Searcharxiv⌕ Search

arXiv subjects

Xiaowu Dai

Publications and source records attributed to Xiaowu Dai.

40 records · Page 3Linked to original sources

Towards Theoretical Understanding of Large Batch Training in Stochastic Gradient Descent

Stochastic gradient descent (SGD) is almost ubiquitously used for training non-convex optimization tasks. Recently, a hypothesis proposed by Keskar et al. [2017] that large batch methods tend to converge to sharp minimizers has received increasing attention. We theoretically justify this hypothesis by providing new properties of SGD in both finite-time and asymptotic regimes. In particular, we give an explicit escaping time of SGD from a local minimum in the finite-time regime and prove that SGD tends to converge to flatter minima in the asymptotic regime (although may take exponential time to converge) regardless of the batch size. We also find that SGD with a larger ratio of learning rate to batch size tends to converge to a flat minimum faster, however, its generalization performance could be worse than the SGD with a smaller ratio of learning rate to batch size. We include numerical experiments to corroborate these theoretical findings.

stat.ML↗

Another Look at Statistical Calibration: A Non-Asymptotic Theory and Prediction-Oriented Optimality

We provide another look at the statistical calibration problem in computer models. This viewpoint is inspired by two overarching practical considerations of computer models: (i) many computer models are inadequate for perfectly modeling physical systems, even with the best-tuned calibration parameters; (ii) only a finite number of data points are available from the physical experiment associated with a computer model. Following this new line of thinking, we provide a non-asymptotic theory and derive a prediction-oriented calibration method. Our calibration method minimizes the predictive mean squared error for a finite sample size with statistical guarantees. We introduce an algorithm to perform the proposed calibration method and connect it to existing Bayesian calibration methods. Synthetic and real examples are provided to corroborate the derived theory and illustrate some advantages of the proposed calibration method.

stat.ME↗

Selection and Estimation Optimality in High Dimensions with the TWIN Penalty

We introduce a novel class of variable selection penalties called TWIN, which provides sensible data-adaptive penalization. Under a linear sparsity regime and random Gaussian designs we show that penalties in the TWIN class have a high probability of selecting the correct model and furthermore result in minimax optimal estimators. The general shape of penalty functions in the TWIN class is the key ingredient to its desirable properties and results in improved theoretical and empirical performance over existing penalties. In this work we introduce two examples of TWIN penalties that admit simple and efficient coordinate descent algorithms, making TWIN practical in large data settings. We demonstrate in challenging and realistic simulation settings with high correlations between active and inactive variables that TWIN has high power in variable selection while controlling the number of false discoveries, outperforming standard penalties.

stat.ME↗

Minimax Optimal Rates of Estimation in Functional ANOVA Models with Derivatives

We establish minimax optimal rates of convergence for nonparametric estimation in functional ANOVA models when data from first-order partial derivatives are available. Our results reveal that partial derivatives can improve convergence rates for function estimation with deterministic or random designs. In particular, for full $d$-interaction models, the optimal rates with first-order partial derivatives on $p$ covariates are identical to those for $(d-p)$-interaction models without partial derivatives. For additive models, the rates by using all first-order partial derivatives are root-$n$ to achieve the "parametric rate". We also investigate the minimax optimal rates for first-order partial derivative estimations when derivative data are available. Those rates coincide with the optimal rate for estimating the first-order derivative of a univariate function.

math.ST↗