Training and optimization of hardware neural networks

Summary

Machine learning and artificial intelligence (AI) have advanced so rapidly that they can now outperform humans at many (or even most) tasks. However, the large models that perform the most complex tasks use an enormous amount of energy, in particular during the learning or training phase. However, it is also clear that despite the rapid progress in AI performance, there is still something fundamental that we don’t understand about how these tasks are done natively in the brain with such low energy on noisy, variable, biological components, and with a level of adaptability to new circumstances that has not been reproduced in artificial systems. This gap in our knowledge is particularly clear in new, emergent bio-inspired hardware being developed for AI in the hopes of reproducing this robust, energy-efficient operation. While large arrays of such devices have been built, their utility and scaling has been limited by the inability to simulate and program them in the same way that digital hardware can be modeled and simulated. This has spawned a variety of training procedures tuned to particular hardware platforms, and has generally limited the size and scope of the emerging hardware. The goal of this project is to develop and demonstrate a general training technique that can be natively implemented on a variety of hardware neural networks, from feedforward crossbar arrays to recurrent physical networks to spiking neuromorphic hardware.

Description

The goal of this project is to develop a general method that can train many different types of neural networks, and to demonstrate and evaluate their performance on new emerging hardware. We aim to develop and demonstrate training on diverse hardware platforms, and in the presence of realistic noise and device-to-device variations.

Major Accomplishments

Research Highlights

The most promising technique we are investigating measures the gradient rather than calculating it via backpropagation, and therefore is robust against noise and fabrication imperfections that can provide a challenge to analog or emerging hardware. The technique is known as multiplexed gradient descent (MGD) and is a model-free zero-order optimization method. In our recent paper “Multiplexed gradient descent: Fast online training of modern datasets on hardware neural networks without backpropagation”, we demonstrate via simulations that the speed of training on emergent hardware with MGD is competitive with backpropagation on a typical GPU or CPU, while being robust against the noise and fabrication imperfections typical of these hardware platforms. We have also been working to speed up the training process by using more of the known architectural information in a variation known as “node perturbation”. In future work, we aim to extend the MGD framework to recurrent and spiking networks, and link it to the reinforcement learning paradigm.

We have been working with collaborators to implement MGD and train diverse hardware platforms. These collaborations span a variety of different hardware platforms, from photonic hardware (U. Queens) to memristive hardware (RIT, NIST) to FPGA hardware (RIT). The promise and context for this approach in photonic neural computing was described in our paper “Photonic Online Learning: A Perspective”. In an earlier collaboration with University of Tennessee and Oak Ridge National Laboratory we used a different approach that allowed us to simulate spiking networks of superconducting optoelectronic neurons designed by evolutionary optimization. This work was published in our paper “Design of Superconducting Optoelectronic Networks for Neuromorphic Computing”.

One of the major difficulties in the field of neuromorphic computing in particular is how to evaluate and compare the performance of the diverse array of hardware systems being developed. As such, we are interested in developing standard benchmark datasets and metrics for testing. As part of this project, we are participating in a collaboration across industry and academia to try to develop new benchmarks for neuromorphic computing. The whitepaper for this work has been published and will be updated on the website neurobench.ai.

Opportunities

We currently have opportunities for postdocs, undergraduate researchers and visiting researchers. Please contact Sonia Buckley (sonia.buckley [at] nist.gov (sonia[dot]buckley[at]nist[dot]gov)) if you are interested in these opportunities.