Abstract: Event-based eye-tracking systems are revolutionizing fields like consumer electronics and neuroscience. Human eye movements can exceed speeds of 300°/s, necessitating high-speed event cameras for accurate tracking. In consumer electronics, particularly AR/VR, these systems reduce power consumption through sparse data streams, enabling lighter, more efficient headsets with extended usage and enhanced comfort. This advancement significantly improves the immersive AR/VR experience and the capabilities of portable technology. In neuroscience, event-based eye tracking aids in understanding visual attention and diagnosing neurological disorders. This talk will explore developing an event-based eye-tracking system for precise rapid eye movement tracking, promising lighter devices and deeper insights into neuroscience and cognitive research.

Abstract: A major aim in neuromorphic computing is to develop physical systems that learn. Unfortunately, the ubiquitous backpropagation (BP) algorithm is suboptimal given the need to propagate exact gradients throughout the system. In this talk, I present an alternative approach which learns through random perturbations in combination with input decorrelation. This approach is local in nature and ideally suited for implementation on noisy physical devices. We show that input decorrelation allows faster training of networks compared to BP. Furthermore, we show that learning based on random perturbations approaches the performance of BP. In our more recent work we focus on porting this approach to FPGAs and other physical devices with the aim of enabling efficient and effective on-chip learning on neuromorphic devices.

Abstract: Resistive memory devices offer outstanding potential for near- and in-memory artificial intelligence (AI) accelerators due to their nonvolatility, high density, and multilevel capacity. However, these devices also present significant challenges due to their high variability and unpredictability, essentially rendering them functionally similar to random variables. This feature presents an intriguing parallel with machine learning, where Bayesian methodologies are expressly designed to work with random variables, delivering the added advantage of being able to quantify prediction uncertainty—a hurdle that conventional artificial intelligence techniques often struggle to overcome. In this study, we propose that these Bayesian methods could serve as an effective strategy to harness the potential of resistive memory devices, without the accompanying drawbacks. Our exploration introduces three distinct approaches that increasingly integrate the random nature of resistive memories. Each methodology is substantiated through practical implementation using circuits fabricated in a hybrid CMOS/hafnium-oxide memristor process. We first introduce Bayesian machines, which leverage near-memory computing for energy-efficient Bayesian reasoning, enabling explainable decision-making under uncertain conditions by using all available data and knowledge. These machines digitally encode parameters in resistive memory, minimizing energy costs associated with data movement. Two designs were realized using a 130-nm hybrid CMOS/HfOx memristor process, featuring stochastic computing and integer logarithmic computation for efficient processing, both demonstrating remarkable energy efficiency. Then, we show the realization of Bayesian neural networks (BNNs), which treat synapses as random variables to model uncertainty in predictions, and utilize resistive memories to simulate the random behavior of synapses. This approach, validated through the programming of 50 in-memory neural networks for arrhythmia detection, allows the networks to quantify prediction certainty by leveraging the inherent randomness of memristors. The performance of these BNNs remains stable over time, incorporating resistive memory fluctuations into their uncertainty modeling. Finally, we introduce the Metropolis-Hastings Markov Chain Monte Carlo (MCMC) technique, tailored to exploit the random nature of memristors for learning, enabling synaptic weights to be sampled in a way that mimics natural Gaussian distributions [4]. This method was applied to train experimentally a 16k array of resistive memories for cancerous tissue identification, achieving a 98% accuracy rate, comparable to traditional software-based methods, thereby illustrating the potential of memristors in advanced Bayesian learning applications.

Abstract: Currently, neural-network processing in machine learning applications relies on layer synchronization, whereby neurons in a layer aggregate incoming currents from all neurons in the preceding layer, before evaluating their activation function. This is practiced even in artificial Spiking Neural Networks (SNNs), which are touted as consistent with neurobiology, in spite of processing in the brain being, in fact asynchronous. A truly asynchronous system however would allow all neurons to evaluate concurrently their threshold and emit spikes upon receiving any presynaptic current. Omitting layer synchronization is potentially beneficial, for latency and energy efficiency, but asynchronous execution of models previously trained with layer synchronization may entail a mismatch in network dynamics and performance. We present a study that documents and quantifies this problem in three datasets on our simulation environment that implements network asynchrony, and we show that models trained with layer synchronization either perform sub-optimally in absence of the synchronization, or they will fail to benefit from any energy and latency reduction, when such a mechanism is in place. We then “make ends meet” and address the problem with unlayered backprop, a novel backpropagation-based training method, for learning models suitable for asynchronous processing. We train with it models that use different neuron exe-cution scheduling strategies, and we show that although their neurons are more reactive, these models consistently exhibit lower overall spike density, reach a correct decision faster without integrating all spikes, and achieve superior accuracy. Our findings suggest that asynchronous event-based (neuromorphic) AI computing is indeed more efficient, but we need to seriously rethink how we train our SNN models, to benefit from it.