Integrated Calendar

 In recent years, groundbreaking research revealed many surprising roles for astrocytes in addition to their well characterized supportive roles, in modulating neuronal activity and even behavior. I will talk on one hand about manipulating astrocytes to alter neuronal activity and behavior and on the other hand about imaging astrocytes in behaving animals.We chronically imaged CA1 astrocytes using 2-photon microscopy when head-fixed mice were trained mice to run on a linear treadmill and proceed in a virtual environment to obtain water rewards. We found that astrocytic activity persistently ramps towards the reward location in a familiar environment. When the reward location was changed in the same environment or when mice were introduced to a novel context, the ramping was not apparent. Using linear decoders, we accurately reconstructed mice location trajectories in a familiar environment from astrocyte activity alone. This is the first indication that astrocytes can encode position related information in learnt spatial contexts, thus broadening their known computational abilities, and their role in cognitive functions.To directly and specifically modulate astrocytic activity we employed a chemogenetic approach: We expressed the Gq-coupled designer receptor hM3Dq in astrocytes, which allowed their time-restricted manipulation, and discovered that astrocytic activation is sufficient to induce de-novo long term potentiation, enhance memory allocation and augment memory recall on the following day. I will talk about these published works and about non-published results on the role of astrocytes in Alzheimer's disease

Understanding how cortical circuits give rise to perception-allowing us, for example, to hear and see the world-remains a central challenge in neuroscience. The application of concepts from cognitive science, such as Representational Similarity Analysis, has proven valuable for interpreting large-scale neuronal recordings, including in rodent models. In this work, I present recent efforts from our laboratory to characterize the structure of auditory representations in the mouse cortex and demonstrate how these representations can be used to predict behavioral phenomena such as stimulus generalization and perceptual choice biases. Moreover, leveraging neuronal activity recordings at single-cell resolution, I describe our findings on the circuit mechanisms that organize sound-evoked activity into structured representational maps and maintain their integrity in the face of perturbations, including synaptic volatility and neuronal loss.

After an earthquake occurs, slip models of the event may be estimated from geodetic observations. This process generally requires static coseismic Green's functions, which must be calculated via a forward model which includes an approximation of the material properties, topography, and fault geometry in the region of interest. Until recently, the lack of seafloor geodetic instrumentation and the use of unrealistically simple forward models have resulted in poor resolution of near-trench slip in subduction zone settings.  In this talk, I will present an investigation into the effects of 3D structure, particularly topography, on forward models of earthquake deformation and on earthquake static slip estimates. I will show that models which neglect 3D structure yield inaccurate estimates of near-trench slip, particularly when seafloor geodetic data are utilized in the inversion. 

In the theory part of the talk, we study the statistical performance of maximum likelihood estimation (MLE) and, more generally, empirical risk minimization (ERM). While MLE is known to be minimax-optimal for low-complexity models, classical work showed that it can be suboptimal over “large” function classes, though the canonical examples are somewhat pathological. First, we develop a technique for detecting and quantifying the suboptimality of ERM in regression over high-dimensional nonparametric classes. Second, we show that the variance term of ERM procedures is always upper-bounded by the minimax rate, implying that any minimax suboptimality must arise from bias. Third, we present the first minimax-optimal estimator for convex regression in all dimensions with a polynomial runtime in the sample size. If time permits, we also discuss connections between the local theory of Banach spaces and minimum-norm interpolators, building on an approach initiated by Maurey and Pisier. In the applied part of the talk, we propose an explanation for the empirical success of test-time training (TTT) in foundation models, which we primarily validate through experiments with sparse autoencoders (SAEs). TTT identifies the training points most similar to a given evaluation point and improves predictions by locally adapting the model to this selected neighborhood. Although TTT has been studied previously, only recently has it been shown to deliver substantial gains in foundation models across domains such as image generation, control, and language modeling.