Synthesizing Hebbian and Backpropagation Paradigms: Hybrid Learning for Interpretable Representations

There's a clear divide in the literature between biologically-inspired learning (e.g., Hebbian plasticity in spiking neural networks [Ravichandran et al., 2023]) and the backpropagation-dominated deep learning paradigm (Dampfhoffer et al., 2023). This idea proposes a hybrid framework: use Hebbian-style local updates for the early layers of a network (where feature extraction is critical and interpretability is desirable), then couple this with gradient-based learning in deeper layers for task-specific optimization. Such a model could be evaluated on benchmarks like MNIST and F-MNIST to compare representation quality and interpretability against purely backpropagation-based or Hebbian-based models. This approach builds a bridge between neuroscience and machine learning, potentially resulting in representations that are both more robust and easier to interpret. It also directly addresses the critique that backpropagation is biologically implausible, offering a new theoretical model that aligns more closely with how learning may occur in the brain.

References:

1. Spiking neural networks with Hebbian plasticity for unsupervised representation learning. Naresh B. Ravichandran, A. Lansner, P. Herman (2023). The European Symposium on Artificial Neural Networks.
2. Backpropagation-Based Learning Techniques for Deep Spiking Neural Networks: A Survey. Manon Dampfhoffer, T. Mesquida, A. Valentian, L. Anghel (2023). IEEE Transactions on Neural Networks and Learning Systems.