1. Using machine learning to identify task variables from neuronal activity
How animals combine sensory information with context and memory to make informed decisions? Modern neuronal recording techniques allow the simultaneous monitoring of the activity of hundreds of neurons during decision making providing us a unique opportunity to track information processing in the brain. A first step towards understanding how activity at a particular brain area contributes to behaviour is to identify the task variables represented by the neuronal activity. Thus, the goal of the project is to use modern machine learning techniques (mainly supervised learning: neural networks, Bayesian modelling, support vector machines, Generalised linear models, etc.) to discover the represented task variables and characterise the form of the representation. We use Ca imaging data acquired from mice navigating in virtual reality environments in order to obtain food reward. For further information, see our research page.
This project requires interest in machine learning and data analysis and good programming skills.
2. Neuronal manifolds underlying navigation in real and abstract spaces
Current experimental techniques allow the simultaneous recording of hundreds of neurons, but extracting the behaviourally relevant and easily interpretable variables represented by the data is challenging. The goal of the project is to use modern machine learning techniques (generative models and variational inference) to identify the low-dimensional nonlinear manifold in population activity that accurately captures variability in the data and provides interpretable latent factors. In this project we use both synthetic datasets and neuronal activity recorded from the hippocampus during various tasks including spatial navigation, as some of the latent factors (e.g. the spatial location of the animal) driving hippocampal neuronal activity are relatively well-known. For further information, see our research page.
The project requires interest in deep learning and strong mathematical background.
3. Network dynamics underlying sequential activity in the hippocampus
The hippocampus is a brain area critically involved in model based planning. Hippocampal neurons represent behaviourally relevant events as temporally compressed neuronal activity sequences. Interestingly, these sequences can be replayed during offline periods (sleep or immobility) either in a forward or in a reverse direction. It has been speculated that forward and reverse sequences has a distinct computational function (planning versus reinforcement learning), but the critical features of the network structure responsible for this flexibility is unknown. The goal of the project is to analyse experimental data in order to identify differences in the network dynamics during forward and reverse sequences and build a neuronal network that can generate sequential activity consistent with the experimental data.
The project is recommended for those who are interested in neuronal networks and data analysis and requires expertise in linear algebra and excellent programming skills.
4. The role of dendritic spikes in learning and memory
With the help of their excitable dendrites, cortical neurons implement complex nonlinear input-output transformations, similar to signal processing in a feed-forward neuronal network. Dendritic excitability is mainly manifested in the generation of dendritic spikes, but the role of these spikes in behaviourally relevant computations is not well known. The focus of the project are the various forms of Ca-spikes recorded in CA3 pyramidal neurons. To understand how they contribute to circuit computation we can take either a bottom up or a top down approach. The bottom up approach starts by building a detailed biophysical model where the natural synaptic inputs triggering Ca spikes can be studied. In the top-down approach we build a recurrent network implementing episodic memory and show that certain properties of the Ca spikes are ideal for separating information storage from recall in the network. For further information, see our research page.
The project is recommended to those interested in neuroscience and either in single neuron biophysics or neuronal network dynamics. The project requires good programming skills.
5. Automatic experimental training system for rodents: building the setup and analysing the behavioural data
To understand how the brain can solve interesting computational problems we often need to perform animal experiments where we can record the activity of neurons during task execution. In many case to most time consuming step is to train the animals to perform the task. The goal of the project is to develop an automatic behavioural training apparatus, where mice can execute simple tasks for their daily water reward. It is possible to contribute both at the hardware side (improving the existing design and adding new sensors and actuators), the software (controlling the system) or the data analysis and experimental design.
The project is recommended for those who like working with hardware and build real systems interacting with agents or those who are interested in theory of learning and analysing behavioural data.
Please email me if you are interested.