Dendritic predictive coding: A theory of cortical computation with spiking neurons

Fabian A. Mikulasch, Lucas Rudelt, Michael Wibral, Viola Priesemann

Abstract

Top-down feedback in cortex is critical for guiding sensory processing, which has prominently been formalized in the theory of hierarchical predictive coding (hPC). However, experimental evidence for error units, which are central to the theory, is inconclusive, and it remains unclear how hPC can be implemented with spiking neurons. To address this, we connect hPC to existing work on efficient coding in balanced networks with lateral inhibition, and predictive computation at apical dendrites. Together, this work points to an efficient implementation of hPC with spiking neurons, where prediction errors are computed not in separate units, but locally in dendritic compartments. The implied model shows a remarkable correspondence to experimentally observed cortical connectivity patterns, plasticity and dynamics, and at the same time can explain hallmarks of predictive processing, such as mismatch responses, in cortex. We thus propose dendritic predictive coding as one of the main organizational principles of cortex.

Introduction

For decades, Hierarchical Predictive Coding (hPC) has been a leading theory in neuroscience, suggesting the brain is an "inference machine" that constantly compares sensory inputs against internal expectations. However, classical hPC faces two major hurdles. First, the "Error Unit" Mystery: the theory requires dedicated "error neurons" to signal mismatches between prediction and reality, yet experimental evidence for such a distinct population is thin and inconsistent. Second, the "Spiking Gap": most hPC models rely on continuous firing rates, whereas real neurons communicate via discrete, energy-expensive spikes.

Comparison of classical predictive coding with the proposed dendritic and spiking implementations.

This paper proposes "dendritic predictive coding," a theory where individual neurons compute prediction errors within their own dendritic compartments rather than relying on specialized "error neurons." By shifting the computation of error from the population level down to the sub-cellular level, the model provides a biologically plausible bridge between high-level Bayesian inference and cortical micro-circuitry.

The Dendritic Predictive Coding Model

The authors synthesize classical hPC with efficient coding in balanced networks and neural sampling. Instead of separate error units, the model utilizes the voltage dynamics of different dendritic compartments:

  • Basal Dendrites: These compute "bottom-up" errors. They utilize lateral inhibition from neighboring neurons to "explain away" predictable inputs. If an input is already predicted by the network, the inhibition cancels it out; the remaining voltage represents the prediction error.
  • Apical Dendrites: These compute "top-down" errors by receiving feedback from higher cortical areas, comparing the neuron's current activity to higher-level expectations.
Mechanism of spike-based prediction and error integration within a single neuron.

In this framework, a neuron only fires a spike when its somatic potential—the sum of its dendritic errors—exceeds a threshold. Firing a spike effectively "explains" the input to the rest of the network, immediately reducing the local error. This leads to a highly energy-efficient sparse code. Furthermore, the authors derive learning rules based on local membrane potentials (Voltage-Dependent Plasticity), which are more biologically grounded than standard backpropagation, relying on the local state of the dendrite rather than global error signals.

Biological Plausibility and Results

The model maps onto known cortical architecture with high fidelity. PV+ interneurons provide the fast lateral inhibition required for basal dendritic errors, while SST+ interneurons modulate apical dendrites for top-down feedback.

To validate the theory, the authors simulated a mouse moving through a virtual environment to replicate "mismatch responses"—a phenomenon where visual neurons fire specifically when expected visual flow (based on movement) doesn't match actual input. The model successfully explained why neurons in the primary visual cortex (V1) fire more strongly when a mouse's movement and visual feedback are decoupled. Additionally, the derived learning rules qualitatively match experimental observations of how local inhibition modulates synaptic changes (LTP/LTD) in pyramidal cells.

Discussion

While promising, the theory currently assumes linear dendritic integration, whereas real dendrites exhibit non-linearities like NMDA spikes. Future work must integrate these non-linearities and address the "weight transport problem"—the requirement for symmetric feedforward and feedback weights. The ultimate test of this theory will involve recording from individual dendritic branches in vivo to determine if their local voltage truly tracks prediction error as predicted.

Summary updated

Abstract

Top-down feedback in cortex is critical for guiding sensory processing, which has prominently been formalized in the theory of hierarchical predictive coding (hPC). However, experimental evidence for error units, which are central to the theory, is inconclusive, and it remains unclear how hPC can be implemented with spiking neurons. To address this, we connect hPC to existing work on efficient coding in balanced networks with lateral inhibition, and predictive computation at apical dendrites. Together, this work points to an efficient implementation of hPC with spiking neurons, where prediction errors are computed not in separate units, but locally in dendritic compartments. The implied model shows a remarkable correspondence to experimentally observed cortical connectivity patterns, plasticity and dynamics, and at the same time can explain hallmarks of predictive processing, such as mismatch responses, in cortex. We thus propose dendritic predictive coding as one of the main organizational principles of cortex.
BibTeX
@inproceedings{mikulasch2022dendritic,
  author = {Fabian A. Mikulasch and Lucas Rudelt and Michael Wibral and Viola Priesemann},
  title = {{Dendritic predictive coding: A theory of cortical computation with spiking neurons}},
  booktitle = {Trends},
  year = {2022},
  doi = {10.1016/j.tins.2022.09.007},
}
Figure