No-backprop SNN scores 98.2% on Split-MNIST task-incremental, age 14 (github.com)

A 14-year-old developer named Gustav has created an impressive spiking neural network (SNN) that achieves 98.2% accuracy on the Split-MNIST task-incremental benchmark, notably without using backpropagation. This innovative model demonstrates a robust ability to learn new tasks while retaining knowledge of previously learned tasks, a significant challenge in the field of continual learning. While traditional models like standard multi-layer perceptrons and DeepMind's Elastic Weight Consolidation rely on backpropagation and often forget past information, Gustav's SNN utilizes spike-timing-dependent plasticity (STDP) and Hebbian learning principles to mirror the functions of biological brains, promoting adaptability and efficiency. This development is significant for the AI/ML community as it highlights the potential of bio-inspired learning approaches that can operate on standard hardware like laptop CPUs, in stark contrast to contemporary models that demand substantial computational resources. The architecture includes task-specific output heads while maintaining a shared hidden layer, enhancing its efficiency. By focusing on mechanisms like anticipatory firing and emergent curiosity, the SNN not only learns efficiently but also seeks to understand new stimuli, suggesting a paradigm shift in how AI can be developed and deployed. Gustav's work opens avenues for implementing similar SNNs on neuromorphic hardware, promising lower energy costs compared to traditional deep learning architectures.