Research Topics


Neuro-AI combines Neuroscience and Machine Learning. It aims to create artificial intelligence (AI) systems that mimic the functionality of the brain using biologically plausible networks. At the same time, it uses machine learning tools to advance neuroscience research, including the development of innovative technologies like brain-computer interfaces (BCI). In simpler terms, it’s about making AI more like our brains, and using AI to better understand our brains.

Predictive + Reinforcement Learning

This research focuses on the integration of predictive learning and reinforcement learning (RL) in neural networks. Predictive learning creates maps, or world models, of explored environments. These models, which capture the spatial and temporal relationships within the environment, are then used by RL to perform tasks.

The study has three objectives:

The ultimate goal is to understand how the brain combines predictive and RL-based models to solve tasks, enhancing our knowledge of neural computation and behavior. The focus is on the geometrical structure of neural representations. (Potential student projects listed here).

Brain Computer Interfaces

We’re creating tools for Brain-Computer Interfaces (BCIs), where an animal, like a mouse, is trained to perform tasks based on its neural activity. This could involve, for example, enhancing certain neurons’ activity. These are closed-loop cutting-edge experiments, where the animal’s performance is then reinforced with rewards or sensory feedback.  The tools we develop include decoders to predict, encode, and perturb neural activity. We also use advanced data analysis and simulations to understand how the recorded neural circuits learn the BCI task. (Potential student projects listed here).

Neural State Driven Dynamics

Neural dynamics can be seen as a series of neural states that unfold over time. Transitions between states are traditionally thought to be controlled by factors such as sensory inputs or internal drives. However, one key factor that can also control these transitions is neuromodulators. Neuromodulators are chemicals in the brain that can alter the activity of neurons and synapses. We are developing novel theorical and computational models to understand how neuromodulators can control the dynamic unfolding of neural states.

Significantly, machine learning algorithms have not yet adopted this dynamic control. While certain models like attractor models or LLMs can encapsulate elements of neural dynamics, they usually overlook the influence of neuromodulators. Exploring how to integrate these dynamics into machine learning models could potentially improve their dynamic characteristics, leading to increased flexibility and adaptability.

(Potential student projects listed here).

Previous Topics

Here is a video going over some older projects!