Andreas Andreou – Invited Special Session
B1L-C: Andreas Andreou - Invited Special SessionSession Type: Lecture
Session Code: B1L-C
Location: Room 3
Date & Time: Thursday March 23, 2023 (09:00 - 10:00)
Chair: Andreas Andreou
|Architecture Analysis for Symmetric Simplicial Deep Neural Networks on Chip
|Nicolás Rodríguez, Martín Villemur, Pedro Julián
|Convolutional Neural Networks (CNN) are the dom- inating Machine Learning architecture used for complex tasks such as image classification at the expense of heavy computational resources, large storage space and power-demanding hardware. This motivates the exploration of alternative implementations moving the focus to the design and implementation of efficient neuromorphic hardware for resource constrained applications. Conventional Simplicial Piece-Wise Linear implementations allow the development of efficient hardware to run DNNs by avoiding multipliers, but at a cost of large memory requirements. In particular, Symmetric Simplicial (SymSim) functions preserve the efficiency of the implementation while reducing the number of parameters per layer, and can be trained to replace convolutional layers and native run non-linear filters such as MaxPool. This paper analyzes architectures to implement a Neural Net- work accelerator for SymSim operations optimizing the number of parallel cores to reduce the computational time. For this, we develop a model that takes into account the core processing times as well as the data transfer times.
|Inducing Dynamic Group Sparsity on Vagus Nerve Recordings
|Khaled Aboumerhi, Ralph Etienne-Cummings
|As recording electrodes improve with higher spatial resolution, the large amount of incoming data is becoming difficult to handle on low-resource medical technologies, such as implantable devices. Processing each data sample is costly in terms of energy, memory allocation, bandwidth transmission and more. Accurate neural compression has thus become important in reducing the amount of data that needs to be processed, whether to an implantable device or a distant computer. Ideally, neuromorphic hardware would compress much of this data to handle future processing bottlenecks. Preprocessing the data using compression algorithms, such as EI balance networks, have been successfully implemented, but rely on the physical structure of neurons to have many connections, such as in the cortex. This structure does not exist in the peripheral nervous system and so cannot be accurately deployed, and extremely populated data can still be overwhelming for neuromorphic chips. In this paper, we employ dynamic group sparsity (DGS) on two different nerve recording datasets to demonstrate intelligent compression schemes while retaining signal representation as a preprocessing component for neuromorphic hardware. DGS does not require structural connections, but instead assumes that neurons fire in groups. We assume that there is structure among neuron groups that fire sparsely, both spatially and temporally. We show that under certain conditions, DGS achieves 5x compression while retaining 99.2% signal integrity. We then compare these results to typical matching pursuit algorithms, such as OMP and CoSaMP, and conclude with future implementations of DGS in post-acquisition nerve recordings on event-based devices.
|Retinomorphic Channel Design and Considerations
|Jonah Sengupta, Andreas Andreou
|By extrapolating functionality from the retina, retinomorphic engineering has yielded devices that have shown promise to alleviate the challenges presented in modern computer vision tasks. An incredible amount of work has been devoted in recent years to the development and deployment of these event-based vision sensors in applications requiring low-latency, energy-efficient, high dynamic range sensing solutions. However, not much work has been devoted to the analysis of the various encoding and decoding mechanisms necessary for sensory pipelines that are optimized with respect to area, energy, and speed. This paper outlines an empirical framework that presents a clear tradeoff between the various methodologies to transduce physical information in to spikes (encoding) and reconstruct said stimuli from the incident events (decoding). Software-based models of these methodologies were constructed to evaluate the accuracy of stimuli reconstruction for a variety of input profiles. As a result, it is shown that an optimized retinomorphic architecture for a specific set of system-driven cost metrics requires a heterogenous fabric of encoders and decoders. Much like the composition of ganglion cells in magno- and parvo-cellular pathway, this multi-modal solution provides the most time, area, and power efficient method to convey visual data.