arXiv subjects

Tianze Jiang

Publications and source records attributed to Tianze Jiang.

7 recordsLinked to original sources

arXiv2026

Bayesian Inference with Shaped Deep Non-linear MLPs

A central aim of deep learning theory is to characterize how neural networks make predictions in the regime of simultaneously large model and training set size. Since the limits of diverging number of model parameters and dataset size do not commute it is not clear a priori what limits exist. In this work, we shed new light on these questions by studying Bayesian inference in deep non-linear MLPs in the regime where the number of training samples ($P$), the input dimension ($N_0$), the hidden layer width ($N$), and the number of hidden layers ($L$) can all be large. We build on the Neural Covariance SDE (Li et al., 2022) to analyze predictive posteriors in the regime where $LP/N\inΘ(1)$, playing the role of an effective network depth. Our framework covers both smooth and ReLU activation functions and applies to arbitrary temperature. We find to first order in $LP/N$ a simple criterion for which data generating processes benefit from depth in the sense that larger $LP/N$ increases the Bayesian model evidence. We also give a novel derivation of a prior result from the physics literature that at least to first order in $LP/N$, the Bayesian predictive posterior is remarkably simple and is simply equivalent to that of a data-dependent kernel method.

math.ST↗

arXiv2026

Hyperparameter Transfer with Mixture-of-Expert Layers

Mixture-of-Experts (MoE) layers have emerged as an important tool in scaling up modern neural networks by decoupling total trainable parameters from activated parameters in the forward pass for each token. However, sparse MoEs add complexity to training due to (i) new trainable parameters (router weights) that, like all other parameter groups, require hyperparameter (HP) tuning; (ii) new architecture scale dimensions (number of and size of experts) that must be chosen and potentially taken large. To make HP selection cheap and reliable, we propose a new parameterization for transformer models with MoE layers when scaling model width, depth, number of experts, and expert (hidden) size. Our parameterization is justified by a novel dynamical mean-field theory (DMFT) analysis. When varying different model dimensions trained at a fixed token budget, we find empirically that our parameterization enables reliable HP transfer across models from 51M to over 2B total parameters. We further take HPs identified from sweeping small models on a short token horizon to train larger models on longer horizons and report performant model behaviors.

cs.LG↗

arXiv2025

Density estimation using the perceptron

We propose a new density estimation algorithm. Given $n$ i.i.d. observations from a distribution belonging to a class of densities on $\mathbb{R}^d$, our estimator outputs any density in the class whose "perceptron discrepancy" with the empirical distribution is at most $O(\sqrt{d / n})$. The perceptron discrepancy is defined as the largest difference in mass two distribution place on any halfspace. It is shown that this estimator achieves the expected total variation distance to the truth that is almost minimax optimal over the class of densities with bounded Sobolev norm and Gaussian mixtures. This suggests that the regularity of the prior distribution could be an explanation for the efficiency of the ubiquitous step in machine learning that replaces optimization over large function spaces with simpler parametric classes (such as discriminators of GANs). We also show that replacing the perceptron discrepancy with the generalized energy distance of Székely and Rizzo (2013) further improves total variation loss. The generalized energy distance between empirical distributions is easily computable and differentiable, which makes it especially useful for fitting generative models. To the best of our knowledge, it is the first "simple" distance with such properties that yields minimax optimal statistical guarantees. In addition, we shed light on the ubiquitous method of representing discrete data in domain $[k]$ via embedding vectors on a unit ball in $\mathbb{R}^d$. We show that taking $d \asymp \log (k)$ allows one to use simple linear probing to evaluate and estimate total variation distance, as well as recovering minimax optimal sample complexity for the class of discrete distributions on $[k]$.

math.ST↗

arXiv2025

Global Universality of Singular Values in Products of Many Large Random Matrices

We study the singular values (and Lyapunov exponents) for products of $N$ independent $n\times n$ random matrices with i.i.d. entries. Such matrix products have been extensively analyzed using free probability, which applies when $n\to \infty$ at fixed $N$, and the multiplicative ergodic theorem, which holds when $N\to \infty$ while $n$ remains fixed. The regime when $N,n\to \infty$ simultaneously is considerably less well understood, and our work is the first to prove universality for the global distribution of singular values in this setting. Our main result gives non-asymptotic upper bounds on the Kolmogorov-Smirnoff distance between the empirical measure of (normalized) squared singular values and the uniform measure on $[0, 1]$ that go to zero when $n, N\to \infty$ at any relative rate. We assume only that the distribution of matrix entries has zero mean, unit variance, bounded fourth moment, and a bounded density. Our proofs rely on two key ingredients. The first is a novel small-ball estimate on singular vectors of random matrices from which we deduce a non-asymptotic variant of the multiplicative ergodic theorem that holds for growing matrix size $n$. The second is a martingale concentration argument, which shows that while Lyapunov exponents at large $N$ are not universal at fixed matrix size, their empirical distribution becomes universal as soon as the matrix size grows with $N$.

math.PR↗

arXiv2024

Prediction from compression for models with infinite memory, with applications to hidden Markov and renewal processes

Consider the problem of predicting the next symbol given a sample path of length n, whose joint distribution belongs to a distribution class that may have long-term memory. The goal is to compete with the conditional predictor that knows the true model. For both hidden Markov models (HMMs) and renewal processes, we determine the optimal prediction risk in Kullback- Leibler divergence up to universal constant factors. Extending existing results in finite-order Markov models [HJW23] and drawing ideas from universal compression, the proposed estimator has a prediction risk bounded by redundancy of the distribution class and a memory term that accounts for the long-range dependency of the model. Notably, for HMMs with bounded state and observation spaces, a polynomial-time estimator based on dynamic programming is shown to achieve the optimal prediction risk Θ(log n/n); prior to this work, the only known result of this type is O(1/log n) obtained using Markov approximation [Sha+18]. Matching minimax lower bounds are obtained by making connections to redundancy and mutual information via a reduction argument.

math.ST↗

arXiv2023

Kernel-Based Tests for Likelihood-Free Hypothesis Testing

Given $n$ observations from two balanced classes, consider the task of labeling an additional $m$ inputs that are known to all belong to \emph{one} of the two classes. Special cases of this problem are well-known: with complete knowledge of class distributions ($n=\infty$) the problem is solved optimally by the likelihood-ratio test; when $m=1$ it corresponds to binary classification; and when $m\approx n$ it is equivalent to two-sample testing. The intermediate settings occur in the field of likelihood-free inference, where labeled samples are obtained by running forward simulations and the unlabeled sample is collected experimentally. In recent work it was discovered that there is a fundamental trade-off between $m$ and $n$: increasing the data sample $m$ reduces the amount $n$ of training/simulation data needed. In this work we (a) introduce a generalization where unlabeled samples come from a mixture of the two classes -- a case often encountered in practice; (b) study the minimax sample complexity for non-parametric classes of densities under \textit{maximum mean discrepancy} (MMD) separation; and (c) investigate the empirical performance of kernels parameterized by neural networks on two tasks: detection of the Higgs boson and detection of planted DDPM generated images amidst CIFAR-10 images. For both problems we confirm the existence of the theoretically predicted asymmetric $m$ vs $n$ trade-off.

stat.ML↗

arXiv2023

Detection-Recovery and Detection-Refutation Gaps via Reductions from Planted Clique

Planted Dense Subgraph (PDS) problem is a prototypical problem with a computational-statistical gap. It also exhibits an intriguing additional phenomenon: different tasks, such as detection or recovery, appear to have different computational limits. A detection-recovery gap for PDS was substantiated in the form of a precise conjecture given by Chen and Xu (2014) (based on the parameter values for which a convexified MLE succeeds) and then shown to hold for low-degree polynomial algorithms by Schramm and Wein (2022) and for MCMC algorithms for Ben Arous et al. (2020). In this paper, we demonstrate that a slight variation of the Planted Clique Hypothesis with secret leakage (introduced in Brennan and Bresler (2020)), implies a detection-recovery gap for PDS. In the same vein, we also obtain a sharp lower bound for refutation, yielding a detection-refutation gap. Our methods build on the framework of Brennan and Bresler (2020) to construct average-case reductions mapping secret leakage Planted Clique to appropriate target problems.

math.ST↗