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

TL;DR

Researchers have developed a way to train Restricted Boltzmann Machines (RBMs) using "Ising machines"—specialized hardware that mimics physical systems to solve optimization problems. By moving the training process directly into the hardware, they achieved a 29x speedup and a 1000x reduction in energy consumption compared to a Google Tensor Processing Unit (TPU).

What Problem Does This Paper Address?

Modern Machine Learning (ML) is a power-hungry beast. Training state-of-the-art models requires massive GPU clusters and millions of dollars in electricity. While we’ve spent decades perfecting the von Neumann architecture (the standard CPU/GPU design), it struggles with the "sampling" phase required by energy-based models like RBMs.

RBMs are a type of neural network that learns the probability distribution of data. To train them, you have to perform "Gibbs sampling"—a process of jumping between states to find the network's equilibrium. On a digital chip, this is a slow, iterative calculation. This paper asks: What if we used a physical system that naturally settles into these states on its own?

Methods

The authors leverage Ising Machines, which are nature-inspired computing substrates designed to find the lowest energy state of a system of "spins" (variables that are either +1 or -1). Because the math governing an Ising machine is nearly identical to the energy function of an RBM, the hardware can act as a physical "twin" of the neural network.

The paper proposes two architectural shifts:

  1. Gibbs Sampler (GS) Architecture: A "hybrid" approach where the Ising machine handles the heavy lifting of sampling, but a digital host (like a TPU) still manages the overall training loop and weight updates.
  2. Boltzmann Gradient Follower (BGF): The more radical approach. Here, the authors added charge-redistribution circuits to 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 constantly talk back to a digital host, they eliminated a massive communication bottleneck.

Datasets

The researchers evaluated their design across a diverse set of benchmarks to prove its versatility:

  • Image Classification: MNIST (digits), 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 and Ideas

  • The Mathematical Handshake: The core "aha moment" is that RBM training is essentially a search for low-energy states. Since Ising machines are physical systems that literally seek low-energy states via their own dynamics, they are the perfect "accelerator" for this specific math.
  • Breaking Amdahl’s Law: In the GS architecture, the system is limited by how fast the host can send and receive data. The BGF architecture bypasses this by making the hardware "self-sufficient"—it follows the gradient and updates itself, only reporting the final results to the host.
  • Embracing "Mortal Computation": Inspired by Geoffrey Hinton’s recent talks, the authors show that analog hardware doesn't need to be perfect. Even with 10-30% noise and variation in the electronic components, the BGF architecture still reached the same accuracy as perfect software simulations because the underlying algorithm is inherently stochastic (randomized).

Key Results

  • Speed: The BGF architecture achieved a 29x geometric mean speedup over a TPU (v1).
  • Efficiency: It demonstrated a staggering 1000x reduction in energy compared to the TPU.
  • Density: A 1600x1600 node BGF array would occupy only about $21 mm^2$—a tiny fraction of the size of a modern high-end processor.
  • Quality: Despite the analog nature and modifications to the training algorithm, the final model accuracy remained virtually identical to traditional software-based training (e.g., 98.8% on MNIST).

Limitations and Future Work

While the results are impressive, the authors acknowledge several hurdles:

  • Fixed Capacity: Unlike a GPU, which can run any size model (given enough time), these Ising substrates have a fixed number of nodes. Scaling to massive, billion-parameter models would require complex multi-chip architectures.
  • Hardware Overhead: While the core is efficient, the system still needs Analog-to-Digital Converters (ADCs) to read out the final weights, which are area-intensive.
  • Algorithm Co-evolution: This study focused on RBMs. The next big challenge is determining how to adapt these nature-based substrates to more common architectures, like Transformers or standard Deep Neural Networks, which currently rely on backpropagation—a process that is notoriously difficult to implement in analog hardware.

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 the future. 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.
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},
}

Loading related content...

Figure