Convergent Evolution in Neural Representation Space: Emergent Order in Deep Belief Networks

The paper investigates whether Deep Belief Networks (DBNs), trained in a purely unsupervised manner, spontaneously organize their internal representations according to the unknown data classes. The central finding is that class-specific clustering generally increases with depth across datasets and network widths, despite no label information being available during training. The authors analyze successive layers of DBNs trained on MNIST, Fashion-MNIST, and KMNIST using the Generalized Discrimination Value (GDV), supervised probes applied only after training, a reconstruction-based measure of abstraction distance, effective dimensionality, and free sample generation.

The key results are as follows:

1. Unsupervised class separation: The GDV decreases almost monotonically across successive layers for all three datasets, indicating progressively better-separated class clusters. This effect becomes more pronounced with narrower hidden layers. Control experiments show this effect depends on learned feature structure and cannot be explained by random transformations, weight marginals, dimensionality reduction, or sigmoid saturation.

2. Abstraction distance: The abstraction distance, operationalized as reconstruction MSE, increases monotonically across successive DBN layers. Early increases are mainly associated with the omission of idiosyncratic details (prototypical shift), while higher layers increasingly include distortions that may remove class-discriminative details.

3. Effective dimensionality collapse: The effective dimensionality decreases monotonically across layers, with an approximately 30-fold collapse for narrow networks. The authors decompose this collapse and find it is driven primarily by learned feature redundancy: many responsive neurons acquire increasingly aligned incoming weight directions and consequently produce statistically dependent responses. The weight matrix itself collapses to nearly the same effective dimension as the evoked population activity.

4. GDV vs. class accessibility: The first hidden layers frequently make class identity more accessible to linear and nonlinear probes. However, with greater depth, probe accuracy declines even as GDV continues to improve. This discrepancy is concentrated in a few difficult digit pairs (4–9, 5–8, 3–5) that become geometrically less separated and show increased classification confusion, while the other 42 pairs improve.

5. Origin of enhanced class accessibility: The first RBM can transform raw pixels into compact, class-accessible features without labels. Weight permutation destroys this effect, showing the learned spatial arrangement of weights is essential. The learned affine features already improve nonlinear class accessibility, while the sigmoid nonlinearity is needed for the linear-probe gain.

6. Free generation: Free dreams show a dataset-dependent depth progression, with greater depth generally allowing a wider repertoire of recognizable samples to emerge, although the trajectory is neither strictly monotonic nor equally successful for all datasets.

The authors conclude that layer-wise generative learning can spontaneously uncover and progressively amplify class-related structure in unlabeled data. They interpret this as evidence for algorithmic or Platonic attractors — preferred regions of parameter or representation space that independently initialized learning systems tend to approach under common constraints. The study demonstrates that abstraction distance, discriminative accessibility, and generative quality are related but non-equivalent properties of an unsupervised deep model.

Improvements to AI Systems:

1. Unsupervised Class-Discovery Layers: Add a class-clustering module to deep generative models that explicitly monitors GDV (Generalized Discrimination Value) across layers during unsupervised pretraining. The system can automatically halt layer-wise abstraction when GDV improvement stalls, preventing over-abstraction that destroys class-discriminative details (e.g., the 4–9, 5–8, 3–5 confusion). This yields a self-terminating feature extractor that maximizes class separability without labels.

2. Redundancy-Aware Compression: Implement a dynamic regularization term that penalizes alignment of incoming weight directions across neurons in each hidden layer, directly targeting the observed effective dimensionality collapse driven by feature redundancy. The improved system maintains high representational diversity (higher effective dimensionality) while still achieving class clustering, leading to more robust downstream transfer learning and less catastrophic forgetting.

3. Dual-Path Probe Architecture: Design a hybrid classifier that uses two parallel heads: one trained on early-layer features (high linear probe accuracy) and one on late-layer features (high GDV but lower linear accuracy). A gating network learns to route inputs based on input complexity (e.g., ambiguous digits) to the appropriate head. This improves classification accuracy on hard pairs (4–9, 5–8, 3–5) by leveraging early-layer geometric separation while retaining late-layer abstraction for easy samples.

4. Abstraction-Distance-Aware Generation: Modify the generative sampling process to use a layer-wise abstraction budget derived from the monotonic reconstruction MSE increase. The system can generate samples at a user-specified abstraction level (e.g., prototypical vs. distorted) by controlling how many top layers are traversed during free generation. This enables controllable creative output—e.g., producing idealized class prototypes vs. varied realistic samples—without retraining.

5. Label-Free Early Stopping for Discriminative Transfer: Use the first hidden layer’s linear probe accuracy (measured post-hoc without labels) as a meta-feature to decide when to stop unsupervised pretraining for a given downstream task. The system can automatically select the layer that maximizes class accessibility for a target dataset, avoiding the depth-dependent accuracy decline seen in the paper. This reduces compute by skipping unnecessary layers and improves fine-tuning performance.

6. Permutation-Invariant Weight Regularization: Add a training objective that encourages spatial weight structure (e.g., local receptive fields) in the first RBM, since weight permutation was shown to destroy class accessibility. The improved system explicitly enforces spatial smoothness or topographic organization in early layers, leading to more label-free class discovery in non-image domains (e.g., audio spectrograms, sensor data) where spatial structure is less obvious.

What the Improved AI System Can Do:

• Self-supervised clustering on unlabeled data with automatic depth selection, producing class-separated representations without any labels, while avoiding the hard-pair confusion problem.

• Adaptive feature extraction that switches between early and late layers based on input ambiguity, achieving higher accuracy on difficult classification tasks without supervised fine-tuning.

• Controllable generative modeling that lets users specify whether to generate prototypical or detailed samples, useful for data augmentation, anomaly detection (prototypes vs. outliers), and creative AI.

• Compute-efficient pretraining that stops at the optimal layer for a given downstream task, reducing training time and memory usage while maintaining or improving transfer learning performance.

• Robust representation learning in non-visual domains by enforcing spatial weight structure, enabling label-free discovery of meaningful categories in time-series or graph data.

Deep Belief Networks (DBNs) learn hierarchical generative models without class supervision. Here, we ask whether this purely unsupervised process nevertheless organizes internal representations according to the unknown data classes. We analyze successive layers of DBNs trained on MNIST, Fashion-MNIST, and KMNIST using the Generalized Discrimination Value (GDV), supervised probes applied only after training, a reconstruction-based measure of abstraction distance, effective dimensionality, and free sample generation. Remarkably, class-specific clustering generally increases with depth across datasets and network widths, although no label information is available during DBN training. Control experiments show that this effect depends on the learned feature structure and cannot be explained by random transformations, weight marginals, dimensionality reduction, or sigmoid saturation. The first hidden layers also frequently make class identity more accessible to linear and nonlinear probes. With greater depth, representations become increasingly compact and prototype-like as neurons acquire correlated feature directions. At the same time, GDV and probe accuracy reveal complementary aspects of class structure: improved average clustering can coexist with reduced accessibility for a few difficult class pairs. These findings demonstrate that layer-wise generative learning can spontaneously uncover and progressively amplify class-related structure in unlabeled data.