Biologically Plausible Training of Deep Neural Networks Using a Top-down Credit Assignment Network

TL;DR

This paper introduces a biologically inspired top-down credit assignment network that replaces backpropagation, demonstrating improved training efficiency and performance across various neural network tasks and architectures.

Contribution

The study proposes a novel two-level training framework using a Top-Down Credit Assignment Network with a brain-inspired credit diffusion mechanism, enhancing biological plausibility and training speed.

Findings

Outperforms backpropagation in non-convex optimization, supervised, and reinforcement learning tasks.

Effectively bypasses local minima in the loss landscape.

Shows strong generalization across datasets and architectures.

Abstract

Despite the widespread adoption of Backpropagation algorithm-based Deep Neural Networks, the biological infeasibility of the BP algorithm could potentially limit the evolution of new DNN models. To find a biologically plausible algorithm to replace BP, we focus on the top-down mechanism inherent in the biological brain. Although top-down connections in the biological brain play crucial roles in high-level cognitive functions, their application to neural network learning remains unclear. This study proposes a two-level training framework designed to train a bottom-up network using a Top-Down Credit Assignment Network (TDCA-network). The TDCA-network serves as a substitute for the conventional loss function and the back-propagation algorithm, widely used in neural network training. We further introduce a brain-inspired credit diffusion mechanism, significantly reducing the TDCA-network's parameter complexity, thereby greatly accelerating training without compromising the network's performance.Our experiments involving non-convex function optimization, supervised learning, and reinforcement learning reveal that a well-trained TDCA-network outperforms back-propagation across various settings. The visualization of the update trajectories in the loss landscape indicates the TDCA-network's ability to bypass local minima where BP-based trajectories typically become trapped. The TDCA-network also excels in multi-task optimization, demonstrating robust generalizability across different datasets in supervised learning and unseen task settings in reinforcement learning. Moreover, the results indicate that the TDCA-network holds promising potential to train neural networks across diverse architectures.