Guo Yu

Guo Yu, Wenlin Liu, Yulan Hu et al.

On-policy distillation (\textsc{OPD}) has recently become a prominent post-training recipe as it combines two desirable ingredients: on-policy student trajectories and dense teacher supervision, yet how this hybrid changes a model's parameters remains unclear. Across several language and vision-language model pairs and use cases, our analysis yields two main findings. On sparsity, \textsc{OPD}-style updates are small and coordinate-sparse. They are distributed across layers and are usually FFN-heavy. This sparse structure is operationally useful: training only the discovered subnetwork recovers nearly the same performance as full \textsc{OPD}. However, the sparsity-inducing SGD optimizer underperforms AdamW in our optimizer ablation, likely because dense teacher supervision preserves heterogeneous coordinate-wise gradient scales where AdamW's adaptive scaling remains useful. On geometry, the updates are numerically full-rank but spectrally concentrated; they lie mostly away from the principal singular subspaces of the source weights and fall disproportionately on coordinates where the source weights are close to zero. These findings suggest that dense teacher supervision does not turn \textsc{OPD} into ordinary dense parameter rewriting; instead, \textsc{OPD} retains important geometric signatures of on-policy post-training.

Mingfei Gao, Zizhao Zhang, Guo Yu et al.

Active learning (AL) combines data labeling and model training to minimize the labeling cost by prioritizing the selection of high value data that can best improve model performance. In pool-based active learning, accessible unlabeled data are not used for model training in most conventional methods. Here, we propose to unify unlabeled sample selection and model training towards minimizing labeling cost, and make two contributions towards that end. First, we exploit both labeled and unlabeled data using semi-supervised learning (SSL) to distill information from unlabeled data during the training stage. Second, we propose a consistency-based sample selection metric that is coherent with the training objective such that the selected samples are effective at improving model performance. We conduct extensive experiments on image classification tasks. The experimental results on CIFAR-10, CIFAR-100 and ImageNet demonstrate the superior performance of our proposed method with limited labeled data, compared to the existing methods and the alternative AL and SSL combinations. Additionally, we study an important yet under-explored problem -- "When can we start learning-based AL selection?". We propose a measure that is empirically correlated with the AL target loss and is potentially useful for determining the proper starting point of learning-based AL methods.

Guo Yu, Jacob Bien

The lasso has been studied extensively as a tool for estimating the coefficient vector in the high-dimensional linear model; however, considerably less is known about estimating the error variance in this context. In this paper, we propose the natural lasso estimator for the error variance, which maximizes a penalized likelihood objective. A key aspect of the natural lasso is that the likelihood is expressed in terms of the natural parameterization of the multiparameter exponential family of a Gaussian with unknown mean and variance. The result is a remarkably simple estimator of the error variance with provably good performance in terms of mean squared error. These theoretical results do not require placing any assumptions on the design matrix or the true regression coefficients. We also propose a companion estimator, called the organic lasso, which theoretically does not require tuning of the regularization parameter. Both estimators do well empirically compared to preexisting methods, especially in settings where successful recovery of the true support of the coefficient vector is hard. Finally, we show that existing methods can do well under fewer assumptions than previously known, thus providing a fuller story about the problem of estimating the error variance in high-dimensional linear models.

Guo Yu, Jacob Bien

In many applications, data come with a natural ordering. This ordering can often induce local dependence among nearby variables. However, in complex data, the width of this dependence may vary, making simple assumptions such as a constant neighborhood size unrealistic. We propose a framework for learning this local dependence based on estimating the inverse of the Cholesky factor of the covariance matrix. Penalized maximum likelihood estimation of this matrix yields a simple regression interpretation for local dependence in which variables are predicted by their neighbors. Our proposed method involves solving a convex, penalized Gaussian likelihood problem with a hierarchical group lasso penalty. The problem decomposes into independent subproblems which can be solved efficiently in parallel using first-order methods. Our method yields a sparse, symmetric, positive definite estimator of the precision matrix, encoding a Gaussian graphical model. We derive theoretical results not found in existing methods attaining this structure. In particular, our conditions for signed support recovery and estimation consistency rates in multiple norms are as mild as those in a regression problem. Empirical results show our method performing favorably compared to existing methods. We apply our method to genomic data to flexibly model linkage disequilibrium. Our method is also applied to improve the performance of discriminant analysis in sound recording classification.