Modern artificial intelligence is built on backpropagation, an algorithm that calculates how to change every weight in a network to reduce error. While backpropagation is remarkably effective, it faces a fundamental "biological implausibility" problem. Biological brains achieve similar or superior learning feats without the specific mathematical requirements that backpropagation demands.
Researchers are currently bridging this gap by developing learning rules that respect the physical and local constraints of biological neurons while maintaining the power of deep learning.
The Weight Transport Problem
The most significant hurdle for biological backpropagation is the weight transport problem. In standard gradient descent, the backward pass must use the exact transpose of the forward weight matrix ($W^T$) to propagate error signals.
As Hongyin Luo and James R. Glass explain in Bidirectional Backpropagation, biological synapses are unidirectional. There is no known mechanism for a "backward" synapse to perfectly mirror the strength of a "forward" synapse across different layers. This physical routing constraint makes scaling decentralized architectures, like neuromorphic hardware, incredibly difficult (Augustine et al.).
Feedback Alignment: Learning to Use Random Signals
A surprising solution to the weight transport problem is Feedback Alignment (FA). In Direct Feedback Alignment Provides Learning in Deep Neural Networks, Arild Nøkland demonstrated that the backward weights do not need to be the transpose of the forward weights. Instead, they can be fixed, random matrices.
The intuition behind why this works is "alignment": during training, the forward weights adapt so that they eventually align with the fixed random feedback weights. The network essentially learns how to make the random error signals useful. Pierre Baldi and colleagues further showed in Learning in the Machine that networks are remarkably robust to these "noisy" feedback signals, even when they are sparse or low-precision.
Predictive Coding and Local Credit Assignment
Another major direction is Predictive Coding (PC), which reframes learning as a local process of minimizing "prediction errors." Instead of a global backward pass, each layer tries to predict the activity of the layer below it.
Beren Millidge and Anil K. Seth describe PC as a framework where the brain's primary job is to minimize the difference between its internal model and incoming sensory data (Predictive Coding: a Theoretical and Experimental Review). Crucially, Rafał Bogacz has shown that PC can mathematically replicate backpropagation using only local Hebbian plasticity—the "cells that fire together, wire together" principle (Reverse Differentiation via Predictive Coding). This provides a bridge between the engine of modern AI and neuroscientific theories of the cortex.
Physical Realization: Dendrites and Equilibrium
To make these algorithms truly biological, researchers look at the actual structure of neurons. Jordan Guergiuev and Blake Aaron Richards propose that segregated dendrites allow neurons to solve the credit assignment problem (Towards deep learning with segregated dendrites). By separating feedforward and feedback inputs into different compartments of the same cell, a neuron can calculate its own local error without needing a global supervisor.
Similarly, Equilibrium Propagation (EP) treats the neural network like a physical system seeking a state of minimum energy. Maxence Ernoult and Benjamin Scellier have extended this to convolutional architectures, showing that weight updates can occur in real-time as the system settles into equilibrium (Continual Weight Updates and Convolutional Architectures for Equilibrium Propagation).
The Scaling Challenge
Despite these theoretical advances, a "scaling gap" remains. In Assessing the Scalability of Biologically-Motivated Deep Learning Algorithms and Architectures, Sergey Bartunov and Geoffrey E. Hinton found that while methods like Feedback Alignment work on simple datasets like MNIST, they often struggle to match backpropagation's performance on massive datasets like ImageNet.
The current research frontier focuses on closing this gap, moving beyond simple synaptic weights to embrace the complexity of neural circuits, evolutionary "pre-wiring," and self-supervised objectives (Prince et al.). The goal is not just to make AI more "brain-like," but to unlock the energy efficiency and robustness inherent in biological learning.