Supporting Energy-Based Learning With An Ising Machine Substrate: A Case Study on RBM

Uday Kumar Reddy Vengalam, Yongchao Liu, Tong Geng, Hui Wu, Michael C. Huang

Abstract

Nature apparently does a lot of computation constantly. If we can harness some of that computation at an appropriate level, we can potentially perform certain types of computation (much) faster and more efficiently than we can do with a von Neumann computer. Indeed, many powerful algorithms are inspired by nature and are thus prime candidates for nature-based computation. One particular branch of this effort that has seen some recent rapid advances is Ising machines. Some Ising machines are already showing better performance and energy efficiency for optimization problems. Through design iterations and co-evolution between hardware and algorithm, we expect more benefits from nature-based computing systems. In this paper, we make a case for an augmented Ising machine suitable for both training and inference using an energy-based machine learning algorithm. We show that with a small change, the Ising substrate accelerates key parts of the algorithm and achieves non-trivial speedup and efficiency gain. With a more substantial change, we can turn the machine into a self-sufficient gradient follower to virtually complete training entirely in hardware. This can bring about 29x speedup and about 1000x reduction in energy compared to a Tensor Processing Unit (TPU) host.

Introduction

Modern Machine Learning (ML) is increasingly constrained by the energy demands of standard computing architectures. While the von Neumann architecture (the foundation of modern CPUs and GPUs) has been perfected over decades, it faces significant bottlenecks when executing energy-based models like Restricted Boltzmann Machines (RBMs). RBMs are neural networks that learn probability distributions of data through an iterative process called "Gibbs sampling." On digital chips, this sampling requires slow, repetitive calculations to find a network's equilibrium.

This research explores a nature-inspired alternative: Ising machines. These specialized hardware substrates are designed to find the lowest energy state of a system of "spins" (variables of +1 or -1). Because the mathematical energy function of an RBM is nearly identical to that of an Ising machine, the hardware can function as a physical "twin" of the neural network, allowing the system to naturally settle into the required states through its own physical dynamics.

Proposed Architectures

The researchers propose two distinct architectural shifts to integrate Ising machines into the ML training pipeline:

  1. Gibbs Sampler (GS) Architecture: This is a hybrid approach where the Ising machine handles the computationally intensive sampling phase. However, a digital host (such as a TPU) remains responsible for the overall training loop and weight updates. While efficient, this model is still limited by the communication overhead between the host and the accelerator.
  2. Boltzmann Gradient Follower (BGF): This more radical design introduces charge-redistribution circuits directly into the coupling units. This allows the hardware to update its own weights in-place based on the samples it generates. By removing the need to communicate with a digital host during the training loop, the BGF architecture eliminates the primary communication bottleneck, becoming a self-sufficient gradient follower.

Experimental Methodology

The design was evaluated across a diverse set of benchmarks to demonstrate its versatility across different ML tasks:

  • Image Classification: MNIST, KMNIST (Japanese characters), FMNIST (fashion), EMNIST (letters), CIFAR10, and SmallNorb.
  • Recommendation Systems: The MovieLens 100k dataset.
  • Anomaly Detection: European Credit Card Fraud Detection dataset.

Key Insights

The Mathematical Handshake

The core efficiency of this approach stems from the alignment between RBM training and Ising physics. Since RBM training is essentially a search for low-energy states, and Ising machines are physical systems that seek low-energy states via their own dynamics, the hardware acts as a native accelerator for the underlying mathematics.

Bypassing Amdahl’s Law

In the hybrid GS architecture, system performance is capped by the speed at which the host can exchange data with the accelerator. The BGF architecture bypasses this limitation by making the hardware self-sufficient. By following the gradient and updating itself internally, it only needs to report final results to the host, significantly increasing throughput.

"Mortal Computation" and Robustness

Inspired by the concept of "mortal computation," the authors demonstrate that analog hardware does not require digital perfection. Even with 10-30% noise and variation in electronic components, the BGF architecture achieved the same accuracy as perfect software simulations. The stochastic nature of the RBM algorithm inherently compensates for hardware imperfections.

Results

The BGF architecture demonstrated transformative improvements over traditional hardware:

  • Performance: Achieved a 29x geometric mean speedup over a Google TPU (v1).
  • Energy Efficiency: Demonstrated a 1000x reduction in energy consumption compared to the TPU.
  • Area Efficiency: A 1600x1600 node BGF array would occupy approximately $21 mm^2$, a small fraction of the area required by modern high-end processors.
  • Model Quality: Despite the analog implementation, final model accuracy remained high (e.g., 98.8% on MNIST), matching traditional software-based training.

Limitations and Future Work

While the BGF architecture offers a path toward ultra-efficient AI, several challenges remain. Unlike GPUs, Ising substrates have a fixed capacity; scaling to billion-parameter models would require complex multi-chip interconnects. Additionally, the system requires Analog-to-Digital Converters (ADCs) for weight readout, which are area-intensive components. Future research will focus on adapting these nature-based substrates to more common architectures, such as Transformers, which currently rely on backpropagation—a process more difficult to implement in analog hardware than the energy-based learning used by RBMs.

Summary updated

Abstract

Nature apparently does a lot of computation constantly. If we can harness some of that computation at an appropriate level, we can potentially perform certain type of computation (much) faster and more efficiently than we can do with a von Neumann computer. Indeed, many powerful algorithms are inspired by nature and are thus prime candidates for nature-based computation. One particular branch of this effort that has seen some recent rapid advances is Ising machines. Some Ising machines are already showing better performance and energy efficiency for optimization problems. Through design iterations and co-evolution between hardware and algorithm, we expect more benefits from nature-based computing systems. In this paper, we make a case for an augmented Ising machine suitable for both training and inference using an energy-based machine learning algorithm. We show that with a small change, the Ising substrate accelerate key parts of the algorithm and achieve non-trivial speedup and efficiency gain. With a more substantial change, we can turn the machine into a self-sufficient gradient follower to virtually complete training entirely in hardware. This can bring about 29x speedup and about 1000x reduction in energy compared to a Tensor Processing Unit (TPU) host.
BibTeX
@inproceedings{vengalam2023supporting,
  author = {Uday Kumar Reddy Vengalam and Yongchao Liu and Tong Geng and Hui Wu and Michael C. Huang},
  title = {{Supporting Energy-Based Learning With An Ising Machine Substrate: A Case Study on RBM}},
  booktitle = {MICRO},
  year = {2023},
  doi = {10.1145/3613424.3614315},
}
Figure