Can the Brain Do Backpropagation? - Exact Implementation of Backpropagation in Predictive Coding Networks

Yuhang Song, Thomas Lukasiewicz, Zhenghua Xu, Rafal BogaczNeurIPS 2020 · OpenReview

Whether the brain actually performs backpropagation (BP)—the engine behind modern AI—is one of the most enduring debates in neuroscience. While BP is incredibly effective, it is often criticized as "biologically implausible" because it requires global information and external control signals that real neurons don't seem to have.

In this paper, researchers present a breakthrough: Zero-divergent Inference Learning (Z-IL). They demonstrate, for the first time, a framework where a biologically plausible network can implement the exact same weight updates as backpropagation using only local computations and autonomous control.

TL;DR

The authors prove that Predictive Coding Networks (PCNs) can implement backpropagation exactly, rather than just approximating it. By introducing "Zero-divergent Inference Learning" (Z-IL) and its fully autonomous variant (Fa-Z-IL), they bridge the gap between the global requirements of AI training and the local, autonomous constraints of the biological brain.

What Problem Does This Paper Address?

Artificial Neural Networks (ANNs) rely on BP to calculate how to change weights to reduce error. However, BP has three major "biological gaps":

  1. Exactness: It has been unclear if biological learning (BL) can match BP’s precision or if it only produces rough approximations.
  2. Local Plasticity: In BP, updating a weight requires information from distant layers. In the brain, a synapse should only change based on the activity of the two neurons it connects.
  3. Autonomy: BP requires a "global controller" to switch the network between a prediction phase and a learning phase. The brain operates continuously and autonomously.

This paper aims to close these gaps using Predictive Coding Networks (PCNs), a popular neuroscience framework where neurons try to predict the activity of other neurons and minimize the resulting "prediction error."

Methods

The authors build upon Inference Learning (IL), a supervised learning algorithm for PCNs. Their contribution consists of two main stages:

1. Zero-divergent IL (Z-IL)

They identify three specific, "easy-to-satisfy" conditions (C1–C3) under which the weight updates in a PCN become mathematically identical to those in an ANN trained with BP:

  • C1 (Initialization): The network starts in a state where it has already converged on a prediction.
  • C2 (Timing): Weights at layer $l$ are updated at a specific "inference moment" $t=l$.
  • C3 (Integration Step): The step size for internal inference is set to 1.
Comparison of ANN and PCN structures. While ANNs propagate activity directly, PCNs use error nodes to mediate the flow of information.
Mapping of variables between Artificial Neural Networks (ANNs) and Predictive Coding Networks (PCNs).

2. Fully Autonomous Z-IL (Fa-Z-IL)

To remove the need for an external signal to trigger weight updates (Condition C2), they introduce a local function, $\phi(\cdot)$. This function acts like a "spike" detector: it monitors local error nodes and triggers a weight update only when it detects a specific change in error dynamics. This allows the network to perform all computations locally, simultaneously, and fully autonomously.

Datasets

The authors evaluated their models on two standard benchmarks to prove both correctness and scalability:

  • MNIST: Used for ablation studies to show that violating conditions C1–C3 leads to divergence from BP.
  • ImageNet: Used to measure runtime and demonstrate that Z-IL can scale to large-scale, real-world computer vision tasks.

Key Insights and Ideas

  • Equivalence, Not Approximation: Most previous "biologically plausible" models only claimed to approximate BP. This paper provides a mathematical proof that PCNs can be exactly equivalent.
  • Local Error as a Trigger: The most clever insight is using the temporal dynamics of local error nodes to replace a global "update" signal. By "watching" how local errors change, a synapse can "know" exactly when to update itself without being told by a central controller.
  • Efficiency: Because Z-IL updates weights at specific moments rather than waiting for the entire network to settle into a global equilibrium (as standard IL does), it is significantly faster and more computationally efficient.

Key Results

  • Mathematical Proof: The authors successfully proved that under conditions C1–C3, the PCN weight update exactly matches the BP update ($\Delta\theta = \Delta w$).
  • Ablation Success: On MNIST, they showed that if any of the three conditions are relaxed, the PCN's performance begins to diverge from the ANN's performance.
  • Scalability on ImageNet: Z-IL and Fa-Z-IL achieved runtimes nearly identical to standard BP on both CPU and GPU, whereas previous versions of Inference Learning (IL) were up to 15x slower.

Limitations and Future Work

While this work is a major step forward, some biological hurdles remain:

  • Weight Symmetry: Like most BP-equivalent models, this one assumes that feedback connections have the same weights as feedforward connections (the "weight transport problem"). While the authors mention recent work suggesting this can be mitigated, it wasn't the primary focus here.
  • Non-Spiking Neurons: The current model uses continuous-valued neurons. Translating this exact equivalence to a spiking neural network (SNN) is a logical next step for even greater biological realism.
  • The Prediction Phase: The model still requires a "prediction phase" to converge before learning starts, which the authors acknowledge is a requirement of the learning setup itself.

Overall, this research provides a powerful theoretical bridge, suggesting that the brain doesn't need to "invent" a new algorithm—it might just be implementing backpropagation in a very clever, localized way.

Summary updated

Abstract

Backpropagation (BP) has been the most successful algorithm used to train artificial neural networks. However, there are several gaps between BP and learning in biologically plausible neuronal networks of the brain (learning in the brain, or simply BL, for short), in particular, (1) it has been unclear to date, if BP can be implemented exactly via BL, (2) there is a lack of local plasticity in BP, i.e., weight updates require information that is not locally available, while BL utilizes only locally available information, and (3)~there is a lack of autonomy in BP, i.e., some external control over the neural network is required (e.g., switching between prediction and learning stages requires changes to dynamics and synaptic plasticity rules), while BL works fully autonomously. Bridging such gaps, i.e., understanding how BP can be approximated by BL, has been of major interest in both neuroscience and machine learning. Despite tremendous efforts, however, no previous model has bridged the gaps at a degree of demonstrating an equivalence to BP, instead, only approximations to BP have been shown. Here, we present for the first time a framework within BL that bridges the above crucial gaps. We propose a BL model that (1) produces \emph{exactly the same} updates of the neural weights as~BP, while (2)~employing local plasticity, i.e., all neurons perform only local computations, done simultaneously. We then modify it to an alternative BL model that (3) also works fully autonomously. Overall, our work provides important evidence for the debate on the long-disputed question whether the brain can perform~BP.
BibTeX
@inproceedings{song2020can,
  author = {Yuhang Song and Thomas Lukasiewicz and Zhenghua Xu and Rafal Bogacz},
  title = {{Can the Brain Do Backpropagation? - Exact Implementation of Backpropagation in Predictive Coding Networks}},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2020},
  url = {https://openreview.net/forum?id=EhYXAmRJSV},
}
Figure