D-SELD: Dataset-Scalable Exemplar LCA-Decoder

TL;DR

D-SELD introduces a Dataset-Scalable Exemplar LCA-Decoder that leverages sparse coding and the Locally Competitive Algorithm to train Spiking Neural Networks on neuromorphic hardware without gradient-based backpropagation. By building a dictionary directly from dataset features extracted by CNN backbones, the method enables scalable decoding that can incorporate new data online and reduce memory and compute demands. Empirically, D-SELD achieves state-of-the-art top-1 accuracy on ImageNet (80.75%) and CIFAR-100 (79.32%) with SNNs and demonstrates highly sparse neuron activity, favorable FLOP profiles, and robust reconstruction performance. The work offers a practical pathway for deploying large-scale SNNs on neuromorphic chips, balancing accuracy, efficiency, and scalability for dynamic datasets.

Abstract

Neuromorphic computing has recently gained significant attention as a promising approach for developing energy-efficient, massively parallel computing systems inspired by the spiking behavior of the human brain and natively mapping Spiking Neural Networks (SNNs). Effective training algorithms for SNNs are imperative for increased adoption of neuromorphic platforms; however, SNN training continues to lag behind advances in other classes of ANN. In this paper, we reduce this gap by proposing an innovative encoder-decoder technique that leverages sparse coding and the Locally Competitive Algorithm (LCA) to provide an algorithm specifically designed for neuromorphic platforms. Using our proposed Dataset-Scalable Exemplar LCA-Decoder we reduce the computational demands and memory requirements associated with training SNNs using error backpropagation methods on increasingly larger training sets. We offer a solution that can be scalably applied to datasets of any size. Our results show the highest reported top-1 test accuracy using SNNs on the ImageNet and CIFAR100 datasets, surpassing previous benchmarks. Specifically, we achieved a record top-1 accuracy of 80.75% on ImageNet (ILSVRC2012 validation set) and 79.32% on CIFAR100 using SNNs.

Figures (7)

• Figure 1: Dataset-Scalable Exemplar LCA-Decoder architecture. The features ($\phi_k$) are extracted once and stored on neuromorphic platforms. The spiking neurons have membrane potential $u_k$ and activation $a_k$, and S is the input vector.
• Figure 2: Trade-off between accuracy and workload across CNN models used as feature extractors in D-SELD
• Figure 3: The reconstructed images of an automobile from the CIFAR-10 dataset. The "Dataset" dictionary performs better in reconstructing the input image. Input image is sourced from unseen test dataset.
• Figure 4: Energy function $E$ during image reconstructions (a and b).
• Figure 5: Top-1 Accuracy comparison across three decoders with increasing dictionary over-completeness on MNIST dataset (image size: 28 * 28). Here, #2 represents dictionary size of 28 * 28 * 2 and "Complete" denotes the utilization of all training data points.