OpenNovelty.org

Bidirectional Predictive Coding

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.5Download Report PDF
predictive codingsensory processingdiscriminative and generative tasks
Abstract:

Predictive coding (PC) is an influential computational model of visual learning and inference in the brain. Classical PC was proposed as a top-down generative model, where the brain actively predicts upcoming visual inputs, and inference minimises the prediction errors. Recent studies have also shown that PC can be formulated as a discriminative model, where sensory inputs predict neural activities in a feedforward manner. However, experimental evidence suggests that the brain employs both generative and discriminative inference, while unidirectional PC models show degraded performance in tasks requiring bidirectional processing. In this work, we propose bidirectional PC (bPC), a PC model that incorporates both generative and discriminative inference while maintaining a biologically plausible circuit implementation. We show that bPC matches or outperforms unidirectional models in their specialised generative or discriminative tasks, by developing an energy landscape that simultaneously suits both tasks. We also demonstrate bPC's superior performance in two biologically relevant tasks including multimodal learning and inference with missing information, suggesting that bPC resembles biological visual inference more closely.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes bidirectional predictive coding (bPC), which integrates both generative (top-down) and discriminative (bottom-up) inference pathways within a single predictive coding architecture. It resides in the 'Bidirectional Predictive Coding Models' leaf, which contains only three papers total, indicating a relatively sparse research direction within the broader taxonomy of 19 papers. This leaf sits under 'Unified Architectures Integrating Generative and Discriminative Processing', suggesting the work addresses a core challenge in the field: explicitly combining both inference modes rather than treating them separately.

The taxonomy reveals neighboring approaches in sibling leaves: 'Diffusion Model Adaptations for Discriminative Tasks' (2 papers) adapts generative diffusion models for perception, while 'Hierarchical Model Combination Frameworks' (2 papers) combines separate models at different levels. The broader 'Specialized Generative or Discriminative Approaches' branch (7 papers) contains work emphasizing one mode without architectural integration. The paper's bidirectional approach distinguishes it from these alternatives by maintaining simultaneous pathways rather than sequential adaptation or hierarchical separation, positioning it closer to neuroscience-inspired theories of cortical processing.

Among 21 candidates examined across three contributions, the core 'Bidirectional predictive coding model' contribution shows one refutable candidate out of 10 examined, suggesting some prior work in this specific architectural space. The 'Energy landscape explanation' contribution found no refutable candidates among 10 examined, indicating this theoretical framing may be more novel. The 'Demonstration of biological relevance' contribution examined only 1 candidate with no refutation. The limited search scope (21 total candidates, not hundreds) means these statistics reflect top semantic matches rather than exhaustive coverage, particularly for the biological validation aspects.

Given the sparse taxonomy leaf (3 papers) and limited search scope, the work appears to occupy a relatively underexplored niche within predictive coding research. The single refutable candidate for the core contribution suggests some architectural precedent exists, though the energy landscape framing and biological task demonstrations show less overlap in the examined literature. A broader search might reveal additional related work, particularly in neuroscience-oriented predictive coding literature not captured by semantic similarity to this machine learning-focused abstract.

Taxonomy

Core-task Taxonomy Papers
19
3
Claimed Contributions
21
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: Combining generative and discriminative visual inference in predictive coding. The field explores how biological and artificial systems can integrate two complementary modes of visual processing—generative models that synthesize or predict sensory input, and discriminative models that classify or recognize patterns—within a unified predictive coding framework. The taxonomy reflects this duality through several main branches. Unified Architectures Integrating Generative and Discriminative Processing encompasses works that explicitly merge both inference modes, often drawing inspiration from neuroscience or hybrid learning paradigms, as seen in Unsupervised Generative Discriminative[3] and Predictive Coding Network[19]. Specialized Generative or Discriminative Approaches groups studies that emphasize one mode over the other, while Cross-Modal and Knowledge Transfer Applications examines how these principles extend beyond single-modality vision. Domain-Specific Applications and Evaluation addresses practical deployments in areas such as medical imaging or robotics, and Conceptual Frameworks and Theoretical Analyses provides the foundational theories linking predictive coding to brain function and machine learning. Within the Unified Architectures branch, a particularly active line of work focuses on bidirectional predictive coding models that iteratively refine both top-down predictions and bottom-up error signals. Bidirectional Predictive Coding[0] sits squarely in this cluster, proposing mechanisms that allow feedback and feedforward pathways to jointly optimize visual representations. This approach contrasts with earlier efforts like Unsupervised Generative Discriminative[3], which combined generative and discriminative objectives but did not emphasize the bidirectional dynamics central to predictive coding theory. Meanwhile, Predictive Coding Network[19] explores similar iterative refinement but with different architectural choices for balancing reconstruction and recognition losses. A recurring theme across these works is the trade-off between computational efficiency and biological plausibility: some models prioritize scalable discriminative performance, while others aim to capture the hierarchical, recurrent structure observed in cortical circuits. The original paper's emphasis on bidirectional flow positions it as a bridge between neuroscience-inspired theories and practical deep learning architectures.

Claimed Contributions

Bidirectional predictive coding model

The authors introduce bPC, a predictive coding model that unifies generative (top-down) and discriminative (bottom-up) inference within a single energy function. This model maintains biological plausibility through local computations and Hebbian learning rules while enabling flexible inference in both directions.

10 retrieved papers
Can Refute
Energy landscape explanation for bidirectional performance

The authors demonstrate that bPC develops an energy landscape optimized for both discriminative and generative tasks simultaneously. This landscape avoids the overconfidence of discriminative models and the class-mean collapse of generative models, resulting in superior performance across both task types.

10 retrieved papers
Demonstration of biological relevance through multimodal and occlusion tasks

The authors show that bPC excels in biologically motivated scenarios: learning associations across sensory modalities (analogous to linking visual and auditory information) and classifying images with missing information (similar to handling retinal blind spots). These results suggest bPC may more faithfully model visual inference in the brain.

1 retrieved paper

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Bidirectional predictive coding model

The authors introduce bPC, a predictive coding model that unifies generative (top-down) and discriminative (bottom-up) inference within a single energy function. This model maintains biological plausibility through local computations and Hebbian learning rules while enabling flexible inference in both directions.

Contribution

Energy landscape explanation for bidirectional performance

The authors demonstrate that bPC develops an energy landscape optimized for both discriminative and generative tasks simultaneously. This landscape avoids the overconfidence of discriminative models and the class-mean collapse of generative models, resulting in superior performance across both task types.

Contribution

Demonstration of biological relevance through multimodal and occlusion tasks

The authors show that bPC excels in biologically motivated scenarios: learning associations across sensory modalities (analogous to linking visual and auditory information) and classifying images with missing information (similar to handling retinal blind spots). These results suggest bPC may more faithfully model visual inference in the brain.

Table of Contents

Bidirectional Predictive Coding | Novelty Validation