Predictive Coding Networks and Inference Learning: Tutorial and Survey

Recent years have witnessed a growing call for renewed emphasis on neuroscience-inspired approaches in artificial intelligence research, under the banner of NeuroAI. A prime example of this is predictive coding networks (PCNs), based on the neuroscientific framework of predictive coding. This framework views the brain as a hierarchical Bayesian inference model that minimizes prediction errors through feedback connections. Unlike traditional neural networks trained with backpropagation (BP), PCNs utilize inference learning (IL), a more biologically plausible algorithm that explains patterns of neural activity that BP cannot. Historically, IL has been more computationally intensive, but recent advancements have demonstrated that it can achieve higher efficiency than BP with sufficient parallelization. Furthermore, PCNs can be mathematically considered a superset of traditional feedforward neural networks (FNNs), significantly extending the range of trainable architectures. As inherently probabilistic (graphical) latent variable models, PCNs provide a versatile framework for both supervised learning and unsupervised (generative) modeling that goes beyond traditional artificial neural networks. This work provides a comprehensive review and detailed formal specification of PCNs, particularly situating them within the context of modern ML methods. This positions PC as a promising framework for future ML innovations.

💡 Research Summary

This paper presents a comprehensive tutorial and survey of Predictive Coding Networks (PCNs) and their learning algorithm, Inference Learning (IL), positioning them as a neuro‑inspired alternative to conventional deep learning. The authors begin by framing predictive coding (PC) as a hierarchical Bayesian inference model of the brain that continuously generates top‑down predictions and minimizes the resulting prediction errors via feedback connections. They contrast this biologically plausible mechanism with the widely used back‑propagation (BP) algorithm, highlighting that PCNs replace the global error signal of BP with locally computed prediction errors that propagate both upward (error) and downward (prediction).

The core technical contribution is a detailed formal specification of PCNs. Each layer ℓ contains an activity vector a^ℓ, a locally predicted activity μ^ℓ = f(W^{ℓ‑1}a^{ℓ‑1}), and a prediction error ε^ℓ = a^ℓ − μ^ℓ. The network’s energy (or variational free energy) is defined as E = ½∑_ℓ‖ε^ℓ‖². During training, the input layer is clamped to the data x and the output layer to the target y; hidden activities are then updated by gradient descent on E (the inference phase) using Δa^ℓ = ‑γ∂E/∂a^ℓ. After convergence, weights are updated locally as ΔW^ℓ ∝ ε^{ℓ+1}(a^ℓ)ᵀ. This two‑step procedure—iterative inference followed by weight adjustment—is the essence of IL.

A key insight is that discriminative PCNs become mathematically equivalent to feed‑forward neural networks (FNNs) at test time: when prediction errors vanish, the hidden activities satisfy a^ℓ = f(W^{ℓ‑1}a^{ℓ‑1}), reproducing the standard forward pass. Consequently, IL can be viewed as a generalized learning rule that reduces to BP under the zero‑error fixed point, while offering a parallelizable alternative during training. Because IL relies only on locally available information, the serial dependence of BP’s forward‑backward sweep can be eliminated; with sufficient hardware parallelism, computation time no longer scales with network depth. The authors cite empirical work showing that, on modern GPUs/TPUs, IL can surpass BP in wall‑clock efficiency for deep architectures.

Beyond the algorithmic perspective, the paper treats PCNs as probabilistic latent‑variable models. The hierarchical Gaussian formulation leads to an Expectation‑Maximization‑like procedure: the inference phase corresponds to the E‑step (inferring hidden states), and the weight update corresponds to the M‑step (maximizing the complete‑data log‑likelihood or variational free energy). This connects PCNs to generative modeling frameworks such as variational autoencoders and diffusion models, and explains why PCNs naturally support both supervised (prediction errors flow from data to labels) and unsupervised learning (errors flow toward data reconstruction).

A major extension discussed is the concept of “PC graphs.” While classic PCNs assume a strict hierarchy, recent work shows that IL can be defined on arbitrary directed graphs, enabling heterarchical, brain‑like connectivity patterns that are not trainable with standard BP. This broadens the expressive power of PCNs, making them a strict superset of traditional ANNs: any feed‑forward architecture can be embedded as a special case, but PCNs also admit non‑hierarchical, sparse, or recurrent topologies.

The authors also address computational complexity. Although IL introduces an inner inference loop, the locality of updates and the possibility of parallel execution mitigate the overhead. They provide a thorough analysis showing that, for a given depth L, the asymptotic time complexity of a fully parallelized PCN scales as O(1) with respect to L, compared to O(L) for BP. This property makes PCNs attractive for neuromorphic hardware where energy and latency are critical.

To support reproducibility, the paper releases PRECO, an open‑source PyTorch library implementing discriminative PCNs, generative PCNs, and arbitrary PC graphs. The library includes utilities for clamping, inference dynamics, and weight updates, allowing researchers to experiment with the full spectrum of PCN models without implementing the mathematics from scratch.

In the literature review, the authors position their work relative to three recent surveys (Millidge et al., 2022/2023; Salvatori et al., 2023). While those surveys provide broad overviews, this paper distinguishes itself by delivering a deep mathematical treatment, explicit algorithmic derivations, and a hands‑on toolkit.

The conclusion emphasizes that PCNs unify three perspectives: (1) a learning algorithm comparable to BP but more biologically plausible, (2) a probabilistic latent‑variable model comparable to modern generative approaches, and (3) a generalized neural architecture that subsumes feed‑forward networks and admits arbitrary graph topologies. The authors argue that this trifecta positions PCNs as a promising foundation for future NeuroAI research, especially in domains requiring energy‑efficient, online, or continual learning. Open challenges include scaling IL to massive datasets, integrating PCNs with reinforcement learning, and deploying them on dedicated neuromorphic chips.