In traditional neuroscience, neural spikes are often viewed as noisy, stochastic events whose meaning is buried in their average firing rate. However, a growing body of research suggests that spikes are far more purposeful: they are discrete, optimal updates designed to minimize a global representation error. In this framework, the "noise" observed in the brain is actually the signature of a highly efficient system that communicates only what is necessary to correct its internal model of the world.
The Membrane Potential as a Running Error
The core intuition of this research direction is that a neuron’s membrane potential is not just a precursor to a spike, but a real-time physical representation of a prediction error.
As shown by Sophie Denève in Bayesian inference in spiking neurons, neurons can be viewed as agents performing Bayesian inference. Instead of transmitting raw data, a neuron tracks the difference between an external signal and the network's current estimate of that signal. When this internal estimate deviates significantly from reality, the membrane potential rises.
A spike is triggered only when this error exceeds a specific threshold. Crucially, the act of spiking "resets" the error. As Ralph Bourdoukan, Christian K. Machens, and colleagues demonstrate in Learning optimal spike-based representations, a neuron fires an action potential if and only if that spike reduces the squared error between the input signal and its linear reconstruction.
Greedy Optimization and Metabolic Efficiency
This mechanism transforms the biological process of neural firing into a greedy optimization task. Each spike is a discrete "correction" that pulls the network's representation back toward the truth.
This approach explains the remarkable metabolic efficiency of the brain. In Predictive Coding of Dynamical Variables in Balanced Spiking Networks, Martin Boerlin and Sophie Denève show that by minimizing a loss function that combines representation error with a penalty for spiking, the network naturally produces sparse, asynchronous activity. Because neurons only fire when they have something "new" to contribute to the error reduction, they avoid the massive redundancy found in traditional rate-based models.
This efficiency is further explored by Jaldert O. Rombouts and Sander M. Bohté in Fractionally Predictive Spiking Neurons, where they demonstrate that using power-law kernels allows neurons to encode signals with even fewer spikes, treating the spike train as a fractional derivative of the signal.
The Role of Excitation-Inhibition Balance
One of the most striking features of the mammalian cortex is the "tight balance" between excitatory (E) and inhibitory (I) inputs. In this predictive framework, E/I balance is the mechanism that implements error calculation.
Inhibition serves to "cancel out" the predictable parts of a signal. When a neuron receives an excitatory input (the signal), it is immediately countered by inhibitory feedback (the network's current prediction). The remaining voltage—the "unbalanced" part—is the prediction error.
Alireza Alemi and Jean-Jacques Slotine expanded this into the realm of control theory in Learning Nonlinear Dynamics in Efficient, Balanced Spiking Networks Using Local Plasticity Rules. They showed that by maintaining this balance, networks can learn complex, nonlinear dynamical systems using only local learning rules, as the global error is always locally available at each neuron's membrane.
Moving Beyond Explicit Error Neurons
While classical predictive coding models often require specialized "error neurons," recent work suggests the brain might be more integrated. Fabian A. Mikulasch and Viola Priesemann propose in Dendritic predictive coding that these errors are computed within individual dendritic compartments.
This "implicit" error representation is a recurring theme. As noted in the survey Predictive Coding with Spiking Neural Networks, the presence of a spike signals an unexpected update, while the absence of a spike (suppression via inhibition) signals that the input was already predicted. This allows the network to communicate "surprisal" without needing a dedicated architecture for every error type.
Evidence from Biology and AI
The predictive nature of spikes isn't just theoretical; it aligns with empirical observations across different brain regions and artificial models:
- Language: Ariel Goldstein and Kenneth A. Norman found that neural spikes in humans correlate with the "surprise" or cross-entropy signals in deep language models like GPT-2 (Shared computational principles for language processing).
- Motor Control: In the cerebellum, Marc Junker and Martin A. Giese observed that climbing fiber signals "reverberate" past errors to guide future actions, acting as a memory of performance discrepancies (Learning from the past).
- Vision: Models like PredNet, discussed by Gabriel Kreiman, show that training a network simply to predict the next video frame leads to the emergence of neural response patterns identical to those in the biological visual cortex (A neural network trained to predict future video frames).
By treating spikes as discrete corrections within a continuous optimization process, this research direction provides a unified explanation for why the brain is so noisy, so balanced, and so incredibly efficient.