Spiking LSTM on Loihi 2: A Neuromorphic Reinterpretation of Recurrent Networks

In recent years, the presence of artificial intelligence (AI) has become increasingly pervasive, with artificial neural networks (ANNs) being applied across a growing number of domains. These models, while powerful, are also becoming larger and more computationally demanding. In addition, an interest has emerged in understanding and emulating the remarkable efficiency and compactness of the human brain. This has led to the rise of neuromorphic computing, which aims to design lightweight and energy-efficient systems inspired by biological neurons, through spiking neural networks (SNNs). By combining ideas from neuroscience and machine learning, neuromorphic computing offers a way to reinterpret traditional AI models in a more biologically grounded way. Among the various approaches to bridging the gap between classical and brain-inspired computation, revisiting well-established ANN architectures within a neuromorphic framework is a viable direction. This work focuses on translating the Long Short-Term Memory (LSTM) network, a widely used recurrent architecture known for its ability to capture long-term dependencies and handle sequential data, into a spiking form, where operations are rephrased in terms of neuron populations and synapses. The goal of this reinterpretation is to make the architecture compatible with neuromorphic hardware, enabling efficient execution on brain-inspired systems. Several spiking LSTM (sLSTM) variants were explored, and the final design replaces conventional activation functions with the spiking dynamic of Leaky Integrate-and-Fire (LIF) neurons, leveraging membrane potential values for the computation of the internal states. The architecture was implemented on two frameworks: snnTorch, used for training and hyperparameter optimization via the Neural Network Intelligence (NNI) framework and NxKernel, Intel's proprietary framework for deployment on the Loihi 2 neuromorphic board. The transition between the two frameworks required the implementation of custom neurons in microcode to reproduce the behavior observed in snnTorch. Additional challenges arose from the fixed-point arithmetic used in Loihi 2’s synapses and neuron models, which demanded careful quantization and scaling strategies. Moreover, Loihi 2’s pipelined execution introduced differences in layer synchronization compared to the software-based simulation, requiring further adaptations to preserve consistent network dynamics across frameworks. Using the profiling tools available in the NxKernel framework, the model performance on the Loihi 2 hardware was evaluated, taking into account the effects of network sparsity and also multiple partitioning and mapping configurations obtained through a heuristic optimization algorithm. For initial experiments, a Human Activity Recognition (HAR) task was used, employing a spike-encoded dataset with six input channels and seven output classes. Once a robust pipeline for training, weight transfer and hardware deployment was established, the architecture was further tested on the Spiking Heidelberg Digits (SHD) dataset, which involves classifying spoken digits. Without any preprocessing, the model was trained on snnTorch, achieving competitive test accuracies up to 90%. When deployed on Loihi 2, it exhibited a slight accuracy drop, yet the overall results highlight the potential of this neuromorphic reinterpretation of LSTM networks.