Counter-Current Learning: A Biologically Plausible Dual Network Approach for Deep Learning

TL;DR

Inspired by the counter-current exchange mechanisms observed in biological systems, counter-current learning (CCL) is proposed, a biologically plausible framework for credit assignment in neural networks and achieves comparable performance to other biologically plausible algorithms while offering a more biologically realistic learning mechanism.

Abstract

Despite its widespread use in neural networks, error backpropagation has faced criticism for its lack of biological plausibility, suffering from issues such as the backward locking problem and the weight transport problem. These limitations have motivated researchers to explore more biologically plausible learning algorithms that could potentially shed light on how biological neural systems adapt and learn. Inspired by the counter-current exchange mechanisms observed in biological systems, we propose counter-current learning (CCL), a biologically plausible framework for credit assignment in neural networks. This framework employs a feedforward network to process input data and a feedback network to process targets, with each network enhancing the other through anti-parallel signal propagation. By leveraging the more informative signals from the bottom layer of the feedback network to guide the updates of the top layer of the feedforward network and vice versa, CCL enables the simultaneous transformation of source inputs to target outputs and the dynamic mutual influence of these transformations. Experimental results on MNIST, FashionMNIST, CIFAR10, and CIFAR100 datasets using multi-layer perceptrons and convolutional neural networks demonstrate that CCL achieves comparable performance to other biologically plausible algorithms while offering a more biologically realistic learning mechanism. Furthermore, we showcase the applicability of our approach to an autoencoder task, underscoring its potential for unsupervised representation learning. Our work presents a direction for biologically inspired and plausible learning algorithms, offering an alternative mechanism of learning and adaptation in neural networks.

Figures (9)

• Figure 1: Overview of the Counter-Current Learning Framework: (a) At initialization, the counter-current learning framework establishes a dual network structure, with a forward network that maps the input to the target output, and a complementary feedback network that mirrors the forward network's architecture but propagates information in the opposite direction. The framework leverages the data processing inequality (DPI) from information theory, which states that information content cannot be increased through signal processing. Consequently, in both networks, information content decreases from the lower to the upper layers. (b) During training, the losses are computed in a layer-wise manner, i.e., by calculating the difference of activations from corresponding layer pairs between the forward and feedback networks, allowing the networks to learn from each other's complementary information. Notably, the dependency of the gradient on earlier layer parameters is interrupted using the gradient detachment operator.
• Figure 2: Code Snippet For Counter-Current Learning With Dual Network Architecture.
• Figure 3: Dynamic Feature Alignment Between Forward and Backward Models During Counter-Current Learning. This series of t-SNE plots demonstrates the evolution of feature space alignment over different stages of training. Circular dots represent features from the forward network processing MNIST images, while squares depict features from the feedback network handling one-hot encoded labels. Each color represents a distinct class, with every subplot providing an independent t-SNE visualization. This emphasizes how distinct classes increasingly converge within and across the forward and backward models as training progresses, highlighting the dynamic and reciprocal nature of learning within the counter-current framework.
• Figure 4: Visualization of the First Layer Convolutional Kernels of the Forward Model Trained with Error Backpropagation (BP) and Counter-Current Learning (CCL). Kernels from models trained with BP have more high-frequency components, as manifested as neighboring white (e.g., weight with high values) and black pixels (e.g., weight with low values). In comparison, those with CCL have more low-frequency components. We posit this might be because the error signal can contain more high-frequency information than the ideal target signal.
• Figure 5: Qualitative Comparison of an Eight-Layered Convolutional Autoencoder Trained Using Error Backpropagation (BP) and Counter-Current Learning (CCL). The network structure does not contain skip connections. Testing set reconstruction results highlight CCL's comparable reconstruction as BP while achieving biological plausibility.