We consider the problem of reconstructing a hidden assignment of random variables (x_1,…, x_n) sitting on the nodes of a graph, given noisy observations of pairs (x_u, x_v) for every edge (u,v) of the graph. Such problems have been extensively studied when the underlying graph is ‘mean-field’ (e.g., the complete graph, Erdos-Renyi, random regular,…), in which case the existence of a ``possible-but-hard’’ phase where reconstruction is possible but computationally difficult is ubiquitous.

In contrast, the picture is dramatically different when the graph is amenable. I will represent a generic result about the optimality of local algorithms for computing marginals under a somewhat strong model of side information. Next, I will focus on Z_2 synchronization (aka planted Ising model) on the Euclidean lattice with weaker side information, and discuss a renormalization-based procedure for reconstructing the assignment.

Joint work with Andrea Montanari.

In this talk I will consider extreme behavior of the extremal eigenvalues of white Wishart matrices, which play an important role in multivariate analysis. I will focus on the case when the dimension of the feature p is much larger than or comparable to the number of observations n, a common situation in modern data analysis. I will discuss asymptotic approximations for the tail probabilities of the extremal eigenvalues. In addition, I will discuss the construction of an efficient Monte Carlo importance sampling algorithm to estimate the tail probabilities. Simulation results show that our method has the best performance among known approximation approaches, and furthermore provides an efficient and accurate way for evaluating the tail probabilities in practice. Based on joint work with Tiefieng Jiang and Gongjun Xu.

We propose a strategy for improving the existing methods for solving synchronization problems that arise from various computer vision tasks. Specifically, our strategy identifies severely corrupted relative measurements based on cycle consistency information. To the best of our knowledge, this paper provides the first exact recovery guarantees using cycle consistency information. This result holds for a noiseless but corrupted setting as long as the ratio of corrupted cycles per edge is sufficiently small. It further guarantees linear convergence to the desired solution. We also establish stability of the proposed algorithm to sub-Gaussian noise.

How many critical points does a smooth random function on a high-dimensional space typically have at a given height? how are their distances distributed? what is the volume or geometry of the level sets? can we design efficient optimization algorithms for the random function? For the spherical spin glass models, those questions are closely related to the structure of the Gibbs measures, which have been extensively studied in physics since the 70s.

I will start with an overview of the celebrated Parisi formula and ultrametricity property. I will then describe an alternative method to analyze the Gibbs measure using critical points, in the setting of the pure spherical models. Finally, I will explain how the latter can be extended to all spherical models, using another (soft) geometric approach, while at the same time making rigorous and generalizing the famous Thouless-Anderson-Palmer approach from physics.

We consider (volume-)frozen percolation on the triangular lattice. The model can be described informally as follows. Fix a large integer $N$. Initially, all vertices are vacant. We let clusters grow (vertices become occupied) as long as their volume is strictly smaller than $N$, and they stop growing (they "freeze") when their volume becomes at least $N$. A vertex $v$ is frozen if it belongs to an occupied cluster with volume at least $N$.

In this model, there exists a sequence of "exceptional scales" $(m_k(N))$: roughly speaking, if we consider frozen percolation in a box of side length $m_k(N)$, then as $N\to\infty$, the probability that $0$ is frozen in the final configuration is bounded away from $0$; while if we consider the process in a box of side length that is far from $m_k(N)$ and $m_{k+1}(N)$ (but between them), then as $N\to\infty$, the corresponding probability will go to $0$. The limiting exception scale, $m_\infty(N)$, is not studied and almost nothing is known. In an ongoing project with Pierre Nolin, we show that if we consider the process in a box of side length $m_\infty(N)$, then there are "avalanches" of freezings: the number of frozen circuits surrounding the origin divided by $\log\log{N}$ converges to an explicit constant in probability. If time allows, I will also talk about the analogous result in the forest fire process.

Modern data are usually high-dimensional with noise and corruption. A useful representation of data has to be robust and address the data structure. In this talk, I will first present a class of robust models called the scattering transform that can be used to generated features from graph data. In graph scattering transforms, the representation is generated in an unsupervised manner based on graph wavelets. It is approximately invariant to permutations and stable to signal or graph manipulations. Numerical results show that it works effectively for classification and community detection problems. Next, I will address how the structure of data can be found using autoencoders. Indeed, in the framework of autoencoders, graph scattering transform can be applied to the important task of graph generation. Specifically, I will illustrate its application in generating molecular samples.

The evolution of a deep neural network trained by the gradient descent can be described by its neural tangent kernel (NTK) as introduced by Jacot et al., where it was proven that in the infinite width limit the NTK converges to an explicit limiting kernel and it stays constant during training. The NTK was also implicit in many other recent papers. In the overparametrization regime, a fully-trained deep neural network is indeed equivalent to the kernel regression predictor using the limiting NTK. And the gradient descent achieves zero training loss for a deep overparameterized neural network. However, it was observed by Arora et al. that there is a performance gap between the kernel regression using the limiting NTK and the deep neural networks. This performance gap is likely to originate from the change of the NTK along training due to the finite width effect.  The change of the NTK along the training is central to describe the generalization features of deep neural networks.In the work, we study the dynamic of the NTK for finite width deep fully-connected neural networks. We derive an infinite hierarchy of ordinary differential equations, the neural tangent hierarchy (NTH) which captures the gradient descent dynamic of the deep neural network. Moreover, under certain conditions on the neural network width and the data set dimension, we prove that the truncated hierarchy of NTH approximates the dynamic of the NTK up to arbitrary precision. This is a joint work with Horng-Tzer Yau.