Master projects in the Neuro TMA Group
1. Generalized Latent Equilibrium: Applications in Robotics
The NeuroTMA team has developed a novel framework (Generalized Latent Equilibrium) for bio-plausible neural networks and synaptic learning rules described by locally coupled differential equations (Backpropagation through space, time and the brain). These equations describe physical systems capable of on-line learning complex signal processing tasks, and this Master's project will put this theory to a new and challenging test.
The project will focus on experimentally validating the GLE framework in a real-world setup involving continuous learning in real time. The tasks are as follows: first, prepare an experimental setup for balancing a simple or double cart-pole; second, construct appropriate network models and learning procedures based on the GLE framework (supervised and reinforcement learning); third, study the results with respect to different experimental parameters. The ideal candidate will be passionate about computational neuroscience and AI, eager to work with real-world setups, in addition to having the required mathematical knowledge and coding skills.
Feel like you're up to the challenge? Then drop us an email at xavier.lesage@unibe.ch, mihai.petrovici@unibe.ch and federico.benitez@unibe.ch.
2. Generalized Latent Equilibrium: Oscillatory Neurons
The NeuroTMA team has developed a novel framework (Generalized Latent Equilibrium) for bio-plausible neural networks and synaptic learning rules described by locally coupled differential equations (Backpropagation through space, time and the brain). These equations describe physical systems of leaky integrator neurons capable of on-line learning complex signal processing tasks, and this Master's project will widen the scope of this theory.
The project will focus on extending the GLE framework to networks of oscillatory neurons. Not only can biological networks exhibit oscillatory behavior, but oscillator networks also represent a promising new venue for physical AI. Incorporating these dynamics has the potential of significantly increasing the model's computational capabilities. This oscillatory behavior can be modelled either by assigning a second internal degree of freedom to leaky integrators, or by considering neurons with complex-valued internal representations. These modifications are expected to increase the memory capacity and computational power of GLE networks.
The tasks are as follows: first, extend the GLE theory to the case of oscillatory neurons; second, construct appropriate network models and learning procedures (supervised learning); third, benchmark the new models on challenging spatiotemporal learning problems. The ideal candidate will be passionate about computational neuroscience and AI, eager to expand their modelling toolset, in addition to having the required mathematical knowledge and coding skills.
Feel like you're up to the challenge? Then drop us an email at mihai.petrovici@unibe.ch and federico.benitez@unibe.ch.
3. Generalized Latent Equilibrium: Recurrent Networks
The NeuroTMA team has developed a novel framework (Generalized Latent Equilibrium) for bio-plausible neural networks and synaptic learning rules described by locally coupled differential equations (Backpropagation through space, time and the brain). These equations describe physical systems of leaky integrator neurons capable of on-line learning complex signal processing tasks, and this Master's project will widen the scope of this theory.
The project will focus on extending the GLE framework to to more complex architectures and tasks, including sequence generation with autoregressive networks, long-term memory and context with densely recurrent networks and state space models, and latent variable inference with temporal autoencoders. Recurrent connections are expected to enhance the temporal processing capabilities of GLE networks, and would be an important step towards impactful applications in robotics and brain-computer interfaces.
The ideal candidate will be passionate about computational neuroscience and AI, eager to expand their modelling toolset, in addition to having the required mathematical knowledge and coding skills.
Feel like you're up to the challenge? Then drop us an email at mihai.petrovici@unibe.ch and federico.benitez@unibe.ch.
4. Generalized Latent Equilibrium for Brain-Computer Interfaces
The NeuroTMA team has developed a novel framework (Generalized Latent Equilibrium) for bio-plausible neural networks and synaptic learning rules described by locally coupled differential equations (Backpropagation through space, time and the brain). These equations describe physical systems of leaky integrator neurons capable of on-line learning complex signal processing tasks, and this Master's project will put this theory to a high-impact, challenging test.
The project will focus on implementing the GLE architecture for training Brain-Computer Interfaces (BCIs), that decode brain activity in persons that have lost some or most of their motor abilities following diseases or accidents. These systems have the potential to substantially increase the quality of life of those affected. Thanks to our colleagues at the University of Pittsburgh, we have access to high quality real-world training data. By harnessing the real-time learning capabilities of GLE, we will develop training protocols of BCI interfaces that can be much more time efficient and precise compared to current methods. The ideal candidate will be passionate about computational neuroscience and AI, eager to work with real-world setups, in addition to having the required mathematical knowledge and coding skills.
Feel like you're up to the challenge? Then drop us an email at mihai.petrovici@unibe.ch and federico.benitez@unibe.ch.
