🤖 machine learning

Generative Models on Analog Hardware with Dynamics

This paper introduces Analog Interaction Systems (AIS), a unified framework that bridges the gap between fixed physics-determined analog hardware and flexible generative modeling by employing time-varying parameters and hidden states, achieving a 2-orders-of-magnitude energy reduction and significantly improved FID scores on MNIST datasets compared to prior hardware-based approaches.

Original authors: Yu-Neng Wang, Sara Achour

Published 2026-06-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yu-Neng Wang, Sara Achour

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot to draw pictures.

The Old Way (Digital Computers):
Currently, we use powerful digital computers (like the GPUs in your gaming PC or data centers) to do this. These computers are like very fast, very precise accountants. To draw a picture, they break the task down into millions of tiny, discrete math steps. They calculate, stop, calculate again, and stop again. This process is incredibly accurate but uses a massive amount of electricity, like running a high-end air conditioner just to draw a sketch.

The New Idea (Analog Hardware):
The authors of this paper propose using "analog hardware" instead. Think of this not as an accountant, but as a giant, interconnected playground of swings and springs.

In this playground, the "computation" happens naturally. If you push one swing, the physics of the ropes and chains automatically pulls on the neighboring swings.
The system doesn't count steps; it just flows. It solves problems by letting the physical laws (like gravity or magnetism) do the work. This is incredibly energy-efficient, like using the wind to move a sailboat instead of burning fuel.

The Big Problem:
There is a mismatch. Modern AI models are like flexible software that can learn any shape or pattern. But our analog playground is rigid. The swings are connected in a specific way by physics, and they can only move in specific patterns. It's like trying to teach a swing set to draw a perfect circle when it's only designed to swing back and forth. The "playground" isn't naturally flexible enough to learn complex images on its own.

The Solution: "Analog Interaction Systems" (AIS)
The authors created a new framework called Analog Interaction Systems (AIS) to bridge this gap. They didn't try to force the swings to act like a digital computer; instead, they found clever ways to make the playground more expressive without breaking the laws of physics.

They introduced three main "tricks" to make the rigid swings capable of drawing complex pictures:

Time-Chunked Controls (The "Conductor"):
Instead of the connections between swings being fixed forever, imagine a conductor who changes the rules every few seconds. For the first second, the swings are connected tightly; for the next second, they are connected loosely. By changing the "rules of the game" over time, the system can create much more complex movements than if the rules stayed the same.
Hidden Swings (The "Ghost Dancers"):
The authors added "hidden" swings that you can't see in the final picture but are connected to the visible ones. These hidden swings act like a secret support team. They dance around, helping the visible swings find the right path to form a picture, even though you never see the hidden ones in the final result. This gives the system much more "brainpower" without needing more visible parts.
Letting Go of the Path (The "Destination-Only" Goal):
Usually, when training AI, we tell it exactly how to move at every single step (like a GPS giving turn-by-turn directions). But because the analog hardware is rigid, forcing it to follow a specific path is impossible.
Instead, the authors used a GAN (Generative Adversarial Network) approach. They told the system: "I don't care how you get there, just make sure you end up at the right destination."
- The Generator: The analog playground tries to turn random noise into an image.
- The Critic: A digital computer looks at the final image and says, "That looks like a cat," or "That looks like garbage."
- The playground learns by trial and error, finding its own unique, winding path from noise to image, rather than being forced to follow a straight line.

The Results:
The team tested this on simple image datasets (like handwritten numbers and fashion items).

Performance: Their analog system produced images that were 3 to 4 times better than previous attempts at analog image generation.
Efficiency: This is the big win. They estimated that their system uses about 23 microjoules of energy to generate one image.
- To put that in perspective: A standard digital computer might use 7 to 79 millijoules for the same task.
- That means the analog system is 100 times more energy-efficient (two orders of magnitude) than the digital baseline.

In Summary:
The paper shows that we can build a new kind of AI that runs on physical hardware (like oscillators and springs) instead of just digital chips. By accepting the hardware's limitations and using clever training methods (changing rules over time, adding hidden helpers, and focusing only on the final result), they created a system that draws images with a fraction of the energy cost of today's computers. It's a proof that nature's physics can be harnessed to do complex AI tasks efficiently.

Technical Summary: Generative Models on Analog Hardware with Dynamics

Problem Statement
Modern generative models, particularly those based on diffusion and flow matching, rely on flexible, software-defined dynamics implemented via neural networks. These models incur substantial energy costs due to iterative numerical integration on digital hardware. Conversely, analog hardware platforms (e.g., coupled oscillators, Analog Ising Machines) naturally solve differential equations with high energy efficiency by leveraging physical laws. However, a fundamental mismatch exists: analog hardware imposes fixed, physics-determined differential equations with limited approximation capacity, whereas modern generative models require flexible vector fields. This paper addresses the challenge of designing a generative model architecture that is amenable to implementation on analog hardware while overcoming the limitations of fixed dynamics, non-standard interaction kernels, and hardware constraints like noise and limited precision.

Methodology
The authors introduce Analog Interaction Systems (AIS), a unified framework for hardware-implementable dynamical systems. The core approach involves three key methodological pillars:

Architecture and Dynamics:
- AIS models are defined by ordinary differential equations (ODEs) where the state evolution is governed by physical interactions (e.g., sinusoidal coupling in oscillators, polynomial or sigmoidal nonlinearities in Ising machines).
- To bridge the expressivity gap between fixed physics and flexible neural networks, the authors propose two hardware-compatible mechanisms:
  - Time-piecewise parameters: Dividing the integration interval into segments with independent parameter sets to increase the complexity of the vector field.
  - Hidden physical states: Augmenting the visible state space with hidden physical elements that interact with the visible nodes, expanding the effective parameter space without requiring direct observation of the hidden trajectory.
- The hardware architecture is designed as a sparse grid of physical elements (PEs) with local and limited long-distance connections to minimize routing overhead. Parameters are quantized to low bit-widths to reduce DAC area and power.
Training Paradigm:
- The paper argues that exact trajectory supervision (e.g., Optimal Transport Conditional Flow Matching), which forces the model to follow a specific path at every time step, is infeasible for AIS due to structural constraints.
- Instead, the authors employ a Wasserstein Generative Adversarial Network (WGAN) training procedure. This method utilizes endpoint supervision, penalizing only the terminal visible state ( $v(T)$ ) rather than intermediate trajectories. This allows the physical system to self-organize and discover its own path from noise to data, optimizing over hidden states that lack ground-truth trajectories.
Hardware Implementation Strategy:
- The proposed architecture supports time-varying weights via three strategies: "Save-and-reprogram" (digitizing intermediate states), "Weight stationary" (spatial unrolling across multiple cores), and "State stationary" (dynamic weight updates within a single core).
- Power estimation is extrapolated from existing 28-nm coupled oscillator Ising machine silicon measurements, scaling for node count, coupling density, and weight precision.

Key Contributions

Unified Framework: Definition of Analog Interaction Systems (AIS) as a class of dynamical systems where computation emerges from device physics, distinct from digital emulations of neural ODEs.
Expressivity Mechanisms: Empirical identification and validation of time-piecewise parameters and hidden physical states as effective methods to narrow the expressivity gap between physics-constrained systems and neural network baselines.
Training Innovation: Development of a WGAN training procedure tailored for analog hardware, which decouples trajectory shape from target quality, enabling the optimization of systems with fixed interaction kernels.
Hardware-Aware Design: Proposal of a sparse, low-bit-width architecture with scaling laws demonstrating that sparse connectivity and quantization are necessary for practical energy efficiency.

Results

Expressivity Study: On low-dimensional synthetic distribution tasks (e.g., transforming Gaussian distributions to multi-modal mixtures), the combination of SWD (Sliced Wasserstein Distance) endpoint supervision, time-piecewise weights, and hidden states significantly reduced the Wasserstein distance ( $W_2^2$ ) compared to fixed-parameter or trajectory-supervised baselines. KuraSHIL (Kuramoto with injection-locking) emerged as the strongest performer.
Generative Quality: On MNIST and Fashion-MNIST datasets, the oscillator-based AIS achieved Fréchet Inception Distance (FID) scores of 27.6 and 80.8, respectively. This outperforms the best prior hardware-implementable analog generative models (DTM and NLM) by a factor of 3–4×.
Robustness: The model maintained high performance with 4-bit quantized weights (FID 27.2/81.9) and demonstrated resilience to transient noise levels ( $\delta=0.025$ ).
Energy Efficiency: The authors estimate an energy cost of 23 $\mu$ J per generated image for a 56×56 AIS core operating at 100 MHz. This represents a two-orders-of-magnitude improvement over digital baselines (estimated at 7–79 mJ per image). Scaling analysis indicates that moving to all-to-all connectivity or higher bit-widths would drastically increase energy consumption, validating the necessity of sparse, low-precision designs.

Significance
The paper claims that the limitations imposed by analog hardware—specifically fixed dynamics and physical constraints—can be overcome to yield expressive, energy-efficient generative models. By shifting the training objective from trajectory matching to endpoint distribution matching via WGANs, and by leveraging hidden states and time-varying parameters, the authors demonstrate that analog hardware can natively implement differential equation-based generative modeling. This work suggests a viable path toward low-power, low-latency generative AI that leverages the intrinsic physics of the substrate rather than emulating digital algorithms. The results indicate that with appropriate architectural and training adaptations, analog systems can achieve competitive generative quality while offering substantial energy savings over digital GPU-based alternatives.

Technical Summary: Generative Models on Analog Hardware with Dynamics

More like this