CATFormer: When Continual Learning Meets Spiking Transformers With Dynamic Thresholds

Although deep neural networks perform extremely well in controlled environments, they fail in real-world scenarios where data isn't available all at once, and the model must adapt to a new data distribution that may or may not follow the initial distribution. Previously acquired knowledge is lost during subsequent updates based on new data. a phenomenon commonly known as catastrophic forgetting. In contrast, the brain can learn without such catastrophic forgetting, irrespective of the number of tasks it encounters. Existing spiking neural networks (SNNs) for class-incremental learning (CIL) suffer a sharp performance drop as tasks accumulate. We here introduce CATFormer (Context Adaptive Threshold Transformer), a scalable framework that overcomes this limitation. We observe that the key to preventing forgetting in SNNs lies not only in synaptic plasticity but also in modulating neuronal excitability. At the core of CATFormer is the Dynamic Threshold Leaky Integrate-and-Fire (DTLIF) neuron model, which leverages context-adaptive thresholds as the primary mechanism for knowledge retention. This is paired with a Gated Dynamic Head Selection (G-DHS) mechanism for task-agnostic inference. Extensive evaluation on both static (CIFAR-10/100/Tiny-ImageNet) and neuromorphic (CIFAR10-DVS/SHD) datasets reveals that CATFormer outperforms existing rehearsal-free CIL algorithms across various task splits, establishing it as an ideal architecture for energy-efficient, true-class incremental learning.

SHaRe-SSM: An Oscillatory Spiking Neural Network for Target Variable Modeling in Long Sequences

In recent years, with the emergence of large models, there has been a significant interest in spiking neural networks (SNNs) primarily due to their energy efficiency, multiplication-free, and sparse event-based deep learning. Similarly, state space models (SSMs) in varying designs have evolved as a powerful alternative to transformers for target modeling in long sequences, thereby overcoming the quadratic dependence on sequence length of a transformer. Inspired by this progress, we here design SHaRe-SSM (Spiking Harmonic Resonate and Fire State Space Model), for target variable modeling (including both classification and regression) for very-long-range sequences. Our second-order spiking SSM, on average, performs better than transformers or first-order SSMs while circumventing multiplication operations, making it ideal for resource-constrained applications. The proposed block consumes $73 \times$ less energy than second-order ANN-based SSMs for an 18k sequence, while retaining performance. To ensure learnability over the long-range sequences, we propose exploiting the stable and efficient implementation of the dynamical system using parallel scans. Moreover, for the first time, we propose a kernel-based spiking regressor using resonate and fire neurons for very long-range sequences. Our network shows superior performance on even a 50k sequence while being significantly energy-efficient. In addition, we conducted a systematic analysis of the impact of heterogeneity, dissipation, and conservation in resonate-and-fire SSMs.