Georgia Institute of Technology College of Sciences

Nonconvex optimization naturally arises in many machine learning problems. Machine learning researchers exploit various nonconvex formulations to gain modeling flexibility, estimation robustness, adaptivity, and computational scalability. Although classical computational complexity theory has shown that solving nonconvex optimization is generally NP-hard in the worst case, practitioners have proposed numerous heuristic optimization algorithms, which achieve outstanding empirical performance in real-world applications.To bridge this gap between practice and theory, we propose a new generation of model-based optimization algorithms and theory, which incorporate the statistical thinking into modern optimization. Specifically, when designing practical computational algorithms, we take the underlying statistical models into consideration. Our novel algorithms exploit hidden geometric structures behind many nonconvex optimization problems, and can obtain global optima with the desired statistics properties in polynomial time with high probability.

Since its inception, neuroscience has largely focused on the neuron as the functional unit of the nervous system. However, recent evidence demonstrates that populations of neurons within a brain area collectively show emergent functional properties ("dynamics"), properties that are not apparent at the level of individual neurons. These emergent dynamics likely serve as the brain’s fundamental computational mechanism. This shift compels neuroscientists to characterize emergent properties – that is, interactions between neurons – to understand computation in brain networks. Yet this introduces a daunting challenge – with millions of neurons in any given brain area, characterizing interactions within an area, and further, between brain areas, rapidly becomes intractable.I will demonstrate a novel unsupervised tool, Latent Factor Analysis via Dynamical Systems ("LFADS"), that can accurately and succinctly capture the emergent dynamics of large neural populations from limited sampling. LFADS is based around deep learning architectures (variational sequential auto-encoders), and builds a model of an observed neural population's dynamics using a nonlinear dynamical system (a recurrent neural network). When applied to neuronal ensemble recordings (~200 neurons) from macaque primary motor cortex (M1), we find that modeling population dynamics yields accurate estimates of the state of M1, as well as accurate predictions of the animal's motor behavior, on millisecond timescales. I will also demonstrate how our approach allows us to infer perturbations to the dynamical system (i.e., unobserved inputs to the neural population), and further allows us to leverage population recordings across long timescales (months) to build more accurate models of M1's dynamics.This approach demonstrates the power of deep learning tools to model nonlinear dynamical systems and infer accurate estimates of the states of large biological networks. In addition, we will discuss future directions, where we aim to pry open the "black box" of the trained recurrent neural networks, in order to understand the computations being performed by the modeled neural populations.pre-print available: lfads.github.io [lfads.github.io]

TBA

One way to analyze a module is to consider its minimal free resolution and take a look its Betti numbers. In general, computing minimal free resolution is not so easy, but in case of some certain modules, computing the Betti numbers become relatively easy by using a Hochster's formula (with the associated simplicial complex. Besides, Mumford introduced Castelnuovo-Mumford regularity. The regularity controls when the Hilbert function of the variety becomes a polynomial. (In other words, the regularity represents how much the module is irregular). We can define the regularity in terms of Betti numbers and we may see some properties for some certain ideals using its associated simplicial complex and homology. In this talk, I will review the Stanley-Reisner ideals, the (graded) betti-numbers, and Hochster's formula. Also, I am going to introduce the Castelnuovo-Mumford regularity in terms of Betti numbers and then talk about a useful technics to analyze the Betti-table (using the Hochster's formula and Mayer-Vietories sequence).