My research focused on understanding the mechanisms of intelligence in biological and artificial systems. With a background in both neuroscience and AI, I specifically focus on investigating how biological and artificial neural networks acquire knowledge and contribute to solving complex tasks. Artificial neural networks take direct inspiration from the brain and have been largely successful at reaching above-human performance in many tasks in recent years. However, there are still unresolved questions in the brain, which once solve, could significantly improve artificial systems. These include the following challenges.
Both artificial and biological neural networks implement neurons as units of information transfer, and synaptic weights as learnable parameters. These synaptic weights can be strengthened or weakened, and can be used to extract knowledge from input data (the outside world). Current artificial neural networks use backpropagation of signed error as a learning signal over the course of training. While effective, it does present some limitations and more importantly: this is not how biological neural networks learn. Understanding how learning occurs in the brain remains a crucial question.
Previous work in neurophysiology shows that activity can drive plasticity, however these findings come from in vitro (e.g. brain slices), require many stimuli, and likely do not reflect in vivo physiological mechanisms.
In this research axis, I explore learning in neural network as an interaction of the following:
Evolutionary priors: without any external signals, biological neural networks have strong evolutionary priors that are usually ignored in artificial models
Neuromodulators: such as dopamine and serotonine, which are not sufficient learning signals on their own (not local enough) but can provide interesting reinforcement feedback while interacting with the world
Inhibitory interneurons: which can gate neural activity, are typically ignored in artificial neural networks, and may play a central role in learning
Top-down feedback: originating from higher-order regions, may provide target representations and/or predictions during learning
Ultimately, learning relies on these different signals which are provided either by interaction with the environment (at various timescale) or through observation
Simplified diagram of the bi-directional hierarchical microcircuit in the cortex. Neurons in layer II/III tend to project to deeper, higher-order neurons in the hierarchy, while neurons in layer V tend to send feedback projections to lower-order neurons.
My research is at the intersection between neuroscience and artificial intelligence and aims at establishing how networks of biological neurons coordinate their activities to propagate errors and update synapses, in turn supporting learning and higher-order cognition.
As we explore the world, we collect information through our sensory modalities (sound, vision, touch, ...). These online signals help us build a model of the external world, and understand the rules that govern it. This internal knowledge can then be leverage to make predictions and hypotheses about the future, which are then tested against incoming sensory signals. This loop of testing predictions based on our internal knowledge could be essential to cognition, and more precisely, learning of new information. A particular brain region, called the hippocampus, contains neurons that constantly represent past and future events in space and time, and could support this function. A portion of my research has been to record the activity of these neurons and understand how they could contribute to learning and memory.
Neurons in the hippocampus tend to fire in a sequential order that depends on connectivity and is established at birth. When animals explore environments, these neurons fire sequentially (~8 times per second) with respect to local oscillations, and represent past, current, and potential future states.
Learning involves updating synapses, the points of contact between neurons. While this phenomenon has been widely described in vitro, the exact mechanisms of synaptic updates in vivo remain misunderstood. One hypothesis is that, when predictions are validated by incoming sensory stimuli, neurons can be depolarized (activated) beyond their typical state, in turn facilitating the formation of new synapses (and enhancing existing ones). The exact mechanisms behind this phenomenon remain unknown, but could involve a local microcircuit including inhibitory interneurons. The aim of this research axis is to test local learning rules using a deep learning framework in silico, and test these hypotheses in vivo using voltage imaging of excitatory and inhibitory neurons within microcircuits.
Top: pyramidal neurons, which are most prevalent in the cortex and hippocampus, integrate information at their based (green) and apical (red) dendrites. Apical dendrites are thought to integrate predictions from higher-order neurons, whereas basal dendrites would integrate sensory signals predominantly. Only when sensory signals are predicted by higher-order neurons, pyramidal neurons emit bursts of action potentials.
Bottom: as animals explore environments, these bursts could in turn trigger plasticity events, update synapses, and induce long-term changes in representations.
We spend an enormous amount of our life sleeping (and dreaming). Yet, the exact function of sleep remains largely misunderstood. Using large scale recording of thousands of neurons during sleep and awake experience, one of my objectives is to determine the computational role of sleep in cognition.
Neurons in the hippocampus and cortex establish their connectivity at birth (structural priors). These priors can dictate the expression of internal sequences that do not depend on sensory inputs or experiences. Experiences acquired during awake exploration can be replayed during subsequent REM sleep episodes. Alternatively, internal sequences present during REM sleep can be re-expressed in awake states (preplay).
As the world population continues to age, more and more people are affected by Alzheimer's disease and associated dementia. This prompts the need for early diagnostic, and non-invasive targeted treatments. Using machine learning, my research aims to discover features in neural recordings (single cell and local field potential) to predict cognitive decline as early as possible. Additionally, I have been developing red-shifted optogenetics, where light can pass through the skull to activate transfected neurons deep in the brain without the need for invasive brain implants.
Left: Typical optogenetic actuators (e.g. ChR2) are most activated by blue light (~450nm) which necessitates light delivery close to targets (using fiber optic implants). On the other hand, red-shifted actuators (e.g. ChrimsonR) can be activated through the cranium using red lasers, thus removing the need for invasive brain implants.
Right: using deep learning models, neural activities from healthy and Alzheimer's disease conditions can be encoded in low-dimensional manifolds. Subsequently, neural signatures and predictors can be extracted from these low-dimensional embeddings.
Using open-source miniaturized head-mounted microscopes (miniscopes), we can record activity from large assemblies (>1,000) neurons over extended periods of time (>2 years) for affordable costs. Here is an example of a few neurons in hippocampal subregion CA1 using 1-photon in vivo imaging.
Optogenetics allow targeted dissection of neural circuits. Using viral transfections with cell-type specificity (e.g. targetting inhibitory interneurons only). From Etter et al., 2023b.
Example of a 16 channels silicon probe recording in the hippocampus during a sharp-wave ripple event. The darker layer corresponds to the pyramidal cell layer (corresponding local field potential signal in green). From Etter et al., 2019.
Whole cell patch clamp electrophysiology of a hippocampal mossy cell, which gives access to synaptic and intracellular mechanisms. Image: Etter et al., 2014.
Blake Richards (McGill University/MILA): bioplausible deep learning learning
Daniel Aharoni (UCLA): miniscope hardware and software development
Massimo Avoli (McGill University): therapeutical uses of optogenetics in epilepsy
Keith Murai (McGill University): gene expression in interneurons