The Backpropagation Revolution in Biological Brains

In the sleek, silicon-driven world of modern artificial intelligence, Backpropagation (BP) is the mathematical engine of every LLM and image generator. Yet, it has long been dismissed as a biological impossibility because the human brain was thought to lack the specialized "wiring" to perform the complex calculus BP requires.

But what if the brain isn't trying to do math at all? What if it is simply trying to stay balanced?

A Bridge Between AI and Biology

A provocative new study from researchers at Johns Hopkins University suggests the gap between biological brains and artificial intelligence is much smaller than we imagined. This discovery matters because it finally bridges the "black box" of AI with the wetware of the human mind.

If our brains naturally perform the same learning algorithms used by GPT-4, we are one step closer to:

Developing low-power, brain-like hardware.
Understanding how neurobiological disorders disrupt our ability to learn.

The Core Discovery: Backpropagation as an Emergent Property

The team has mathematically proven that Backpropagation is not a specialized computational trick, but an emergent property of individual neurons striving to maintain Excitatory-Inhibitory (E-I) balance.

By introducing a novel neuroplasticity rule, they showed that when the brain's internal balance function corresponds to the derivative of its activation function, the system becomes mathematically identical to Gradient Descent.

Bypassing the "Biological BP" Hurdles

The researchers bypassed three classic biological hurdles by focusing on retrograde signaling. In their model, temporal scales are key:

Information flows backward across the synapse in seconds.
Neurons fire in milliseconds.
Long-term plasticity takes seconds to minutes.

This approach eliminated the need for:

Symmetric weights.
Separate learning phases.
The computation of complex derivatives.

How the Learning Rate Shapes Brain Architecture

In simulations with an all-to-all connected network of 300 neurons, the team discovered the learning rate ( $r$ ) acts as a master architect for the brain’s physical structure.

Low $r$ value: Forces the network into block-diagonal structures, promoting connectivity within modules.
High $r$ value: Shifts the architecture toward off-diagonal block connectivity, promoting communication across modules.

Key Assumptions and Biological Challenges

Even elegant math faces the chaos of biology. The model relies on several critical assumptions that future research must validate.

Equilibrium Requirement: The model assumes credit must be redistributed before weights update—a process requiring m-1 timesteps in an m-layered network.

Potential System Instability: Without external credit signals to "clamp" the system, the credit distribution can spiral to zero or diverge to infinity. This mimics the instability seen in cases of extreme sensory deprivation.

Precision Mapping: For perfect equivalence, the biological "balance function" must map precisely to the mathematical derivative. It remains to be confirmed if physical neurons possess this precision or simply approximate the path to equilibrium.

Conclusion: From Mathematical Trick to Biological Principle

"BP can be interpreted as a simple consequence of individual neurons striving to maintain E-I balance," the authors noted. This study reframes backpropagation from an engineered computational trick into a potential natural law of learning, deeply rooted in the brain's fundamental drive for stability.

Reference: Fan, X., & Mysore, S. P. (2025). Neurons that contribute to balance, wire in accordance: Emergence of backpropagation from a simple, bio-plausible neuroplasticity rule. arXiv:2405.14139v2 [q-bio.NC].