D-Wave Chips Beat GPUs in Generative Model Sampling, Drug Discovery

Introduction to Quantum Annealers in Optimization and Machine Learning With D-Wave — Ai4

Ai4

Physical quantum annealers deliver orders-of-magnitude speedups over classical GPU clusters by natively executing discrete Boltzmann sampling at the latent layer of generative models.

AI-assisted summary of Introduction to Quantum Annealers in Optimization and Machine Learning With D-Wave — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 28, 2025, this presentation features Kevin Chern, Senior Benchmarking Researcher at D-Wave. The talk introduces the principles of quantum annealing on multi-thousand-qubit architectures and highlights its real-world benchmarking benefits across NP-hard combinatorial optimization and generative machine learning.

## TL;DR Chern explains how D-Wave quantum annealers optimize energy landscapes to solve binary combinatorial problems represented as Ising and QUBO models, demonstrating multi-order-of-magnitude speedups over classical baselines and GPU simulations. He shows how quantum annealers natively serve as rapid physical Boltzmann samplers to accelerate training in generative energy-based models. Finally, he presents hybrid architectures like Quantum Variational Autoencoders (QVAEs), highlighting a real-world drug discovery implementation with Japan Tobacco that enhances molecular generation through expressive discrete latent spaces.

## Atmosphere & style The presentation carries a charmingly self-deprecating and approachable tone, effortlessly translating intimidating physics equations into relatable machine learning intuitions. Chern keeps the mood lighthearted with witty asides about his introverted personality and references to Geoffrey Hinton, culminating in a candid audience selfie to commemorate overcoming stage fright. It feels like an engaging, unpretentious masterclass that demystifies cutting-edge quantum hardware for a practical data science audience.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) D-Wave's quantum annealers are purpose-built to solve NP-hard combinatorial optimization, discrete optimization problems, and complex quantum simulations. - [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) Quantum annealing finds global minimum energy states by evolving a baseline Hamiltonian into a target binary optimization landscape, natively matching Ising models in physics and QUBO in computer science. - [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) D-Wave hardware architectures have scaled from minimal multi-qubit prototypes up to production quantum processors with over 4,400 interconnected programmable qubits. - [06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) On benchmarked hard optimization problems, quantum annealers can deliver at least a two-order-of-magnitude speedup compared to classical computing baselines. - [06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) Discrete combinatorial problems such as graph matching and the Traveling Salesperson Problem (TSP) naturally map to quadratic unconstrained binary representations. - [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) Beyond deterministic optimization, quantum annealers act as native physical Boltzmann samplers, accelerating the gradient estimation needed to train generative machine learning models. - [11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) Simulating large quantum systems on classical GPU clusters scales into millions of years, while physical QPUs provide more than an order of magnitude speedup over classical MCMC sampling. - [11:54](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=714s) Quantum Variational Autoencoders (QVAEs) bridge continuous data with quantum hardware by using the quantum annealer to optimize and sample within a discrete binary bottleneck layer. - [13:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=794s) In practical drug discovery applications, Japan Tobacco used QVAEs to optimize molecular generation, achieving improved chemical property distributions and QED drug-likeness scores. - [13:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=794s) While classical planar graph latent spaces allow simple closed-form sampling, the high connectivity of quantum annealer topologies provides the expressivity needed for complex generative distributions.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) Introduction and the Role of D-Wave in Quantum Computing

Kevin Chern opens his presentation with a self-deprecating shout-out to his colleague Emma for designing the title slide—jokingly calling it "the most beautiful slide of the entire deck." D-Wave is a quantum computing company focused on building practical quantum hardware. He points out a slide showing D-Wave's massive cryogenic dilution refrigerator, humorously contrasting the giant outer enclosure ("larger than me") with the actual physical quantum processor chip housed inside ("smaller than me"). D-Wave's hardware tackles NP-hard combinatorial optimization, discrete optimization problems, and quantum simulations—allowing researchers to simulate complex quantum systems even if the target system does not match the physical layout of the processor exactly.

```mermaid flowchart TD accTitle: D-Wave Quantum Computing Workflow subgraph Input ["Problem Formulation"] A["Combinatorial Optimization / Discrete Problems"] B["Quantum Simulations / Generative ML"] end

subgraph Hardware ["D-Wave Quantum Annealer"] C["Cryogenic Dilution Refrigerator"] D["Superconducting Qubit Processor"] end

subgraph Output ["Execution & Results"] E["Ising / QUBO Energy Minimization"] F["Boltzmann Sampling Distributions"] end

A --> C B --> C C --> D D --> E D --> F ```

Chern introduces himself as a Senior Benchmarking Researcher at D-Wave, where he has worked for about four years developing and benchmarking quantum-classical hybrid algorithms. Most of his day-to-day work involves sitting at a desk formulating real-world computational challenges into representations compatible with quantum annealers. After completing his undergraduate and graduate degrees at the University of British Columbia (UBC), he jokes that as a quiet Vancouver native who doesn't go out much, if he starts "stuttering like a little language model," the audience will understand why. Framing the presentation playfully as an "unskippable ad," his real goal is to inspire attendees to explore the powerful intersection of quantum annealers, combinatorial optimization, and generative modeling.

- **Role at D-Wave:** Benchmarking researcher designing hybrid quantum-classical algorithms. - **Academic Background:** Undergrad and graduate studies at the University of British Columbia (UBC). - **Core Hardware Capabilities:** Discrete optimization, quantum simulation, and generative machine learning.

---

[02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) Quantum Annealing Mechanics and Binary Optimization (Ising & QUBO)

A quantum annealer is a physical computing device designed to perform quantum annealing—a process modeled around energy landscape optimization. In a typical optimization landscape with complex, ragged valleys, the objective is to locate the global minimum. Quantum annealing accomplishes this by initializing the system in an easily solvable baseline state (where finding the minimum is trivial) and then slowly transitioning the Hamiltonian toward the target problem landscape. Under suitable conditions, physical and mathematical theorems guarantee that the system will evolve and settle directly into the global optimum state.

```mermaid flowchart LR accTitle: Quantum Annealing Formulation subgraph Variables ["Decision Variables"] X1["x_i in {-1, +1}"] X2["x_j in {-1, +1}"] end

subgraph Coefficients ["Hamiltonian Parameters"] H["Linear Coefficients (h_i)"] J["Quadratic Couplings (J_ij)"] end

subgraph Systems ["Hardware Scale"] Q2["2 Qubits"] Q48["48 Qubits"] Q4400["4,400+ Qubits"] end

X1 --- H X1 --- J X2 --- J H --> Q2 J --> Q2 Q2 --> Q48 --> Q4400 ```

From a mathematical perspective, the problems natively solved by quantum annealers are binary optimization problems: - **Physicists** know this formulation as the **Ising model** (expressed using Pauli $X$ and $Z$ matrices and annealing schedules $A(s)$ and $B(s)$). - **Operations researchers and computer scientists** refer to it as **QUBO** (**Quadratic Unconstrained Binary Optimization**).

In QUBO/Ising models, binary decision variables take values $x_i \in \{-1, +1\}$ (or $\{0, 1\}$). The objective function $H(x)$ consists of linear coefficients (represented visually as biases on individual nodes/vertices) and quadratic coefficients (represented by edges/couplings between pairs of variables $x_i$ and $x_j$).

To demystify high-dimensional physics jargon, Chern references a well-known quote by machine learning pioneer Geoffrey Hinton: *"To deal with hyperplanes in a 14-dimensional space, visualize a 3D space and say 14 to yourself very loudly. Everyone does it."* In the same spirit, Chern tells the audience that whenever they feel overwhelmed by the equations, they can simply picture a discrete binary network, "say quantum spins to yourself very loudly," and remember that the processor is solving an optimization problem. He illustrates this scaling trajectory by stepping from a minimal 2-qubit system up to 48 coupled qubits, and finally to massive production architectures featuring over 4,400 interconnected programmable qubits.

---

[06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) Performance Benchmarks and Combinatorial Problem Mapping

When comparing solution quality and runtime on optimization benchmarks, quantum annealers demonstrate significant advantages over classical solvers on hard problem instances. Plotting time on the x-axis against proximity to the energy minimum on the y-axis (where lower is better), D-Wave's quantum annealer achieves at least a two-order-of-magnitude speedup over classical computing baselines in best-case scenarios.

Many critical real-world and computer vision problems naturally map onto quadratic unconstrained binary models:

| Problem Domain | Formulation Mechanics | QUBO Mapping | |:--- |:--- |:--- | | **Graph Matching** | Finding an optimal node permutation between two graph structures (e.g., blue and orange graphs). | Formulated to minimize structural discrepancy and preserve adjacency structure. | | **Traveling Salesperson Problem (TSP)** | Routing a path through multiple destinations (e.g., 7 cities or errands) to minimize total commute time. | Represented as an optimization over permutation matrices where binary variables indicate step assignments. | | **General Combinatorial Problems** | Binary Integer Linear Programming (BILP) and finite discrete optimization problems. | Transformed into QUBO via penalty terms and automated reformulation frameworks. |

While all finite discrete optimization problems can theoretically be reduced to QUBO instances, Chern cautions that doing so efficiently is the real challenge. Manually translating complex operational constraints into penalty functions on paper can be tedious and labor-intensive, which has spurred active research into automated QUBO-formulation toolchains.

---

[08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) Quantum Annealers as Samplers: Boltzmann Machines and Generative Modeling

Quantum annealing is not merely an optimization engine—it is also a physical sampling process. This property creates a direct bridge to machine learning and generative modeling, particularly through **Boltzmann machines** (a foundational architecture closely associated with Geoffrey Hinton's work).

```mermaid flowchart TD accTitle: Boltzmann Machine Unification Across Fields BM["Boltzmann Machine / Energy-Based Model"]

BM -->|Statisticians| EXP["Exponential Family Distribution"] BM -->|Physicists| ISING["Thermal Ising Model"] BM -->|Computer Scientists| MRF["Binary Markov Random Field"]

EXP --> SAM["Iterative Sampling & Gradient Estimation"] ISING --> SAM MRF --> SAM

SAM --> QA["Hardware Acceleration via D-Wave Quantum Annealing"] ```

A Boltzmann machine can be interpreted across multiple disciplines: - **Statisticians** view it as a high-dimensional member of the **exponential family** of distributions, yielding a probability distribution over discrete binary vectors analogous to a multivariate Gaussian for continuous data. - **Physicists** see it as an **Ising spin system** in thermal equilibrium. - **Computer scientists** classify it as a binary **Markov Random Field (MRF)**.

Training generative models on binary data spaces requires drawing samples from the model distribution at each training step to compute data-model expectation gradients. Because classical sampling techniques (such as Markov Chain Monte Carlo) can suffer from severe computational bottlenecks in multimodal energy landscapes, using physical quantum annealing hardware as a native Boltzmann sampler offers a promising alternative. Chern highlights a landmark paper published in *Science* demonstrating these sampling capabilities and physical speedups on multi-qubit systems.

[11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) Quantum Simulation and Fast Sampling Benchmarks vs. GPUs

Benchmarking data demonstrates the vast efficiency gap between quantum annealing hardware and classical GPU simulation. When plotting the number of qubits on the x-axis against the time required for classical GPUs to simulate the system on the y-axis, the computational wall becomes immediately apparent: the "punch line is takes millions of years on GPUs" to simulate large quantum systems. This simulation workload is fundamentally a sampling problem, where researchers must draw samples from the model and compute statistical observables to extract meaningful physical insights about the target system.

```mermaid flowchart LR accTitle: Quantum Annealer vs GPU Sampling Comparison subgraph Classical["Classical Approach"] GPU["GPU Cluster"] --> MCMC["MCMC Sampling"] MCMC --> HighTime["High Latency / Scaling Wall"] end subgraph Quantum["Quantum Annealing"] QA["D-Wave QPU"] --> DirectSample["Native Sampling"] DirectSample --> Speedup["1+ Orders of Magnitude Speedup"] end ```

A direct comparison using Markov Chain Monte Carlo (MCMC) algorithms displays run time versus sample size. When evaluating the quantum computer against state-of-the-art GPU implementations, the quantum hardware achieves more than a full order of magnitude speedup. The primary technical advantage is that quantum annealers excel at rapid, highly efficient sampling over complex energy landscapes.

- GPU simulation times for large quantum systems scale into millions of years. - Benchmarks against classical MCMC algorithms demonstrate an order of magnitude or greater runtime advantage for the QPU. - High-speed sampling directly accelerates physical simulation and statistical inference tasks.

---

[11:54](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=714s) Generative Machine Learning: Discrete and Quantum Variational Autoencoders

Addressing the question of where binary optimization problems occur in real-world workflows, the answer is "in some sense always" because modern machine learning frameworks can translate continuous or structured data directly into discrete domains. By utilizing Discrete Variational Autoencoders (DVAEs) and Quantum Variational Autoencoders (QVAEs), practitioners can bridge the gap between arbitrary data types and binary optimization hardware.

```mermaid flowchart LR accTitle: Quantum Variational Autoencoder Pipeline subgraph Pipeline["QVAE Architecture"] Input["Input Data"] --> Enc["Encoder Network"] Enc --> Latent["Binary Latent Space"] Latent --> QA["Quantum Annealer<br/>(Optimization / Sampling)"] QA --> Latent Latent --> Dec["Decoder Network"] Dec --> Output["Generated / Optimized Output"] end ```

In a QVAE architecture, an encoder maps complex input data into a binary latent space. The quantum annealer serves as both a high-performance optimization engine and a generative model at this discrete bottleneck layer, optimizing or sampling the latent variables before a classical decoder reconstructs them back into the original data domain. This enables researchers to optimize data generation directly over the binary representation.

- DVAEs and QVAEs map high-dimensional training data into discrete binary latent spaces. - The quantum annealer operates at the bottleneck layer to handle generative sampling and binary optimization. - The learned binary representations are decoded back into the original input space, making quantum annealing applicable to domains beyond native binary graphs.

---

[13:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=794s) Drug Discovery Case Study and Latent Graph Expressivity

A concrete implementation of this architecture comes from Japan Tobacco, one of the largest tobacco companies in Japan, which deployed quantum variational autoencoders for pharmaceutical drug discovery. By embedding a quantum computer directly at the bottleneck layer of a variational autoencoder, they optimized molecular candidates. The resulting distributions showed marked improvements across chemical property metrics: the quantum system produced more favorable generation scores on property histograms and showed clear shifts in drug-likeness metrics such as the Quantitative Estimate of Drug-likeness (QED) score.

```mermaid flowchart TD accTitle: Decision Flow for Latent Graph Modeling Start["Design Binary Latent Space"] --> Planar["Try Classical Planar Graph<br/>(Closed-form / simple sampling)"] Planar --> Check{"Expressive Enough<br/>for Target Task?"} Check -- Yes --> DeployClassical["Deploy Classical Pipeline"] Check -- No --> DeployQuantum["Leverage D-Wave QPU<br/>(High connectivity & rich expressivity)"] ```

When designing binary latent representations, practitioners can initially experiment with classical graph structures like planar graphs, where sampling can sometimes be solved in closed form due to low connectivity. However, planar graphs frequently suffer from limited expressivity. When these classical graph topologies fail to capture complex data distributions, switching to the rich, highly connected hardware topology of a quantum annealer provides the necessary capacity to optimize and generate high-quality candidates.

- Japan Tobacco integrated D-Wave quantum annealers into the bottleneck layer of a VAE for molecular generation. - The quantum-enhanced pipeline generated superior compound candidate distributions and improved QED drug-likeness scores. - Planar graph latent spaces offer low-connectivity baselines with closed-form solutions, but quantum annealers are needed when higher graph connectivity and expressivity are required.

---

[14:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=892s) Audience Wrap-Up and Commemorative Selfie

Concluding the presentation, Chern pauses before taking final questions to capture a personal memento of the session. He admits candidly to the room, "Well, actually before before I let everyone go, I'm going to take a selfie because this is quite scary for me. So, I want to remember this."

Inviting the audience members to either cover their faces or make a goofy expression for the camera, he snaps the stage selfie and thanks the attendees for their time and attention.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) the most beautiful slide of the entire deck. — Kevin Chern

> [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) larger than me — Kevin Chern

> [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) smaller than me — Kevin Chern

> [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) stuttering like a little language model — Kevin Chern

> [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) unskippable ad — Kevin Chern

> [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) To deal with hyperplanes in a 14-dimensional space, visualize a 3D space and say 14 to yourself very loudly. Everyone does it. — Geoffrey Hinton

> [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) say quantum spins to yourself very loudly — Kevin Chern

> [11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) punch line is takes millions of years on GPUs — Kevin Chern

> [11:54](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=714s) in some sense always — Kevin Chern

> [14:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=892s) Well, actually before before I let everyone go, I'm going to take a selfie because this is quite scary for me. So, I want to remember this. — Kevin Chern

## Mentioned Kevin Chern, Emma, D-Wave, University of British Columbia (UBC), Geoffrey Hinton, Science, Japan Tobacco, Ising Model, QUBO (Quadratic Unconstrained Binary Optimization), Boltzmann Machine, Discrete Variational Autoencoders (DVAEs), Quantum Variational Autoencoders (QVAEs), Markov Chain Monte Carlo (MCMC)

## Video description Kevin Chern, Senior Benchmarking Researcher, D-Wave

Abstracts

Quantum annealing is a quantum computing paradigm known for its applications in combinatorial optimization and machine learning. In optimization, quantum annealers provide heuristic solutions to NP-hard combinatorial optimization problems. In machine learning, quantum annealers can be leveraged as generative models in various domains ranging from drug discovery to image generation. In this presentation, I will introduce the principles of quantum annealing with a focus on D-Wave's implementations featuring thousands of programmable qubits. I will also share recent applications and benchmarking results highlighting its practical benefits.

[Source](https://www.youtube.com/watch?v=s9sB2ZPOO04)

中文简报

D-Wave芯片在生成模型采样与药物研发中超越GPU 物理量子退火机通过在生成模型的隐层原生执行离散玻尔兹曼采样，相比传统GPU集群实现了数个数量级的加速。 Chern解释了D-Wave量子退火机如何通过优化能量图谱来求解表示为Ising和QUBO模型的二元组合问题，展现出相比传统基准和GPU模拟数个数量级的速度提升。他说明了量子退火机如何原生充当高速物理玻尔兹曼采样器，以加速基于能量的生成模型的训练。最后，他介绍了量子变分自编码器（QVAE）等混合架构，并重点展示了与日本烟草（Japan Tobacco）合作的真实药物发现应用，该应用通过富有表现力的离散隐空间增强了分子生成能力。 D-Wave 量子退火机专为求解 NP-hard 组合优化、离散优化和量子模拟问题而设计，硬件规模已扩展至超过 4,400 个互连的可编程量子比特。 (2:42) 在复杂优化基准测试中，量子退火机将图匹配和旅行商问题（TSP）等映射为 QUBO 或伊辛模型，相比经典计算基线可实现至少两个数量级的性能加速。 (6:14) 量子退火机可作为原生物理玻尔兹曼采样器加速生成式机器学习模型的梯度估计，在模拟任务中相比经典 GPU 运行的 MCMC 采样实现了一个数量级以上的加速。 (11:07) 量子变分自编码器（QVAE）将连续数据映射至离散二进制瓶颈层进行量子优化与采样，日本烟草公司已成功将其应用于药物分子生成并显著提升了类药性（QED）评分。 (13:14)

Reading summary

D-Wave Senior Benchmarking Researcher Kevin Chern outlines how quantum annealing hardware addresses NP-hard combinatorial optimization, quantum simulation, and generative modeling by functioning as both a physical energy minimizer and a rapid statistical sampler. Quantum Annealing Mechanics and QUBO Formulations (2:42) D-Wave's quantum computing systems house superconducting qubit processors inside large cryogenic dilution refrigerators, scaling from small 2-qubit and 48-qubit systems up to production architectures featuring over 4,400 interconnected programmable qubits. A quantum annealer operates by initializing a physical system in an easily solvable baseline state and slowly transitioning its Hamiltonian toward a complex target energy landscape. Under suitable conditions, physical and mathematical theorems guarantee that the system evolves and settles into the global minimum state. Mathematically, these processors natively solve discrete binary optimization problems where decision variables take values in $\{-1, +1\}$ or $\{0, 1\}$. The objective function combines linear coefficients, which act as biases on individual nodes, with quadratic couplings representing interactions between pairs of variables. Physicists refer to this formulation as the Ising model, while computer scientists and operations researchers classify it as Quadratic Unconstrained Binary Optimization (QUBO). Combinatorial Problem Mapping and Optimization Speedups (6:14) Many real-world operational and computer vision problems map directly onto QUBO structures. For instance, graph matching can be formulated to find an optimal node permutation that minimizes structural discrepancy between two graphs, while the Traveling Salesperson Problem (TSP) can be represented over permutation matrices where binary variables assign specific routing steps. More broadly, Binary Integer Linear Programming (BILP) and any finite discrete optimization problem can theoretically be reduced to QUBO instances using penalty terms. When benchmarking runtime against proximity to the energy minimum on hard optimization instances, Chern reports that D-Wave's quantum annealer achieves at least a two-order-of-magnitude speedup over classical computing baselines in best-case scenarios. However, he cautions that manually translating complex operational constraints into penalty functions on paper is tedious and labor-intensive, which has spurred active research into automated QUBO-formulation toolchains. Accelerating Boltzmann Sampling and Quantum Simulations (8:52) Beyond finding optimal states, quantum annealing is a physical sampling process that connects directly to energy-based generative models, particularly Boltzmann machines. Across disciplines, a Boltzmann machine is interpreted by statisticians as a high-dimensional exponential family distribution over binary vectors, by physicists as a thermal Ising spin system, and by computer scientists as a binary Markov Random Field. Training these models requires drawing samples at each step to compute expectation gradients, where classical Markov Chain Monte Carlo (MCMC) methods often face severe computational bottlenecks in multimodal energy landscapes. Chern highlights research published in *Science* demonstrating that quantum annealers can act as native Boltzmann samplers to accelerate these workloads. In quantum simulation—which relies on drawing samples to compute statistical observables of target physical systems—Chern notes that simulating large qubit systems on classical GPUs scales into millions of years. In direct runtime-versus-sample-size benchmarks against state-of-the-art GPU MCMC implementations, the quantum processor achieves more than an order of magnitude speedup. Quantum Variational Autoencoders in Drug Discovery (11:54) To apply binary optimization hardware to arbitrary continuous or structured data, practitioners use Discrete and Quantum Variational Autoencoders (DVAEs and QVAEs). In a QVAE architecture, an encoder network maps complex input data into a discrete binary latent space. The quantum annealer operates at this bottleneck layer to perform binary optimization and generative sampling before a classical decoder reconstructs the binary variables back into the original data domain. Demonstrating this hybrid pipeline in an industrial setting, Japan Tobacco embedded a D-Wave quantum annealer into the bottleneck layer of a variational autoencoder to optimize molecular candidates for pharmaceutical drug discovery. According to Chern, the quantum-enhanced model produced superior chemical property distributions and improved Quantitative Estimate of Drug-likeness (QED) scores compared to classical baselines. Despite the speedups demonstrated in optimization and generative sampling, deploying quantum annealers effectively requires navigating practical formulation and architectural trade-offs. Alongside the ongoing challenge of automating QUBO translations, Chern advises that machine learning practitioners designing binary latent spaces should initially test low-connectivity classical structures, such as planar graphs, which permit closed-form or simple classical sampling. Quantum annealing hardware is primarily warranted when these simpler classical topologies lack the connectivity and expressivity needed to model complex data distributions.

精炼阅读摘要

D-Wave 基准测试研究员 Kevin Chern 指出，量子退火硬件不仅能高效求解以 QUBO 模型表征的 NP-hard 离散优化问题，还能作为物理玻尔兹曼采样器加速生成式机器学习模型训练。通过将量子处理器嵌入量子变分自编码器的离散瓶颈层，该技术在分子生成与药物研发等场景中展现出超越经典计算的表达力与采样效率。 量子退火机制与 QUBO 二进制映射 (2:42) 量子退火机通过物理演化在复杂的能量曲面中寻找全局最优解。系统首先初始化在基态易求的基线哈密顿量状态，随后平缓过渡至目标问题的哈密顿量，依据物理与数学定理引导系统收敛至目标全局最小值。在数学上，该设备原生解决二进制变量优化问题，物理学界将其定义为基于泡利矩阵与退火时刻表的伊辛（Ising）模型，计算机科学与运筹学领域则将其称为无约束二次二进制优化（QUBO）。目标函数由单个节点的线性偏置与节点间配对的二次耦合项构成。目前硬件系统已从 2 个、48 个超导量子比特扩展至包含 4400 多个可编程互联量子比特的生产级架构。 组合优化基准测试与公式化瓶颈 (6:14) 许多核心现实问题均可自然映射为 QUBO 结构，包括图匹配（Graph Matching，通过最小化结构差异寻找最优节点置换）、旅行商问题（TSP，通过置换矩阵最小化多目的地总通勤时间）以及一般的二进制整数线性规划（BILP）。Chern 报告的基准测试数据显示，在衡量求解时间与逼近能量最低点的测试中，D-Wave 量子退火机在最佳场景下相比经典计算基线实现了至少两个数量级的计算加速。然而，Chern 同时也提醒，尽管所有有限离散优化问题理论上均可转化为 QUBO，但人工将复杂业务约束转化为数学惩罚函数的流程极其繁重，亟需自动化重构工具链的研发支持。 硬件原生采样加速与经典 GPU 算力对比 (8:52) 量子退火本质上也是一种物理采样过程，这使其与基于玻尔兹曼机（Boltzmann Machines）的生成建模产生了直接技术协同。玻尔兹曼机跨越了多个学科：统计学家视其为离散二进制向量的指数族分布，物理学家将其视为热平衡状态下的伊辛自旋系统，计算机科学家则称之为二进制马尔可夫随机场（MRF）。在传统生成模型训练中，依赖经典马尔可夫链蒙特卡洛（MCMC）方法在多峰能量曲面上采样会遭遇严峻的计算瓶颈。基准对比表明，利用传统 GPU 集群模拟大规模量子系统所需的时间可达数百万年，而 D-Wave 硬件在采样速度上比经典 MCMC 算法快一个数量级以上，显著加速了生成学习中的期望梯度估计。 量子变分自编码器与药物分子研发落地 (11:54) 为将量子退火机的二进制处理能力应用于连续或复杂数据流，研究人员构建了离散变分自编码器（DVAE）与量子变分自编码器（QVAE）架构。该架构利用编码器将高维输入映射至离散二进制隐空间，并在该瓶颈层直接调用量子退火机执行优化或生成采样，最终由经典解码器重构输出。在实际产业实践中，日本烟草公司（Japan Tobacco）将量子退火嵌入 VAE 瓶颈层进行药物分子优化，实验生成的分子候选结构在理化性质直方图与定量药物相似性评价（QED）评分上均取得了显著提升。尽管经典平面图在低连通性下具备闭式求解优势，但其表达力有限；量子退火硬件的高连通拓扑为捕捉复杂高维数据分布提供了关键的表示容量。 量子退火技术在组合优化加速与生成模型采样方面已展现出明确的物理硬件优势，并在分子发现等领域实现了概念验证。然而，其规模化工业落地仍受制于严苛的约束转化门槛，未来发展取决于自动化 QUBO 映射工具链的成熟度，以及在高连通性硬件拓扑下持续扩大相对经典高级启发式算法优势的能力。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 28 日发布。视频内容为 D-Wave 高级基准测试研究员 Kevin Chern 的主题演讲，主要介绍量子退火技术的基本原理、数千量子比特硬件实现，以及其在组合优化与机器学习生成模型（如药物研发）中的前沿应用与基准测试表现。 ## TL;DR 演讲首先从物理能量曲面演化与绝热定理切入，解析了量子退火机如何将 NP-hard 离散优化问题映射为 Ising 模型与 QUBO 进行求解，并在特定基准场景下展现出超越经典算法两个数量级的加速。随后，演讲者指出量子退火不仅是优化引擎，更是天然的高效物理采样器，能颠覆性地加速玻尔兹曼机等概率图模型的训练，远超经典 GPU 与 MCMC 采样。最后，报告探讨了量子变分自编码器（QVAE）如何将复杂连续数据嵌入离散隐空间，并展示了日本烟草公司（JT）利用该技术优化药物分子属性（QED 打分）的落地案例与图建模工程策略。 ## Atmosphere & style 整场演讲氛围幽默轻松且兼具硬核技术深度，演讲者频繁穿插自嘲式调侃、Hinton 诺奖玩梗以及对硬件规模的形象吐槽，极大化解了前沿量子物理与高维数学的晦涩感。台上台下互动生动亲切，不仅有严谨的学术图表与基准对比，末尾还以邀请全场自拍合影欢乐收尾，让观众在轻松惬意的体验中领略量子计算在 AI 领域的巨大潜能。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) D-Wave 致力于研发超导物理量子计算机，核心应用方向涵盖 NP-hard 离散组合优化、微观量子系统模拟以及与机器学习的交叉融合。 - [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) 量子退火基于量子力学绝热定理，系统从构造简单的平缓势阱初始基态出发，缓慢演化并高概率收敛至目标复杂能量曲面的全局最优解。 - [03:58](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=238s) 量子退火的核心数学抽象是 QUBO（二次无约束二值优化），在理论物理中直接对应带有线性偏置场与二次耦合项的 Ising 自旋模型。 - [06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) 在特定组合优化的最佳基准测试场景中，D-Wave 量子退火机相较经典计算基线实现了超过两个数量级（100倍以上）的速度优势。 - [06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) 计算机视觉中的图拓扑匹配以及路径规划中的旅行商问题（TSP），本质上都可抽象为寻找最优二值置换矩阵并通过 QUBO 形式表述求解。 - [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) 量子退火不仅是一种寻找极值的优化过程，其物理终止测量机制使其天然成为从底层吉布斯-玻尔兹曼分布中抽取样本的高效采样引擎。 - [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) 玻尔兹曼机连接了统计物理与生成式 AI，利用量子退火机的物理直接采样能力，可有效突破经典 MCMC 算法在模型训练梯度更新上的计算瓶颈。 - [11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) 在数千量子比特规模的量子模拟与采样任务中，经典 GPU 的模拟耗时会恶化至数百万年级别，而量子退火机表现出大幅领先经典硬件的采样速度。 - [11:57](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=717s) 量子变分自编码器（QVAE）利用神经网络将复杂数据编码至离散二进制隐空间，并在瓶颈层由量子退火机协同执行生成采样与属性优化。 - [12:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=762s) 日本烟草（Japan Tobacco）将量子退火成功应用于新药分子发现，在 VAE 瓶颈层嵌入量子采样生成的候选药物在活性及 QED 药物相似性评分上表现更优。 - [12:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=762s) 面向业务建模建议先从可经典闭式求解的平面图（Planar Graph）起步验证，当遇到模型表达能力与连通度瓶颈时，再迁移至高连通度硬件拓扑的量子退火芯片。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) D-Wave 介绍与演讲者背景 演讲在一片轻松幽默的氛围中拉开序幕。演讲者 Kevin Chern 首先调侃了自己的标题幻灯片，笑称这是整个 PPT 里最漂亮的一页，并特别鸣谢了制作该幻灯片的同事 Emma。随后，他切入正题介绍了自己所在的公司——D-Wave。作为一家深耕量子计算的先驱企业，D-Wave 致力于制造真实的物理量子计算机。Kevin 幽默地指着幻灯片上的稀释制冷机（fridge）照片说，这个庞然大物“具体有多大我不清楚，反正比我整个人都要大”，而真正的核心芯片其实远比他要小，从而给观众划定了一个尺寸参考界限。在业务与研究领域上，D-Wave 主要攻坚 NP-hard 等离散组合优化问题，同时开展量子模拟研究——即利用自身可控的量子物理系统，在一定程度上去模拟研究人员希望探究的其他复杂量子体系。随后，他将话题引向当下炙手可热的领域：量子退火（Quantum Annealing）究竟与机器学习（ML）有着怎样的联系？ Kevin Chern 接着介绍了自己的个人经历与研究方向。他在 D-Wave 担任高级基准测试研究员（Senior Benchmarking Researcher），目前已在公司工作约四年。他的日常核心工作是基准测试以及开发量子-经典混合算法（quantum-classical hybrid algorithms）。他自嘲说，这项工作的本质其实就是“大部分时间枯燥地坐在书桌前，绞尽脑汁地把各类实际问题重新表述为能够在量子退火机上求解的形式”。在加入 D-Wave 之前，他在不列颠哥伦比亚大学（UBC）完成了本科和研究生学业。作为典型的温哥华科研人员，他打趣自己平时极少出门社交，“如果我待会儿讲着讲着开始像个小语言模型一样卡顿口吃，大家就知道是怎么回事了”。至于业余爱好，他列举了摄影摄像、拳击以及“做白日梦”，并坦言做白日梦其实就是“拖延家务的借口”。关于本次演讲的目标，Kevin 直率而幽默地向观众宣布：“我今天其实是来打广告的，非常感谢大家自愿坐在这里收听这条‘无法跳过的贴片广告’。”但在玩笑之余，他严肃地表示，自己最核心的期许是激发大家去深入思考量子退火机、组合优化以及生成式建模（generative modeling）三者交叉融合所蕴含的巨大潜能。 ```mermaid flowchart LR accTitle: D-Wave Core Capabilities subgraph Core["D-Wave 核心能力与应用领域"] A["量子计算机硬件研发<br/>(超导量子芯片与制冷系统)"] --> B["组合优化<br/>(NP-hard 离散优化问题)"] A --> C["量子系统模拟<br/>(物理体系微观动力学)"] A --> D["机器学习与生成模型<br/>(量子玻尔兹曼机与采样)"] end ``` ### [02:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=162s) 量子退火的基本原理与工作机制 在解释量子退火机（Quantum Annealer）的概念时，Kevin 指出，它本质上是一个执行量子退火物理过程的专用硬件设备，而量子退火本身就是一种寻找全局最优解的物理优化过程。为了直观解释这一抽象机制，他展示了一幅高低起伏、形态险峻的复杂“优化能量曲面”（optimization landscape）。在实际问题中，算法的目标就是在这片险峻的地形中找到能量最低点（即图中右下角的蓝色全局最优点）。在经典优化方法中，算法极容易被困在周围无数的高能局部极小值凹坑中无法自拔。 量子退火采用了一种极其巧妙的物理演化策略：系统并不直接从这个极其复杂的原始问题开局，而是从一个构造极其简单、基态一目了然的问题哈密顿量出发。在初始状态下，能量景观如同一个平缓的碗底，系统仅需进行简单的随机游走便能迅速滑落并稳定在全局最低点。随后，物理系统通过缓慢降低横向磁场、逐步引入待求解的目标问题。根据量子力学中的绝热定理（Adiabatic Theorem），只要这一演化过程足够缓慢且满足特定绝热物理条件，量子系统就会始终保持在瞬时基态上，最终自然而然地精确落入目标复杂问题的全局最优点。Kevin 强调，观众在脑海中可以始终保留这种“从平滑势阱渐变到复杂崎岖曲面”的图像，因为这正是量子退火在微观物理层面上真实发生的量子力学演化。 - 量子退火机是专门用于执行量子退火的物理硬件，属于专用型量子计算架构。 - 优化核心在于将目标函数的极值求解映射到物理系统的基态（最低能态）搜寻。 - 绝热定理提供了理论保证：在充分绝热条件下，系统从初态基态演化即可高概率收敛至目标问题的全局最优解。 ### [03:58](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=238s) 优化问题建模：Ising 模型与 QUBO 那么，究竟哪些类型的优化问题可以直接放入量子退火机中求解？Kevin 指出，物理学家与计算机科学家的称谓虽有不同，但本质指向同一数学结构：对于物理学家而言，这就是著名的伊辛模型（Ising Model）；而对于运筹学专家与计算机科学家来说，它被称为二次无约束二值优化（Quadratic Unconstrained Binary Optimization，简称 QUBO）。在 QUBO 目标函数 $H(x)$ 中，所有的决策变量 $x_i$ 均为二值变量（在物理体系中取值为 $\{-1, +1\}$，在经典二值逻辑中则对应 $\{0, 1\}$）。他用包含两个蓝色节点的极简拓扑图进行了解析：每个节点周围的橙色标记代表一阶线性系数（偏置场），而连接两个节点之间的青绿色边线则代表变量 $x_i$ 与 $x_j$ 之间的二阶交互项（二次耦合系数）。 为了缓解台下观众对高维数学与物理术语的焦虑，Kevin 引用了深度学习先驱 Geoffrey Hinton 的经典名言：“‘想要想象 14 维空间中的超平面，你只需要在脑海中想象一个 3 维空间，然后对着自己大声喊 14！每个人都是这么干的’。”借着这个梗，Kevin 笑着建议大家：如果看公式觉得头晕，不妨看着那个数学表达式，然后在心里大声默念“量子自旋”（quantum spins）！对于物理背景的听众，那些项无非就是泡利矩阵（Pauli matrices）$\sigma_x$ 和 $\sigma_z$，附带控制退火节奏的时间调节参数 $A(t)$ 和 $B(t)$；而对于没有物理背景的人，只需记住这就是一个二值决策变量的优化问题。随后，他展示了规模不断扩大的量子硬件拓扑图：从 2 个量子比特开始，观众只需在心里大声默念“48 个量子比特”——图上已经显现出密集的节点与耦合边；紧接着再在心里大声默念“4400 个量子比特”——幻灯片上瞬间呈现出一张宏大而极其密集的硬件拓扑网络。Kevin 坦言，面对 4400 个量子比特的真实芯片连通图，“其实我自己也根本分不清里面具体哪条线连着哪条线了，它实在太庞大了”。 ```mermaid flowchart TD accTitle: Problem Mapping to Quantum Hardware subgraph ProblemLevel["数学建模层面"] P1["实际业务场景<br/>(物流路径/图像匹配/分子构型)"] --> P2["QUBO / 二次无约束二值优化<br/>(决策变量: 二值 x_i ∈ {0, 1} 或 {-1, +1})"] P2 --> P3["Ising 模型哈密顿量<br/>(线性场 h_i + 二次耦合 J_ij)"] end subgraph HardwareLevel["硬件映射与演化层面"] P3 --> H1["量子比特与耦合器编程<br/>(对应 Pauli 矩阵与物理连通拓扑)"] H1 --> H2["物理退火演化<br/>(4400+ 量子比特协同动力学)"] H2 --> H3["基态测量与二值解输出"] end ``` ### [06:14](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=374s) 性能基准与典型组合优化场景 在谈到退火硬件的实际性能时，Kevin 展示了一张基准测试对比图表。横轴代表求解问题所需的时间消耗，纵轴代表当前求解质量与真实理论最小值的距离（越低代表结果越优越、越接近最优解）。橙色曲线代表 D-Wave 量子退火机，蓝色曲线代表经典计算对比基线。测试数据清楚表明，在最佳应用场景下，量子退火相比于经典求解器展现出了至少两个数量级（2 orders of magnitude，即超过 100 倍）的速度优势。 既然量子退火机擅长处理二值优化问题，那么现实世界中有哪些关键问题可以天然映射为二值模型？Kevin 重点列举了两个典型场景： 1. **计算机视觉中的图匹配（Graph Matching）**：给定一张蓝色图和一张橙色图，算法需要寻找蓝色节点的一种最佳置换排列（permutation），使得匹配后图的邻接拓扑结构能够最大程度地予以保留。这一过程可以通过穷举评估不同配置的能量代价来寻找最优匹配，属于标准的 QUBO 问题。 2. **旅行商问题（Traveling Salesperson Problem, TSP）**：这是大众最容易感同身受的组合优化经典问题。例如要在城市中规划拜访 7 个目的地，或者在一个下午先后前往 7 家不同的杂货店购物，如何规划途经顺序才能使总通勤时间最短。对于技术受众而言，这在本质上就是寻找一个最优的置换矩阵（Permutation Matrix），同样能够完美表述为 QUBO。 Kevin 指出，理论上诸如二值整数线性规划（0-1 Integer Linear Programming）以及所有的有限离散优化问题，都可以转化为 QUBO 形式。但在工程实践中，“理论上可行”与“求解是否高效”完全是两码事。而且，手动在草稿纸上把实际业务约束一点点推导并转化为 QUBO 代数表达式的过程极其繁琐枯燥。为此，学术界和工业界目前都在积极研发自动化编译转换框架，试图自动完成从高阶业务逻辑到底层二值哈密顿量的编译。 - 在基准测试的最佳场景中，量子退火相比经典算法实现了至少两个数量级的计算加速比。 - 图匹配是计算机视觉领域的典型 NP-hard 问题，旨在寻找保持拓扑不变性的最优节点对应关系。 - 旅行商问题在数学上可抽象为寻找最优二值置换矩阵，是验证 QUBO 建模能力的基石案例。 - 自动化建模编译器是降低 QUBO 应用门槛的关键技术，能够避免手动推导罚函数项带来的繁重人工计算。 ### [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) 从优化到采样：玻尔兹曼机与生成式模型 在探讨量子计算如何赋能药物研发（Drug Discovery）等 AI 核心热点领域时，Kevin 纠正了一个普遍的认知偏差。他指出：“我前面一直强调量子退火是一种优化过程，但这并不是全貌；我之前没有透露的是，量子退火同时也是一个天然的采样过程（sampling process）。”正是这种物理采样能力，在量子退火与现代机器学习之间架起了一座桥梁。 为了阐明这一点，他引入了玻尔兹曼机（Boltzmann Machine）的概念。由于该模型正是深度学习先驱 Geoffrey Hinton 获得诺贝尔物理学奖的重要贡献之一，Kevin 笑着调侃“如果大家最近这几天经常听说玻尔兹曼机，那正是因为 Hinton 刚拿了诺奖”。他随后从不同学科的视角解构了玻尔兹曼机： - 对统计学家而言，它是标准指数族分布（Exponential Family Distribution）的一员； - 对物理学家而言，它依然是熟悉的伊辛模型在有限温度下的热力学平衡态分布； - 对计算机科学家而言，它是一个二值马尔可夫随机场（Binary Markov Random Field, MRF）。 正如多元高斯分布用钟形曲线刻画连续随机变量的高维概率分布一样，玻尔兹曼机刻画的是高维二值向量空间上的概率分布，其数学表达式给出了观测到任意特定二值状态向量的物理概率。 玻尔兹曼机可以作为生成模型进行训练。当研究人员手头拥有一批二值数据集（例如药物分子的结构指纹）时，生成模型的目标就是学习这批数据背后的真实概率分布。而在传统的经典机器学习训练循环中，为了更新生成模型的网络参数，必须频繁地从当前模型中生成大量样本以计算梯度——这一步通常依赖经典马尔可夫链蒙特卡洛（MCMC）采样，计算开销极其巨大，常常成为整个模型训练的致命瓶颈。而这正是量子退火机的用武之地：量子物理系统在退火终止时，其读出结果本质上就是从底层物理吉布斯-玻尔兹曼分布中直接抽取出的量子统计样本。随后，Kevin 引用了 D-Wave 在国际顶级学术期刊《Science》上发表的重要论文成果，证明量子退火机在硬件层面直接执行物理采样的能力，能够为高维复杂生成模型的训练提供颠覆性的加速基础。 | 学科视角 | 对玻尔兹曼机（Boltzmann Machine）的理解与定义 | |:--- |:--- | | **统计学** | 高维离散变量上的指数族概率分布（Exponential Family） | | **理论物理学** | 处于热平衡态的伊辛自旋玻璃系统（Ising Model / Gibbs Distribution） | | **计算机科学 / AI** | 节点状态受限为二值的马尔可夫随机场（Markov Random Field） | | **生成式机器学习** | 基于物理采样驱动的无监督生成模型，训练需依赖高效模型采样更新梯度 | ### [11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) 量子退火在采样任务中的绝对加速与 GPU 对比 在评估量子退火机的实际表现时，基准测试重点关注了量子模拟与采样任务。量子系统模拟本质上是一个采样问题——研究人员必须从模型中抽取样本并计算统计量，以回答有关物理系统的各种关键问题。如果使用经典 GPU 来模拟这类具有一定量子比特规模（Qubit count）的系统，随着系统尺寸的增加，所需时间呈指数级暴涨。这场对比的核心结论非常直观："在 GPU 上需要耗费数百万年才能完成模拟"，而量子退火机则能直接利用自身物理特性进行高效演化。 在对比马尔可夫链蒙特卡洛（Markov Chain Monte Carlo, MCMC）采样的基准测试中，横轴代表时间，纵轴代表样本量。数据显示，代表量子计算机的采样曲线相较于经典 GPU 实现了整整一个数量级以上的速度飞跃（Speedup）。这一系列测试证明，量子退火机在处理复杂的概率采样与统计物理模拟时具备极高的速度与效率。 - **GPU 模拟瓶颈**：在达到数千量子比特规模的量子模拟任务中，经典硬件（包括高性能 GPU 集群）的计算耗时会迅速恶化到数百万年级别。 - **MCMC 采样对比**：在单位时间内获取同等质量样本的测试中，D-Wave 量子退火机比 GPU 展现出超过一个数量级的加速优势。 --- ### [11:57](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=717s) 结合机器学习：量子变分自编码器（QVAE）与离散隐空间 既然量子退火机天生擅长解决二进制/离散变量问题，那么这种能力如何扩展到更广泛的实际应用中？答案是通过现代机器学习技术。利用离散变分自编码器（Discrete VAE）与量子变分自编码器（Quantum Variational Autoencoder, QVAE），可以将任意连续或复杂的真实世界数据映射到二进制隐空间（Binary latent space）中。 ```mermaid flowchart LR accTitle: Quantum Variational Autoencoder Workflow subgraph Encoding ["编码映射"] A["输入数据 / 分子特征"] --> B["编码器 (Encoder)"] B --> C["二进制隐空间 (Binary Space)"] end subgraph QuantumSampling ["量子采样与优化"] C --> D["D-Wave 量子退火机 (Bottleneck)"] D --> E["优化后的二进制表征"] end subgraph Decoding ["重构生成"] E --> F["解码器 (Decoder)"] F --> G["目标分子 / 生成数据"] end ``` 这种架构的核心逻辑十分明确：首先使用深度神经网络编码器将输入数据压缩映射为二进制表示；随后，利用量子退火机在瓶颈层（Bottleneck layer）充当强大的优化引擎和生成模型先验，直接在离散隐空间中执行采样与性质优化；最后，解码器将优化后的二进制向量重构回原始输入空间，从而生成满足特定物理或化学属性的新数据。 - **双重角色定位**：量子退火机在机器学习中既是二进制空间的高性能优化引擎，也是具备离散隐空间的强力生成模型（Generative model）。 - **优化与生成结合**：通过在潜变量空间直接引导采样分布，能够有针对性地最大化生成样本的特定目标属性。 --- ### [12:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=762s) 日本烟草（Japan Tobacco）药物发现案例与图结构建模建议 在实际商业落地方面，日本最大的烟草及制药企业之一——日本烟草公司（Japan Tobacco, JT）将 D-Wave 量子技术成功应用于新药研发（Drug Discovery）。在该场景下，研究团队在 VAE 的瓶颈层嵌入了量子计算机。实验直方图数据显示，量子退火辅助生成的候选分子在化合物活性指标以及定量药物相似性评估（QED score）等打分上，均显著优于传统经典生成方案，展现出更加理想的成药潜能。 ``` +--------------------------+-----------------------+---------------------------------------+ | 应用方案 | 瓶颈层先验 / 优化机制 | 药物属性打分表现 | +--------------------------+-----------------------+---------------------------------------+ | 传统经典生成模型 (VAE) | 连续高斯 / 经典采样 | 基准分布，优良分子产出率有限 | +--------------------------+-----------------------+---------------------------------------+ | 量子变分自编码器 (QVAE) | D-Wave 量子退火采样 | 活性打分分布更优，QED 药物相似性提升 | +--------------------------+-----------------------+---------------------------------------+ ``` 对于想要将业务问题映射到二进制空间的研究人员，在系统拓扑与图结构选择上存在递进式的工程策略： 1. **经典平面图起步（Planar Graph）**：可先尝试将数据映射到结构简单的平面图上。由于平面图连通度低，其采样任务往往能够直接以闭式解（Closed-form）或低复杂度算法在经典计算机上快速求解。 2. **评估表达能力瓶颈**：如果发现平面图由于连通性限制而无法捕捉复杂的变量依赖（即模型表达能力不足、性能不达标）。 3. **引入量子退火机**：转而利用 D-Wave 具备高连通度硬件拓扑（如 Pegasus/Zephyr 架构）的量子芯片执行全局采样与联合优化。 在报告尾声，演讲者轻松地向全场听众表示，在结束之前他想和大家拍一张合影自拍："因为对我来说站在台上还挺紧张的，我想把这一刻记录下来"，并打趣邀请现场观众无论是捂脸还是扮鬼脸都随意，随后在全场致谢中圆满结束了演讲。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) 具体有多大我不清楚，反正比我整个人都要大 — Kevin Chern > [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) 大部分时间枯燥地坐在书桌前，绞尽脑汁地把各类实际问题重新表述为能够在量子退火机上求解的形式 — Kevin Chern > [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) 如果我待会儿讲着讲着开始像个小语言模型一样卡顿口吃，大家就知道是怎么回事了 — Kevin Chern > [00:04](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=4s) 我今天其实是来打广告的，非常感谢大家自愿坐在这里收听这条‘无法跳过的贴片广告’ — Kevin Chern > [03:58](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=238s) ‘想要想象 14 维空间中的超平面，你只需要在脑海中想象一个 3 维空间，然后对着自己大声喊 14！每个人都是这么干的’ — Geoffrey Hinton > [03:58](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=238s) 其实我自己也根本分不清里面具体哪条线连着哪条线了，它实在太庞大了 — Kevin Chern > [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) 我前面一直强调量子退火是一种优化过程，但这并不是全貌；我之前没有透露的是，量子退火同时也是一个天然的采样过程 — Kevin Chern > [08:52](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=532s) 如果大家最近这几天经常听说玻尔兹曼机，那正是因为 Hinton 刚拿了诺奖 — Kevin Chern > [11:07](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=667s) 在 GPU 上需要耗费数百万年才能完成模拟 — Kevin Chern > [12:42](https://www.youtube.com/watch?v=s9sB2ZPOO04&t=762s) 因为对我来说站在台上还挺紧张的，我想把这一刻记录下来 — Kevin Chern ## Mentioned Kevin Chern, Emma, D-Wave, 不列颠哥伦比亚大学 (UBC), Geoffrey Hinton, Ising Model (伊辛模型), QUBO (二次无约束二值优化), Boltzmann Machine (玻尔兹曼机), Science, MCMC (马尔可夫链蒙特卡洛), QVAE (量子变分自编码器), Japan Tobacco (日本烟草公司) ## Video description Kevin Chern, Senior Benchmarking Researcher, D-Wave Abstracts Quantum annealing is a quantum computing paradigm known for its applications in combinatorial optimization and machine learning. In optimization, quantum annealers provide heuristic solutions to NP-hard combinatorial optimization problems. In machine learning, quantum annealers can be leveraged as generative models in various domains ranging from drug discovery to image generation. In this presentation, I will introduce the principles of quantum annealing with a focus on D-Wave's implementations featuring thousands of programmable qubits. I will also share recent applications and benchmarking results highlighting its practical benefits. [Source](https://www.youtube.com/watch?v=s9sB2ZPOO04)

Source quotes

[0:04] Kevin Chern: 具体有多大我不清楚，反正比我整个人都要大 [0:04] Kevin Chern: 大部分时间枯燥地坐在书桌前，绞尽脑汁地把各类实际问题重新表述为能够在量子退火机上求解的形式 [0:04] Kevin Chern: 如果我待会儿讲着讲着开始像个小语言模型一样卡顿口吃，大家就知道是怎么回事了 [0:04] Kevin Chern: 我今天其实是来打广告的，非常感谢大家自愿坐在这里收听这条‘无法跳过的贴片广告’ [3:58] Geoffrey Hinton: ‘想要想象 14 维空间中的超平面，你只需要在脑海中想象一个 3 维空间，然后对着自己大声喊 14！每个人都是这么干的’ [3:58] Kevin Chern: 其实我自己也根本分不清里面具体哪条线连着哪条线了，它实在太庞大了 [8:52] Kevin Chern: 我前面一直强调量子退火是一种优化过程，但这并不是全貌；我之前没有透露的是，量子退火同时也是一个天然的采样过程 [8:52] Kevin Chern: 如果大家最近这几天经常听说玻尔兹曼机，那正是因为 Hinton 刚拿了诺奖 [11:07] Kevin Chern: 在 GPU 上需要耗费数百万年才能完成模拟 [12:42] Kevin Chern: 因为对我来说站在台上还挺紧张的，我想把这一刻记录下来

Source quotes

[0:04] Kevin Chern: the most beautiful slide of the entire deck. [0:04] Kevin Chern: larger than me [0:04] Kevin Chern: smaller than me [0:04] Kevin Chern: stuttering like a little language model [0:04] Kevin Chern: unskippable ad [2:42] Geoffrey Hinton: To deal with hyperplanes in a 14-dimensional space, visualize a 3D space and say 14 to yourself very loudly. Everyone does it. [2:42] Kevin Chern: say quantum spins to yourself very loudly [11:07] Kevin Chern: punch line is takes millions of years on GPUs [11:54] Kevin Chern: in some sense always [14:52] Kevin Chern: Well, actually before before I let everyone go, I'm going to take a selfie because this is quite scary for me. So, I want to remember this.

Original transcript

[0:04] So, I don't really know what to say [0:05] about a title slide, but this one's the [0:07] most beautiful slide of the entire deck. [0:10] So, shout out to Emma, my colleague, for [0:12] making this slide. Uh, so you can enjoy [0:14] it for the next half second. Here's [0:16] D-Wave. Uh, D-Wave is a quantum [0:18] computing company, and we build quantum [0:21] computers. So, here's a quantum computer [0:24] or rather this this is a this is a [0:26] fridge. It's it's I don't know how big [0:28] it is. it's larger than me. Um, and we [0:31] also solve optimization problems. So, [0:34] think np hard problems or discrete [0:37] optimization problems. [0:40] Okay. And we also perform quantum [0:42] simulations. So that means if we have a [0:45] quantum system that doesn't look exactly [0:47] like the system that you want to [0:49] simulate, then we can do that uh to some [0:51] extent. Right? And now I'm going to talk [0:55] a little bit more about machine [0:56] learning. So what is quantum? What does [0:58] quantum uh analing have to do with [1:01] machine learning? [1:03] And this is this is the actual quantum [1:05] computer. And I don't know what size it [1:07] is, but it's smaller than me. So you [1:09] have a bound. All right. And my name is [1:12] Kevin Chern. I'm a uh benchmarking [1:14] researcher. So at D-Wave, I uh benchmark [1:18] and develop quantum classical hybrid [1:21] algorithms. So what that means is I'm [1:24] sitting in front of my desk most of the [1:25] time and I try to formulate problems in [1:28] a way that can be solved on a quantum [1:31] analer [1:33] and before I joined D-Wave and I've been [1:35] at DW for about four years now before I [1:37] joined D-Wave I did my undergrad and [1:39] graduate degree at the University of [1:42] British Columbia. So shout out to those [1:43] Vancouver folks and that that's [1:47] definitely not a Vancouver folk. We we [1:50] would we we wouldn't make any noise. Uh [1:52] so um but uh yeah, so all that to say I [1:58] don't go out a lot. So you know if I [2:00] start stuttering like a little language [2:01] model, you know what's up. Okay. So well [2:05] I also have some hobbies. Videography, [2:07] boxing, daydreaming. That just means [2:11] procrastinating from my chores. Okay. So [2:14] presentation goal. Um quite [2:16] straightforward. I'm here to deliver an [2:18] ad. And so, thank you all for uh [2:21] voluntarily sitting in to this uh [2:24] unskippable ad. Okay. But on a more [2:28] serious note, uh I hope to excite [2:31] everyone and inspire everyone to start [2:33] thinking about the intersection of [2:35] quantum analers, combinatorial [2:37] optimization, and generative modeling. [2:42] Okay. So, what is a quantum analer? A [2:44] quantum analer is a physical device that [2:47] performs quantum analing. And quantum [2:50] analing is an optimization process. So [2:53] here I'm showing a uh a a an example of [2:57] a optimization landscape. So you're [3:00] trying to find the minimum point in this [3:03] in this uh very ragged shape here. [3:07] And the way quantum analing finds that [3:10] point. So the the minimum point would be [3:12] the blue point at the bottom right. [3:14] Right. So the way quantum analing works [3:16] is by starting off with a simple [3:19] problem. So in this case you can you [3:21] know you can imagine doing a random walk [3:22] or just walking around the space and [3:24] you're just going to drop right into the [3:26] global optimum right and what quantum [3:28] analing does is slowly introduces your [3:31] original problem. And if you continue to [3:33] do this random walk on that space then [3:36] there's a theorem that guarantees that [3:39] if you do this in the right conditions [3:42] then you'll land in the global optimum [3:44] and that's what's happening uh you or [3:47] you can think of this you can keep this [3:48] image in mind when you when I when I say [3:50] quantum analing that's what's happening [3:52] physically but quantum mechanically [3:55] okay so what sort of uh optimization [3:58] problems or what sort of optimization [4:00] problems can I solve [4:02] with a quantum analer. And for the [4:04] physicists out there, that's the easing [4:06] model or ising model, whichever [4:08] pronunciation you like. And for the [4:10] operations researchers or computer [4:12] scientists, that's quadratic [4:14] unconstrained binary optimization. [4:18] Okay, so it's called binary because as [4:20] you see here, this is the expression [4:21] you're trying to minimize h of x and the [4:24] x i values, the decision variables are [4:27] binary. So they can take values minus [4:29] one or plus one. And that's it. And you [4:31] can visualize that using two vertices uh [4:34] in blue over there. Here I'm showing a [4:37] system with two bits. Right? So you can [4:40] have + one minus one + one + one and so [4:43] on. And the orange terms correspond to [4:47] linear coefficients. Okay? So that's [4:49] denoted with the little orange around [4:50] the circles. And quadratic means there's [4:54] an interaction between x i and xj. So [4:58] that's denoted by the turquoise line. [5:03] All right. So I'm going to interject [5:04] with a quote here. Uh you might have [5:06] heard of uh this researcher. Um to deal [5:09] with hyperplanes in a 14-dimensional [5:11] space, visualize a 3D space and say 14 [5:15] to yourself very loudly. Everyone does [5:17] it. Um great. So tada. You can say you [5:23] can look at that expression and say to [5:24] yourself quantum spins very loudly. And [5:26] for the physicist that's your poly [5:28] matrices x and zed and uh as and bs [5:32] terms just correspond to the kneeling. [5:34] Okay. So say quantum spins to yourself [5:36] very loudly and just take the the if if [5:38] you're getting a little uh confused with [5:41] the jargon you're solving an [5:44] optimization problem. That optimization [5:46] problem is binary. [5:48] And then you say 48 cubits to yourself [5:49] very loudly. Here's a picture of that. [5:51] So instead of looking at that two cubit [5:53] system, you have 48 cubits here. Um each [5:56] edge again corresponds to the uh [5:58] quadratic coefficients that you can [6:00] program [6:02] and then you say 4,400 cubits to your [6:04] cell very loudly again and that's yeah I [6:07] can't I can't really tell what's [6:08] happening there either it's it's quite [6:10] large [6:12] okay [6:14] so how well do we solve optimization [6:17] problems on the x-axis you don't need to [6:20] pay attention to the details so you can [6:21] just follow what I'm saying uh the top [6:24] the the x-axis is shows you the time. So [6:26] the amount of time it takes to solve [6:28] this uh problem and the y- axis tells [6:30] you how close you are to the minimum. So [6:33] lower the better and in orange obviously [6:36] is the quantum analer and in blue it's a [6:40] classical computer. So here in the best [6:43] case scenarios we see there's at least a [6:44] two order of magnitude speed up. [6:48] Right? So the next question is you know [6:51] when do we actually have binary problems [6:54] and quite often so one example for the [6:56] computer vision folks is graph matching [6:59] right so here we have two graphs one in [7:01] blue one in orange how do you find a [7:05] matching or a permutation of the blue [7:07] nodes such that most of the adjacency [7:10] structure is preserved [7:12] so that's one configuration you can try [7:15] and then you try another one and you see [7:17] which one is lower right so this is a [7:19] type of problem that can be formulated [7:22] as a quadratic unconstrained binary [7:24] optimization problem [7:28] another example that I think most people [7:30] can uh relate to more is the traveling [7:34] salesperson's problem okay so you have [7:37] multiple destinations here I've plotted [7:39] seven cities or maybe you have like [7:41] grocery trips where you need to visit [7:43] seven different stores uh that's one way [7:46] you can visit them you want to minimize [7:48] your uh commuting time. So, oops, that's [7:52] one way. Here's another way. Again, you [7:55] can formulate this as a quadratic [7:57] unconstrained binary optimization [7:58] problem. Uh for the technical folks, uh [8:02] you can think of it as a permutation [8:05] matrix, right? You're finding the [8:06] optimal permutation matrix here. [8:10] Okay, so what else? Um I can keep going [8:12] on. Um but you know there's binary [8:14] integer linear programming yada yada and [8:18] all finite optimization problems. So [8:21] technically that's true but is it [8:23] efficient? That's a separate question [8:24] right? But there are also automated [8:26] frameworks. Oh yeah and and to to [8:28] actually formulate these final [8:31] optimization problems as a cubo or sorry [8:33] as a quadratic constraint binary [8:35] optimization problem. I'm just going to [8:36] say cubo from now on. But to formulate [8:39] it as a cubo it's it's it's tedious. [8:41] there's some uh manual um work on paper [8:44] that you have to do. So there are also [8:46] some efforts to automate that process [8:49] and I've linked some of the papers [8:50] below. [8:52] Uh but what about drug discovery? That's [8:54] that's that's you know that's a hot [8:56] topic in machine learning. So what can [8:57] we do in that space? [9:00] Earlier I said quantum analing is an [9:02] optimization process but that's that's [9:04] not the full picture. Uh what I didn't [9:07] say is it's also a sampling process. So, [9:11] for those who are familiar with Boltzman [9:14] machines, and if you've heard of [9:15] Boltzman machines over the past couple [9:17] days, it's because it's something that [9:19] uh Jeffrey Hinton has worked on. Um, but [9:22] a Bolsman machine for the statisticians, [9:24] it's [9:26] a member of the exponential family [9:29] distribution. For the physicists, it's [9:30] still an model, easing model. Uh, for [9:33] computer scientists, mark of random [9:35] field, but limited to binary. Okay, so [9:38] what that means is it's a distribution. [9:40] So, so it's you know it's uh it's a [9:43] binary distribution highdimensional just [9:45] like the just the same way a multivaried [9:47] gausian distribution is like a bell [9:49] curve is a distribution we have binary [9:51] data instead right so it's got this [9:53] expression that tells you the [9:54] probability of observing a particular [9:58] binary vector realization [10:01] okay and at that point I I I described a [10:03] ball machine and you just say quantum [10:05] students to yourself references below [10:08] you can you can read more about the [10:09] quantum aspect ect of that um oh okay [10:12] okay so here's here's the point of this [10:14] slide [10:16] the quantum balsin machine or balsman [10:17] machines [10:20] can be trained right so you have a [10:22] binary data set you can learn that data [10:25] set so you can learn a generative model [10:28] by using sampling right so uh what that [10:32] means is you need to in order to train a [10:34] generative model on that binary space [10:37] you need to sample from your model so [10:39] iteratively update that. Okay. And you [10:42] don't need to understand this [10:42] expression. It's more of a clip art than [10:44] than something you need to absorb right [10:46] now. Uh point is to generate data to to [10:50] train your model. You need to sample [10:52] from that. And the quantum analer does [10:55] that. Here's a figure from one of our [10:57] papers uh published in science. This [11:01] figure you don't need to digest every [11:04] detail. Point is x-axis is the system [11:07] size. So the number of cubits and y-axis [11:10] is uh the amount of time it takes for a [11:12] GPU to simulate that system. Okay, [11:17] punch line is takes millions of years on [11:19] GPUs. [11:21] This is this is for quantum simulation. [11:23] Okay, and it's a sampling task because [11:25] you need to sample from your model and [11:28] compute statistics in order to answer [11:30] some question about your system. Here's [11:33] another one. U Markov chain Monte Carlo [11:36] X-axis is time Y-axis is the sample [11:39] size. So here again we see the quantum [11:42] computer in blue and orange uh we see a [11:45] full order of magnitude speed up even [11:48] more uh against a GPU. [11:52] Point is very good at sampling very fast [11:54] very efficient. Uh so when else do we [11:57] have binary problems? The answer is in [12:00] some sense always, right? Because tada, [12:02] we have machine learning to the rescue. [12:04] So we have discrete variational [12:06] auto-enccoders and quantum variational [12:07] auto-enccoders. And the the the punch [12:11] line here is basically you have some [12:14] data and you have an autoenccoder. So [12:16] what do you do with that? You map it to [12:19] binary space. You map your data to [12:20] binary space. So now you can work with [12:22] optimization in that space and decode [12:25] that back to your original input space. [12:29] I've attached a couple links. Uh you [12:31] don't need to take a picture of that. [12:32] You can Google it. You'll find it. [12:35] Okay. So, quantum analing optimization [12:39] and machine learning. [12:42] You know, it's a we we we have this [12:44] quantum analer that is a powerful [12:46] optimization optimization engine in [12:49] binary space and it's also doubled as a [12:53] generative model with binary latent [12:55] space. So the idea here is to sort of [12:58] optimize your data generation over that [13:01] binary space. [13:05] Again, map data to binary space. Use [13:07] quantum computer to optimize that space. [13:09] So you want to optimize some aspect of [13:12] your data generation. Okay. And here's a [13:15] here's a slide from one of our clients [13:18] uh Japan Tobacco. They're one of the [13:20] largest tobacco companies in Japan and [13:23] they have some drug discovery [13:24] application [13:26] and what they do is well they use a [13:28] varational autoenccoder with the quantum [13:31] computer at your at at the bottleneck [13:33] layer. Okay. And the point here is [13:36] better is to the left for the first two [13:38] histograms and then they show that the [13:39] quantum computer is you know it's [13:42] generates it generates scores that are [13:44] that are more favorable for whatever [13:46] compound or application they have. And [13:48] QED score again it's in the opposite [13:50] direction. Cool. And here are again some [13:53] more useful URLs. You can Google it. I I [13:56] trust you'll find it. Uh takeaway map [14:00] data to binary space. Use quantum [14:02] computer to perform optimization. And um [14:05] maybe I'll just leave one one one more [14:07] note here which is [14:10] you can you can you can map data to [14:11] binary space with a different structure [14:13] from the quantum computer right you can [14:15] you can use a simple graph structure [14:17] like a planer graph so you can solve [14:19] your sampling tasks in you know closed [14:22] form right you can do that or you can [14:26] use and and and and maybe you'll find [14:28] that it's not expressive enough right [14:29] this planer graph has limitations in its [14:31] connectivity so you you know, give that [14:34] a try and then try to optimize in that [14:36] space on that planer graph. Again, [14:37] that's that should be quite simple [14:39] because it's it's planer uh low [14:41] connectivity. Um, and then if if if uh [14:45] if you're finding that it's not very [14:46] performant, then try the try using a [14:49] quantum computer. [14:52] Okay, so that's everything. Um, I'll be [14:56] avail Well, actually before before I let [14:57] everyone go, I'm going to take a selfie [14:59] because this is [15:02] quite scary for me. So, I'm I want to [15:06] remember this. [15:08] Feel free to cover your face or make [15:09] some goofy face. [15:13] All right. Thank you all for uh Yeah. [15:16] Thank you.