Title : Deep learning based on spiking neural networks and its acceleration
일시 : 2024년 11월 19일(화), 17시
Speaker : 정두석 Doo Seok Jeong (한양대학교 신소재공학부)
Abstract :
Deep learning (DL) is pervasive in diverse application domains in which instruction-based inference and decision making used to be used as key algorithms. DL uses various network models as per application domains, which include dense-layer based networks like multilayer perceptron and convolution-layer based networks like AlexNet, VGG, ResNet, etc. These are referred to as deep neural networks (DNNs) and commonly static models consisting of static (i.e., time-independent) activations, so that they are suitable for static input encoding as also static outputs. Unlike DNNs, spiking neural networks (SNNs) are a dynamic (time-dependent) graph whose dynamics is mainly endowed with spiking neurons as dynamic nodes and synaptic currents as dynamic inputs. Other defining features of SNNs compared with DNNs include (i) the use of 1-bit input data (spikes) that are encoded as also 1-bit output data (spikes) unlike DNNs using real-valued input and output data, and (ii) asynchronous event-based operations unlike layer-wise operating DNNs. Spiking neurons (with synaptic current models) encode input spikes and as output spikes using a preset code, e.g., rate code, spike-count code, and temporal code. Subsequently, the output spikes are delivered to the postsynaptic neurons and subject to encoding. SNN-based DL encounters challenges to (i) learning weights in the DL framework and (ii) prohibitive computational cost when using general-purpose graphics processing units (GPGPUs). Within the DL framework, SNNs are unrolled over discrete timesteps to approximate them to recurrent neural networks, and their weights are optimized using backpropagation through time (BPTT). Regarding (i), the main difficulty lies in the fact that the spike function (Boolean function) is not differentiable so that the conventional backprop pipeline cannot be applied. Regarding (ii), unrolling SNNs duplicates a given SNN network over all timesteps, so that its computational complexity scales with the number of timesteps, which causes prohibitive workloads on GPGPUs, particularly, when using long event-stream data. In my talk, I will address these challenges in detail and viable solutions to efficient SNN operations with particular emphasis on event-based digital neuromorphic hardware as a forward looking alternative to GPGPUs.