Multiple Projects in Computational Neuroscience

We're offering several master's thesis projects at the intersection of mechanistic interpretability, cognition, reinforcement learning, and efficient vision models.

The full project descriptions are available here.

Available projects

• - Can you beat a language model?

  Design minimal probabilistic-reasoning tasks, compare humans and language models, and analyze how behavior reveals latent reasoning strategies.

• - Continuous behavior clustering using Recurrent Neural Networks

  Build an RNN-based pipeline for continuous behavior clustering and study latent temporal structure in biological datasets without imposing discrete states a priori.

• - Understanding reasoning dynamics in Large Language Models

  Analyze activations and attention patterns in LLMs on psychophysics-inspired tasks to identify internal mechanisms related to belief, uncertainty, memory, and reasoning.

• - Adaptive Computation in Reinforcement Learning

  Implement and evaluate adaptive world-model rollout algorithms for reinforcement learning in complex environments.

• - Recursive Vision Models

  Validate biologically inspired recurrent or linear-attention vision architectures on modern visual reinforcement learning benchmarks.

General profile

Most projects require solid Python skills. Depending on the topic, experience with PyTorch, reinforcement learning, recurrent neural networks, or language models is helpful but not mandatory.

If you're interested contact Gonçalo Guiomar via email.