TL;DR
This paper proposes a unified framework for novelty detection (ND) using predictive coding networks (PCNs). They demonstrate that PCNs can naturally discriminate novel stimuli with high capacity, mirroring human abilities, and can do so at multiple levels of abstraction, from low-level sensory features to high-level semantic ones. This work suggests that ND might not require specialized brain circuits but can emerge from general-purpose learning and representation mechanisms.
What Problem Does This Paper Address?
Humans possess an remarkable ability to detect novel stimuli, a skill crucial for intelligent behavior. While it's hypothesized that this ability could arise from existing memory and representation learning networks, current computational models struggle to match the high capacity of human recognition memory, especially with complex, real-world data. Specifically, models often fail when dealing with correlated input patterns, a common characteristic of natural images. This paper aims to bridge this gap by showing that predictive coding networks, known for their representation learning and associative memory capabilities, can also achieve high-capacity novelty detection, even with correlated data, and can do so across different levels of abstraction.
Methods
The core of this work lies in applying predictive coding networks (PCNs) to novelty detection. PCNs are characterized by their hierarchical structure and the presence of "prediction error" neurons. The authors leverage an energy-based approach where the model learns to minimize an energy function for familiar patterns. During testing, the energy value of a new pattern serves as a novelty signal: higher energy indicates greater novelty.
Two main types of PCNs are explored:
- Recurrent PCN (rPCN): A single-layer model that learns to predict incoming input based on recurrent connections. It's shown to be analytically tractable and robust to pixel correlations.
- Hierarchical PCN (hPCN): A multi-layer model that allows for the detection of novelty at different levels of abstraction. Lower layers detect sensory novelty (e.g., pixel arrangement), while higher layers detect semantic novelty (e.g., object identity).
The authors also benchmark their PCN models against established novelty detection models like Hopfield Networks (HN), Modern Continuous Hopfield Networks (MCHN), Autoencoders (AE), and Variational Autoencoders (VAE).
Datasets
The paper uses the following datasets for training and evaluation:
- Gaussian Uncorrelated Data: 500-dimensional images with randomly generated pixels.
- Gaussian Correlated Data: 500-dimensional images with correlated pixels (0.4 covariance). This dataset is crucial for testing robustness to natural image characteristics.
- Tiny ImageNet: 64x64 pixel images, representing real-world visual data.
- MNIST Dataset: Used specifically for testing the hierarchical PCN's ability to detect novelty at different abstraction levels (digits 4, 5, and 9).
No new benchmarks were introduced; rather, existing datasets were used to compare the performance of PCNs against other ND models.
Key Insights and Ideas
- Unified Framework: Predictive coding offers a single, elegant framework that can naturally account for novelty detection, associative memory, and representation learning, suggesting these cognitive functions might be deeply intertwined.
- Error Neurons as Novelty Detectors: The activity of prediction error neurons in PCNs directly correlates with stimulus novelty. Higher error neuron activity signals a more novel stimulus, providing a clear neural correlate for novelty detection.
- Robustness to Correlation: Unlike many previous ND models, PCNs, particularly rPCNs, demonstrate remarkable robustness to correlated input patterns, a critical feature for processing natural images. This is attributed to the way PCNs transform the input space.
- Hierarchical Abstraction: hPCNs can detect novelty at multiple levels of abstraction. This means they can identify a new pixel pattern (sensory novelty) as well as a new object category or semantic meaning (semantic novelty), mirroring the hierarchical processing in the visual cortex.
- Repetition Suppression Explained: The paper provides a computational mechanism for repetition suppression observed in the brain, linking it to the minimization of prediction errors by error neurons.
Key Results
- Human-like Capacity: rPCNs achieve novelty detection accuracy that closely matches human performance, particularly in discriminating familiar from novel natural images.
- Superior Performance with Correlation: When tested on correlated data, rPCNs significantly outperform HN, MCHN, AE, and VAE, maintaining high accuracy where other models degrade to chance levels.
- Multi-level Novelty Detection: hPCNs successfully differentiate between sensory novelty (e.g., a slightly different image of the same digit) and semantic novelty (e.g., an image of a different digit), with different layers of the hierarchy specializing in these tasks. For instance, in the MNIST experiment, the top layer of the hPCN could distinguish between digits 4 and 5 (semantic novelty), while lower layers were more sensitive to pixel-level differences.
- Negligible Variability: The PCN models exhibit very low variability in their performance across simulations, indicating a stable and reliable mechanism.
Limitations and Future Work
- Computational Cost: While PCNs are shown to be effective, the paper acknowledges that scaling them to very large datasets or complex architectures might still present computational challenges, though recent work in improving PCN scalability is mentioned.
- Experimental Predictions: The authors propose that neurons exhibiting repetition suppression in the brain could be identified as error neurons in PCNs, opening avenues for experimental validation.
- Temporal Novelty: The paper suggests that extending the PCN framework to temporal predictive coding networks (tPCNs) could address the detection of novelty in sequences and videos.
- Out-of-Distribution Detection: The authors note that hPCNs can be viewed as performing out-of-distribution detection, suggesting potential applications in machine learning anomaly detection tasks.
- AM and ND Relationship: The work provides a computational model where AM and ND are handled by the same underlying mechanism, offering a concrete way to think about their complex relationship in the brain.
Summary updated