Brain-Inspired Perspective on Configurations: Unsupervised Similarity and Early Cognition

Juntang Wang, Yihan Wang, Hao Wu, Dongmian Zou, Shixin XuarXiv 2025 · arXiv · OpenReview

TL;DR

This paper introduces "configurations," a novel clustering framework inspired by early human cognition, that naturally handles hierarchical organization, novelty detection, and dynamic category adaptation with a single parameter. To evaluate these brain-like capabilities, they developed "mheatmap," a visualization and alignment tool, demonstrating that configurations outperform traditional methods on these cognitive tasks.

What Problem Does This Paper Address?

Modern machine learning, particularly in unsupervised learning and clustering, struggles to replicate the innate abilities of infants and animals to discover structure, form categories, and detect novelty without explicit labels or complex engineered tasks. Existing methods often require multiple hyperparameters, fail to capture hierarchical relationships, or cannot gracefully handle evolving categories. This gap limits AI's ability to generalize in open-world scenarios and adapt to dynamic environments. The authors aim to bridge this gap by proposing a computational framework that mimics early cognitive processes.

Methods

The core of the paper is the "configurations" framework, a finite-resolution clustering approach that uses a single parameter, $\gamma$, to control the granularity of clusters. This parameter influences an energy function that balances "attraction" (favoring larger, coarser clusters at low $\gamma$) and "repulsion" (favoring smaller, finer clusters at high $\gamma$). This attraction-repulsion dynamic is inspired by how infants might organize information.

To address the challenge of evaluating multi-resolution and dynamic clustering, the authors introduce mheatmap. This is a proportional visualization tool that uses geometric properties (cell width and height) and color to represent overlap between predicted clusters and ground truth categories. It offers a more intuitive understanding of merge-split dynamics than traditional confusion matrices.

Complementing mheatmap is the Reverse Merge/Split (RMS) Alignment Algorithm. This algorithm is designed to align cluster assignments across different resolutions, accounting for the semantic coherence of merge-split transitions rather than treating them as arbitrary reassignments. This is crucial for fair evaluation in dynamic clustering scenarios.

Datasets

The paper evaluates its methods on several datasets, noting their hierarchical structure:

  • CIFAR-10 Subset
  • Infant Study Stimuli (from developmental studies)
  • ImageNet Hierarchical
  • Salinas (a hyperspectral dataset)

Image embeddings are generated using a pre-trained ViT-B/16 model.

Key Insights and Ideas

  • Single Parameter for Granularity: The idea of using a single parameter ($\gamma$) to control the entire hierarchy of clusters is a significant simplification compared to methods requiring multiple resolution parameters or fixed cluster counts.
  • Energy-Based Novelty Detection: The framework's energy function naturally signals novelty. When a new, dissimilar item is introduced, it increases both attraction and repulsion costs, leading to a higher energy state, mirroring how novel stimuli disrupt infant expectations.
  • Intrinsic Hierarchical Organization: The attraction-repulsion dynamics inherently create a hierarchical structure, allowing for the discovery of both superordinate (e.g., animals vs. vehicles) and basic-level (e.g., cats vs. dogs) categories by simply adjusting the $\gamma$ parameter.
  • Dynamic Adaptation: The framework's ability to handle "merge-split dynamics" as $\gamma$ changes allows it to adapt to evolving categories, a capability lacking in many static clustering methods.
  • mheatmap for Fair Evaluation: The development of mheatmap and the RMS algorithm is a crucial contribution, providing a much-needed tool for evaluating dynamic and multi-resolution clustering in a cognitively meaningful way, overcoming limitations of traditional metrics like ARI.

Key Results

  • Competitive Clustering Performance: Configurations achieve competitive or superior performance on standard clustering metrics (ARI, NMI) across datasets, often outperforming baselines by a significant margin (e.g., 7-49% improvement).
  • Novelty Detection: The framework achieves an 87% AUC for novelty detection, aligning with infant habituation studies.
  • Dynamic Stability: Configurations demonstrate 35% better stability during dynamic category evolution compared to other baselines, as measured by 1/ARI.
  • Infant Behavior Validation: When tested on internal vs. external cues (mimicking infant studies), configurations show poor performance with isolated cues but strong performance with combined cues, mirroring infant-like behavior and validating their capture of holistic perceptual information.

Limitations and Future Work

  • Lack of Neural Realism: The authors acknowledge that the framework is a computational model and lacks direct neural-level biological realism. Future work could explore Hebbian plasticity and neural architectures that implement configuration dynamics.
  • Scalability: Testing on larger datasets and broader cognitive domains beyond vision is suggested for future work.
  • Developmental Modeling: Capturing how cognitive abilities emerge over time is another area for future research.
  • Applications: The authors suggest potential applications in educational technology, human-computer interfaces, developmental robotics, and as foundational AI components. They also propose exploring neural implementation and online learning for more adaptive systems.

Summary updated

Abstract

Infants discover categories, detect novelty, and adapt to new contexts without supervision-a challenge for current machine learning. We present a brain-inspired perspective on configurations, a finite-resolution clustering framework that uses a single resolution parameter and attraction-repulsion dynamics to yield hierarchical organization, novelty sensitivity, and flexible adaptation. To evaluate these properties, we introduce mheatmap, which provides proportional heatmaps and reassignment algorithm to fairly assess multi-resolution and dynamic behavior. Across datasets, configurations are competitive on standard clustering metrics, achieve 87% AUC in novelty detection, and show 35% better stability during dynamic category evolution. These results position configurations as a principled computational model of early cognitive categorization and a step toward brain-inspired AI.
BibTeX
@misc{wang2025brain,
  author = {Juntang Wang and Yihan Wang and Hao Wu and Dongmian Zou and Shixin Xu},
  title = {{Brain-Inspired Perspective on Configurations: Unsupervised Similarity and Early Cognition}},
  year = {2025},
  eprint = {2510.19229},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.19229},
}
Figure