut. [54:40] We've said that because we have distributed [54:44] representations, you do not need to rely on the accuracy [54:48] of any scalar value anywhere in the system. [54:51] And it's much quicker and simpler to implement this as [54:54] binary activations and binary synaptic weights. [54:59] I'm not saying it's realistic. [55:01] But given the principles-- [55:02] I understand what's going on-- [55:04] I can back off to it. [55:04] And it makes my system run much more reliably. [55:07] We've spent a great deal of time trying to [55:09] make Grok run fast. [55:10] We can do a learning inference cycle in 10 milliseconds. [55:14] And we need to do this. [55:16] Because if you were going to build a practical system, you [55:18] have to make the things before them. [55:19] So that was an engineering choice we made. [55:21] It's not a biological choice. [55:23] And I'm not disagreeing with you. [55:24] It probably would get better results if I [55:25] did it with a scalar. [55:27] RAY: Just a quick question on your view of Markram's project [55:30] because I've had this debate with him over the past summer. [55:33] And he expects simulating at the molecular level, which of [55:37] course is not the right way to build AI. [55:38] But it may be a good way to verify our models of biology. [55:43] But by 2020, he'll be able to simulate it at [55:47] 100 times real time. [55:48] And you'll be able to actually have a conversation with it. [55:51] And I said, how are you going to have a [55:52] conversation with it? [55:54] Because if you're absolutely perfect in your model, it's [55:59] not to do anything, just like human brain doesn't do [56:01] anything unless it's gone through years of learning. [56:04] And if you're at 100 times real time, how are you going [56:06] to have it learn about the world and have a conversation? [56:11] JEFF HAWKINS: Yeah. [56:12] There's a lot of fundamental issues with that project. [56:15] I would agree. [56:16] But look, I'm excited anybody wants to do any of this stuff. [56:20] And maybe we'll learn something from it. [56:22] And so look at the positive side. [56:24] I think they've now taken that project and view it more as a [56:27] way of simulating drug interactions and all different [56:29] kinds of things like that. [56:31] Yes? [56:31] AUDIENCE: So I want to extend Ray's question a little bit, [56:34] which is I appreciate that you've taken the last 15 years [56:37] of advanced neural architectural understandings [56:40] and built that into your model. [56:41] Fantastic. [56:43] In the past 15 years, there's also been an advance in things [56:45] like slope [INAUDIBLE], hormonal neural [56:48] response, and so on. [56:49] And people like Antonio Damasio and Joe LeDoux are [56:52] saying that that's actually fundamental to cognition. [56:55] And yet it doesn't appear to be in your model anyplace. [56:57] So how do you reconcile that? [56:59] JEFF HAWKINS: Can I sum it up in sort of the emotional [57:01] affective content of the brain. [57:02] Is that a reasonable-- [57:03] AUDIENCE: But there's also underlying neural and hormonal [57:06] mechanisms-- [57:06] JEFF HAWKINS: Yeah, sure. [57:07] AUDIENCE: --account for a lot of that. [57:08] JEFF HAWKINS: Yeah. [57:08] So I want to make sure I address the question. [57:13] So there's a lot of stuff in the brain-- [57:15] a lot. [57:16] You've picked one that I didn't talk about-- the [57:17] synchronies. [57:18] There's all kinds of rhythms and so on. [57:21] And so this particular one is that there [57:25] are hormonal aspects. [57:26] There are multiple neural modulators that are [57:29] distributed throughout the neocortex that are fundamental [57:31] for learning. [57:33] But again, when you look at it, and you can say, well, is [57:36] it important for machine intelligence to have those? [57:41] We have ones that are based on fear and based on [57:43] reward and so on. [57:45] We have an emotional system in our system and in our models. [57:49] It's a switch. [57:49] It says, learn or don't learn. [57:51] And it's a very crude emotional system. [57:55] But it's good enough for what we need to do. [57:58] If I were to build a system that's interacting in social [58:02] networks with people and having conversations and [58:05] trying to get food and trying to have sex and trying to stay [58:07] warm, there's a whole bunch of other things you might want to [58:10] have-- other affective things. [58:12] And then, just to be clear. [58:13] In the brain, the parts of the brain that actually evaluate [58:17] emotional content or emotional saliency [58:19] are not in the neocortex. [58:19] There are small areas. [58:21] They're subcortical. [58:22] They project the neural transmitter [58:24] throughout the neocortex. [58:25] So they have a global affect, like learn this. [58:27] Don't forget this. [58:28] You just nearly died. [58:31] This was a bad piece of chicken. [58:32] Don't eat it again. [58:35] But from what we're trying to do, which is figure patterns [58:39] and data and structure data and so on, it's not necessary [58:41] at this point in time. [58:46] It's a fundamental aspect of being a human. [58:47] But it's not a fundamental aspect of intelligence. [58:50] AUDIENCE: All right. [58:51] JEFF HAWKINS: That's how I'd probably put that. [58:55] AUDIENCE: In all of your examples of practical [58:59] applications, you seem to be predicting [59:02] one-dimensional data. [59:03] Did you use only one dimension of input, as well? [59:07] JEFF HAWKINS: Now, so, when you say one dimension, I [59:09] assume you mean like multiple factors. [59:11] AUDIENCE: I mean that it's a scalar value. [59:14] JEFF HAWKINS: Well, the scalars, we handle one or more [59:18] fields of scalars-- [59:20] enumerated types, dates, and times. [59:23] And so did I answer that question yet? [59:29] I didn't ex-- [59:30] what we do is we end up forming a signal-- [59:32] AUDIENCE: Does your model perform significantly worse if [59:35] it doesn't have that [INAUDIBLE]? [59:36] JEFF HAWKINS: Well, we find out. [59:37] So I didn't tell you about how this works with the way the [59:39] actual user would use Grok. [59:41] If they provide multiple data streams, they tell us what [59:44] they want to predict. [59:45] We do an evolutionary search through model space to figure [59:48] out which factors help make better predictions and which [59:52] don't and then how to encode those factors. [59:55] So in the end, you end up with-- [59:57] and sometimes we actually run these as ensembles of models. [59:59] We have multiple models running at the same time. [60:01] And they are competing with each other. [60:06] So the answer to the question is Grok tries to figure out [60:09] what, of all the data you give it, which ones are the best [60:11] and most predictive. [60:12] Or which factors help-- [60:14] how to use them. [60:15] Sometimes some factors help. [60:17] Sometimes it doesn't. [60:18] For example, in that particular case with the [60:20] windmill, if you included wind speed as well as oil [60:24] temperature, it's usually helpful. [60:26] But it does pretty well without it. [60:28] So it depends. [60:30] It really depends. [60:30] AUDIENCE: All right. [60:31] Now, what you've described is a system for predicting data. [60:37] That does not seem to me to be a complete solution to, shall [60:42] we say, general intelligence or an autonomist system. [60:47] What do you see as being necessary to complete the [60:51] system in that sense? [60:52] JEFF HAWKINS: I hope everyone heard that question. [60:54] So there's a lot. [60:56] I only showed three of my six elements. [60:59] We did not have a motor component. [61:01] We don't have attention. [61:02] And we don't have a hierarchy. [61:04] We built something that's very small. [61:05] We built something that's one-millionth the size of a [61:07] human neocortex-- [61:08] one-thousandth the size of a mouse neocortex. [61:10] It is really small. [61:11] It's 60,000 neurons. [61:13] It's teeny. [61:15] I'm not calling that a sentient being. [61:17] This is like a teeny little piece of cortex that's [61:19] learning patterns. [61:21] But the key elements-- [61:23] my argument is that if you get all six of those elements-- [61:27] all six of those things-- including-- so the hierarchy [61:29] is basically a scale-- [61:30] necessary for scale. [61:32] The sensory motor is a huge component of this because you [61:34] have to explore the world. [61:36] And I'm working on that right now. [61:37] We're making good progress on that-- [61:39] and then the [INAUDIBLE] components. [61:41] So you have to add all this stuff together before you can [61:44] start claiming you've got something that's close to [61:45] machine intelligent. [61:46] AUDIENCE: OK, thank you. [61:47] ALAN: Great. [61:48] It's 12:00 right now. [61:49] So thank you all for coming. [61:51] And thank you, Jeff. [61:52] It's been a really, really, great talk. [61:53] JEFF HAWKINS: Thank you. [61:56] [APPLAUSE] [61:57] [MUSIC]