Deepity – A C++/CUDA framework replacing backprop with predictive coding (github.com)

Deepity has introduced a novel C++/CUDA framework for training neural networks using predictive coding, a brain-inspired learning method that replaces traditional backpropagation. This new framework features a user-friendly Python interface and has achieved impressive results, reaching 98.21% accuracy on the MNIST dataset in under 500 seconds on a standard laptop CPU. Deepity's SimplePCN outperforms existing predictive coding libraries, such as jpc (JAX) and ngc-learn, in both speed and accuracy, highlighting its potential as a competitive alternative for neural network training. The significance of Deepity lies in its departure from backpropagation, where neurons rely on a global error signal. Instead, each layer in a predictive coding network adjusts its weights based on local prediction errors, which aligns more closely with biological learning processes. This framework not only reduces memory requirements—eliminating the need for storing the backward computation graph—but can also lead to faster training times, especially with techniques like Direct Kolen-Pollack (DKP). As interest in models that mimic human learning increases, Deepity's approach positions it well for future developments in AI and could be particularly beneficial for neuromorphic and edge computing architectures.