Description
- Abstract:
- Objective: This thesis explores the viability of biologically constrained artificial-intelligence architecture based on Numenta, Inc.’s Thousand Brains Theory (TBT) for object recognition, and initial steps into hierarchy. Methods: Learning modules (LMs) emulating layers II/III, IV, and VI were implemented with Monty v0.0.3 and embedded in physics-enabled Habitat-Sim scenes. Each LM received sensations from either a distant “eye” agent or a surface “finger” agent and encoded features as distributed graphs. Four supervised architectures were trained on six Yale–CMU–Berkeley objects for one epoch (14 canonical views) and 500 sensorimotor steps per view: (1) single-LM distant; (2) single-LM surface; (3) two horizontally connected LMs; and (4) five horizontally connected LMs, making decisions from 3 LMs agreeing through a voting system. Complementary experiments evaluated a two-level hierarchy and an unsupervised agent that updated its memory after every encounter. Results: After one epoch on each object, a single LM learned a graph model of each object and later reidentified learned items; though thin objects, like cutlery, caused ambiguities in both learning and inference. Surface agents produced smoother graphs and faster inference than distant agents. Although it was expected that surface agents would model full transparent objects and not just the visible parts, this did not occur. Lateral connections (a voting system) delivered benefits: faster inference depending on the number of LMs that have to agree, and more robustness from multiple LMs agreeing. Extending episodes to 1,000–1,500 steps refined graphs, providing slightly faster inference but no accuracy improvement. The unsupervised agent progressively updated its models over three encounters, and the hierarchical architecture produced sparser models on the higher-level LM, hinting at abstraction. Conclusions: These experiments show that sensory-motor architectures grounded on the Thousand Brains Theory pose a promising direction for advancing AI into real-world interactions, being able to create accurate object models from minimal experience and refine them online. It also showed advantages of a voting system, as well as what could be the early stages of abstraction when applying hierarchy structures. TBT motivates further investigation into learning and inference of transparent objects, as well as compositional objects and scenes.
- Notes:
- Thesis (Sc. M.)--Brown University, 2025
Citation
Enriquez, Santiago,
"Exploring Models of Cortical Columns for Biologically Constrained AI in Object Recognition"
(2025).
Biomedical Engineering Theses and Dissertations.
Brown Digital Repository. Brown University Library.
https://repository.library.brown.edu/studio/item/bdr:fbhc6dpp/
Relations
Collection:
-
Biomedical Engineering Theses and Dissertations
Theses and Dissertations for the Biomedical Engineering department....