Representing Part-Whole Hierarchy with Nested Neuronal Coherence

Hao Zheng

Keywords: Part-Whole Relationship, Neural Syntax, Cell Assembly, Nested Oscillation, Neuronal Coherence, Object-Centric Representation, Binding Problem, Hierarchical Grouping, Structured Representation Learning, Spiking Neural Network, Visual Perception, Cortical Computation, Cortical Column, Attractor Network, NeuroAI, Hybrid Approach

TL;DR: This paper introduces a brain-inspired approach for representing objects that intrinsically have a part-whole hierarchy, with continuous neuronal coherence instead of discrete slots, concistent with the neural syntax hypothesis in neuroscience.

Abstract: Human vision flexibly extracts part-whole hierarchy from visual scenes. However, representing such hierarchical structure is a key challenge for neural networks. Most machine learning efforts addressing this issue have focused on slot-based methods, which may be limiting due to their discrete nature and difficulty to express uncertainty. Inspired by how neural syntax is organized in the brain, this paper presents a framework to represent the hierarchical part-whole relationship through hierarchically nested neuronal coherence, which has a continuous and distributed nature. At implementation level, we further developed a cortical-inspired hybrid model, the Composer, which dynamically achieves the emergent nestedness given images. To evaluate the emergent hierarchical structure, 4 synthetic datasets and 3 quantitative metrics are invented, which showed its ability to parse a range of scenes of different complexities. We believe this work, from representation, implementation to evaluation, advances a new paradigm for developing human-like vision in neural network models.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 9008