Schedule

Studying the molecular mechanisms underlying synaptic transmission in the brain is challenging in part because synapses are tiny (less than a micron), exhibit a wide range of shapes and internal structures, undergo activity-dependent plasticity, and their molecular components are dynamic. To better understand the mechanisms responsible of the onset of neurodegenerative disorders we must be able to observe the molecular dynamics and interactions of synaptic proteins at their scale: the nanoscale. We use quantitative super-resolution microscopy techniques to characterize molecular interactions inside neuronal cells with unprecedented spatiotemporal resolution. These techniques come with several layers of complex­ity in their implementation. This has limited their adoption as well as their adaptability to multi-color, multi-modal, and long-term imaging in living neurons. Developing machine learning (ML) assisted frameworks for optical nanoscopy allows real-time optimization of multi-modal live-cell imaging at the nanoscale as well as for quantitative high throughput image analysis. We develop ML-assisted microscopy to characterize activity-dependent remodelling of neuronal proteins. Our ML approaches not only automate image analysis but also enable multi-dimensional analysis of synaptic nanostructures and of the heterogeneity of the synapse population.

Unlike robots, humans learn, adapt and perceive their bodies by interacting with the world. Discovering how the brain represents the body and generates actions is of major importance for robotics and artificial intelligence. Recent advances in computational neuroscience have revealed principles that may be useful for current challenges in robotics, where adaptation to uncertain, complex and changing environments plays a major role. In this talk, I will describe the predictive processing framework extended to generative models of action. I will showcase, through experiments with humanoid and industrial robots, how we can use deep active inference to improve estimation, control, learning and planning. Along with the experiments, I will discuss some interesting findings that may reveal significant mechanisms of how we perceive and act with our bodies, which may be crucial for achieving human-like body intelligence in artificial agents.

Artificial neural networks are becoming more common as modeling frameworks for explaining the underlying function in neuronal circuits from single neurons to cortical regions consisting of millions of neurons. The neuronal fidelity of these models are often assessed through measures of prediction power over unseen stimuli. While their capacity for predicting neural responses in the brain is remarkable, their neuroscience applications beyond response prediction have been viewed with skepticism. This skeptic view is, at least partly, due to the inherent structural complexity of deep neural networks that makes it difficult to reveal the knowledge embedded in the connection weights of these networks. In this talk, I will discuss recent work on using predictive neural network models of the visual cortex to 1) scrutinize established theories of neuronal function in well-studied cortical regions of the human visual cortex and; 2) provide a more precise explanation of the function subserved by these areas. These results further highlight the crucial role of computational models towards a more precise understanding of the neural circuits underlying visual processing.

tbd