Exploiting Neuron and Synapse Filter Dynamics in Spatial Temporal Learning of Deep Spiking Neural Network (2003.02944v2)

Abstract: The recent discovered spatial-temporal information processing capability of bio-inspired Spiking neural networks (SNN) has enabled some interesting models and applications. However designing large-scale and high-performance model is yet a challenge due to the lack of robust training algorithms. A bio-plausible SNN model with spatial-temporal property is a complex dynamic system. Each synapse and neuron behave as filters capable of preserving temporal information. As such neuron dynamics and filter effects are ignored in existing training algorithms, the SNN downgrades into a memoryless system and loses the ability of temporal signal processing. Furthermore, spike timing plays an important role in information representation, but conventional rate-based spike coding models only consider spike trains statistically, and discard information carried by its temporal structures. To address the above issues, and exploit the temporal dynamics of SNNs, we formulate SNN as a network of infinite impulse response (IIR) filters with neuron nonlinearity. We proposed a training algorithm that is capable to learn spatial-temporal patterns by searching for the optimal synapse filter kernels and weights. The proposed model and training algorithm are applied to construct associative memories and classifiers for synthetic and public datasets including MNIST, NMNIST, DVS 128 etc.; and their accuracy outperforms state-of-art approaches.

• The paper introduces a novel representation of SNNs as IIR filters, capturing both neuron and synapse dynamics for enhanced temporal learning.
• The proposed training algorithm optimizes synaptic weights and impulse response kernels, outperforming state-of-the-art models on MNIST and DVS128 Gesture.
• The work demonstrates practical, energy-efficient spiking neural network applications and paves the way for scalable neuromorphic computing solutions.

Exploiting Neuron and Synapse Filter Dynamics in Spatial Temporal Learning of Deep Spiking Neural Networks

The paper by Fang et al. addresses the challenge of effectively training large-scale spiking neural networks (SNNs) by formulating them as systems of infinite impulse response (IIR) filters, leveraging the inherent temporal dynamics of neurons and synapses. This work is fundamentally centered on the potential of SNNs, which have drawn attention due to their biological resemblance and capability for energy-efficient event-driven operation. However, the employment of SNNs in high-performance tasks is hampered by the absence of robust training algorithms that can capitalize on their spatial-temporal processing strengths.

Core Contributions

The authors make several notable contributions:

1. IIR Filter Representation: The paper presents a novel way of representing SNNs as networks of IIR filters, thereby encapsulating both neuron and synapse dynamics, which are often overlooked in conventional models. This formulation aids in capturing the dependencies on historical input spikes, a vital requirement for temporal signal processing.
2. Training Algorithm: A training algorithm that incorporates both spatial and temporal information is proposed. This algorithm is designed to optimize not only synaptic weights but also the impulse response kernels of synapse filters, which the authors argue is akin to learning behaviors observed in biological systems.
3. Experiments and Results: The effectiveness of the proposed model and algorithm is demonstrated across various applications, including associative memory tasks, vision classification tasks such as MNIST and DVS128 Gesture, and temporal pattern classification of datasets like TIDIGITS. The results indicate that the proposed method outperforms state-of-the-art approaches, showcasing the capability of the model to utilize both spike rates and the temporal patterns encoded in spike trains.

Numerical Results and Implications

The paper reports strong numerical results, especially in vision-related tasks. For example, the model achieves an impressive accuracy of 99.46% on the MNIST dataset and 96.09% on the DVS 128 Gesture dataset, surpassing existing methodologies in the spiking domain. These outcomes underscore the potential of the SNN model as a viable alternative to traditional deep neural network (DNN) models, particularly in domains where temporal dynamics play a crucial role.

Implications and Future Directions

The implications of this work are multi-faceted. Practically, the authors' approach provides a pathway for deploying SNNs in real-world applications where energy efficiency and temporal dynamics are critical. Theoretically, the articulation of SNNs as IIR filters offers a robust framework that could be expanded to incorporate more complex neuron models, potentially leading to advances in understanding how biological neural systems process information.

Future developments in this area may involve exploring the scalability of the proposed model and algorithm to even larger networks, thus broadening their applicability to more complex datasets. Moreover, integrating additional types of neuromorphic hardware could harness the energy-efficient properties of SNNs in practical applications. Additionally, further investigation into the biological plausibility of synaptic plasticity mechanisms within this framework might yield deeper insights into learning in neural systems.

In conclusion, this paper makes significant strides in advancing SNN training methodologies by embedding temporal dynamics through IIR filter representations, providing a robust tool for both research and practical applications in neuromorphic computing.