High-accuracy deep ANN-to-SNN conversion using quantization-aware training framework and calcium-gated bipolar leaky integrate and fire neuron

Abstract

Spiking neural networks (SNNs) have attracted intensive attention due to the efficient event-driven computing paradigm. Among SNN training methods, the ANN-to-SNN conversion is usually regarded to achieve state-of-the-art recognition accuracies. However, many existing ANN-to-SNN techniques impose lengthy post-conversion steps like threshold balancing and weight renormalization, to compensate for the inherent behavioral discrepancy between artificial and spiking neurons. In addition, they require a long temporal window to encode and process as many spikes as possible to better approximate the real-valued ANN neurons, leading to a high inference latency. To overcome these challenges, we propose a calcium-gated bipolar leaky integrate and fire (Ca-LIF) spiking neuron model to better approximate the functions of the ReLU neurons widely adopted in ANNs. We also propose a quantization-aware training (QAT)-based framework leveraging an off-the-shelf QAT toolkit for easy ANN-to-SNN conversion, which directly exports the learned ANN weights to SNNs requiring no post-conversion processing. We benchmarked our method on typical deep network structures with varying time-step lengths from 8 to 128. Compared to other research, our converted SNNs reported competitively high-accuracy performance, while enjoying relatively short inference time steps.

Learnable Leakage and Onset-Spiking Self-Attention in SNNs with Local Error Signals

“…59 Fang et al (2021) [62] 128c3-p2-128c3-p2-2048-100-10 8 99.60 FELL (2023) [63] 128c3-p2-128c3-p2-2048-100-10 10 99.38 BELL (2023) [63] 128c3-p2-128c3-p2-2048-100-10 10 99.35 ELL (2023) [63] 128c3-p2-128c3-p2-2048-100-10 10 99. 46 Gao et al (2023) [64] VGG-9 128 99.62 Ours 128c3-p2-128c3-p2-2048-100-10 8 99.67…”

“…Ma et al’s three methods [ 63 ], FELL, BELL, and ELL, all emphasized spike learning based on local classifiers, indicating the effectiveness of local learning in deep networks. Gao et al [ 64 ] adopted the VGG-9 structure and employed a quantized training framework for the conversion from deep ANNs to SNNs. While this approach is technically noteworthy, it required significantly more time steps than our method.…”

“… The visual comparison of different methods [ 19 , 21 , 22 , 39 , 47 , 57 , 58 , 59 , 60 , 61 , 62 , 63 , 64 , 65 , 66 , 67 , 68 , 69 ] cross three datasets. Within each dataset, darker colors of the bar graphs indicate higher accuracy.…”

Deep unsupervised learning using spike-timing-dependent plasticity

“…Note that the CNN layers used for feature-reduction purposes can be converted into SNN layers with various methods as shown in many recent studies [7,[38][39][40][41], or trained using backpropagation through time [42,43] which opens up the potential for adopting the entire feature-reduction in a low-power neuromorphic setting. In this work, we focus on the CNN-SNN conversion method and train it by backpropagation, without unrolling through time, considering a rate-coded network.…”

Incorporating structural plasticity into self-organization recurrent networks for sequence learning

“…Therefore, many researches combine various mathematical optimization techniques with spiking neural networks, trying to propose a new learning paradigm suitable for SNNs. For example, Xing et al adopted the ANN-to-SNN strategy to migrate the parameters of the trained ANNs to the SNNs of the same architecture (Xing et al, 2019 ; Gao et al, 2023 ). Anwar et al applied reinforcement learning to spiking neural networks to perform specific tasks, such as Pong and Cartpole game playing (Bellec et al, 2020 ; Anwar et al, 2022 ; Haşegan et al, 2022 ).…”