Personal projects

Research

Active

Hyperparameters and Representational Geometry in Neural Networks

Do the training choices you make when building a neural network affect what it learns to represent internally? Yes, substantially, even among networks that perform comparably on the same task. This has direct implications for research that uses neural networks as models of the brain.

pythonneuroscienceRSAdeep learningdata analysis
Complete

Hierarchical Reinforcement Learning as Creative Problem Solving

A theoretical and experimental project exploring the relationship between hierarchical reinforcement learning, the psychology of insight, and computational creativity.

reinforcement learningcreativitycognitive sciencepythondeep learning
Complete

Modeling Cognitive Control: RNNs and Representational Similarity Analysis

A published computational neuroscience study using RSA to compare how recurrent neural networks and the human anterior cingulate cortex represent hierarchically structured task sequences. Networks with explicit goal units better account for the representational geometry of human ACC.

pythonneurosciencefMRIdeep learningdata analysisRSA
On hold

Atari ACC Model

A computational model of the anterior cingulate cortex, developed during a postdoc in Clay Holroyd's lab at Ghent University. ACC modules trained on top of a deep RL agent playing Q-bert learn to track and represent the agent's internal state — demonstrating that a separate observer module can recover structured, abstract representations of a complex neural system in action.

pythondeep learningreinforcement learningneuroscienceon hold

CV

Thomas R. Colin, PhD

Publications, professional background, and downloadable PDF.