Knowledge distillation refers to a technique of transferring the knowledge from a large learned model or an ensemble of learned models to a small model. This method relies on access to the original training set, which might not always be available. A possible solution is a data-free adversarial distillation framework, which deploys a generative network to transfer the teacher model's knowledge to the student model. However, the data generation efficiency is low in the data-free adversarial distillation. We add an activation regularizer and a virtual interpolation method to improve the data generation efficiency. The activation regularizer enables the students to match the teacher's predictions close to activation boundaries and decision boundaries. The virtual interpolation method can generate virtual samples and labels in-between decision boundaries. Our experiments show that our approach surpasses state-of-the-art data-free distillation methods. The student model can achieve 95.42% accuracy on CIFAR-10 and 77.05% accuracy on CIFAR-100 without any original training data. Our model's accuracy is 13.8% higher than the state-of-the-art data-free method on CIFAR-100.

Medical time-series datasets have unique characteristics that make prediction tasks challenging. Most notably, patient trajectories often contain longitudinal variations in their input-output relationships, generally referred to as temporal conditional shift. Designing sequence models capable of adapting to such time-varying distributions remains a prevailing problem. To address this we present Model-Attentive Ensemble learning for Sequence modeling (MAES). MAES is a mixture of time-series experts which leverages an attention-based gating mechanism to specialize the experts on different sequence dynamics and adaptively weight their predictions. We demonstrate that MAES significantly out-performs popular sequence models on datasets subject to temporal shift.

Active learning is typically used to label data, when the labeling process is expensive. Several active learning algorithms have been theoretically proved to perform better than their passive counterpart. However, these algorithms rely on some assumptions, which themselves contain some specific parameters. This paper adresses the problem of adaptive active learning in a nonparametric setting with minimal assumptions. We present a novel algorithm that is valid under more general assumptions than the previously known algorithms, and that can moreover adapt to the parameters used in these assumptions. This allows us to work with a larger class of distributions, thereby avoiding to exclude important densities like gaussians. Our algorithm achieves a minimax rate of convergence, and therefore performs almost as well as the best known non-adaptive algorithms.

In 5G Ultra-Dense Networks, a distributed wireless backhaul is an attractive solution for forwarding traffic to the core. The macro-cell coverage area is divided into many small cells. A few of these cells are designated as gateways and are linked to the core by high-capacity fiber optic links. Each small cell is associated with one gateway and all small cells forward their traffic to their respective gateway through multi-hop mesh networks. We investigate the gateway location problem and show that finding near-optimal gateway locations improves the backhaul network capacity. An exact p-median integer linear program is formulated for comparison with our novel K-GA heuristic that combines a Genetic Algorithm (GA) with K-means clustering to find near-optimal gateway locations. We compare the performance of KGA with six other approaches in terms of average number of hops and backhaul network capacity at different node densities through extensive Monte Carlo simulations. All approaches are tested in various user distribution scenarios, including uniform distribution, bivariate Gaussian distribution, and cluster distribution. In all cases K-GA provides near-optimal results, achieving average number of hops and backhaul network capacity within 2% of optimal while saving an average of 95% of the execution time.

We propose a new model that estimates uncertainty in a single forward pass and works on both classification and regression problems. Our approach combines a bi-Lipschitz feature extractor with an inducing point approximate Gaussian process, offering robust and principled uncertainty estimation. This can be seen as a refinement of Deep Kernel Learning (DKL), with our changes allowing DKL to match softmax neural networks accuracy. Our method overcomes the limitations of previous work addressing deterministic uncertainty quantification, such as the dependence of uncertainty on ad hoc hyper-parameters. Our method matches SotA accuracy, 96.2% on CIFAR-10, while maintaining the speed of softmax models, and provides uncertainty estimates that outperform previous single forward pass uncertainty models. Finally, we demonstrate our method on a recently introduced benchmark for uncertainty in regression: treatment deferral in causal models for personalized medicine.

The softmax policy gradient (PG) method, which performs gradient ascent under softmax policy parameterization, is arguably one of the de facto implementations of policy optimization in modern reinforcement learning. For $\gamma$-discounted infinite-horizon tabular Markov decision processes (MDPs), remarkable progress has recently been achieved towards establishing global convergence of softmax PG methods in finding a near-optimal policy. However, prior results fall short of delineating clear dependencies of convergence rates on salient parameters such as the cardinality of the state space $\mathcal{S}$ and the effective horizon $\frac{1}{1-\gamma}$, both of which could be excessively large. In this paper, we deliver a pessimistic message regarding the iteration complexity of softmax PG methods, despite assuming access to exact gradient computation. Specifically, we demonstrate that softmax PG methods can take exponential time -- in terms of $|\mathcal{S}|$ and $\frac{1}{1-\gamma}$ -- to converge, even in the presence of a benign policy initialization and an initial state distribution amenable to exploration. This is accomplished by characterizing the algorithmic dynamics over a carefully-constructed MDP containing only three actions. Our exponential lower bound hints at the necessity of carefully adjusting update rules or enforcing proper regularization in accelerating PG methods.

We propose a novel method to estimate the coefficients of linear regression when outputs and inputs are contaminated by malicious outliers. Our method consists of two-step: (i) Make appropriate weights $\left\{\hat{w}_i\right\}_{i=1}^n$ such that the weighted sample mean of regression covariates robustly estimates the population mean of the regression covariate, (ii) Process Huber regression using $\left\{\hat{w}_i\right\}_{i=1}^n$. When (a) the regression covariate is a sequence with i.i.d. random vectors drawn from sub-Gaussian distribution with unknown mean and known identity covariance and (b) the absolute moment of the random noise is finite, our method attains a faster convergence rate than Diakonikolas, Kong and Stewart (2019) and Cherapanamjeri et al. (2020). Furthermore, our result is minimax optimal up to constant factor. When (a) the regression covariate is a sequence with i.i.d. random vectors drawn from heavy tailed distribution with unknown mean and bounded kurtosis and (b) the absolute moment of the random noise is finite, our method attains a convergence rate, which is minimax optimal up to constant factor.

Tissue microarray (TMA) images have emerged as an important high-throughput tool for cancer study and the validation of biomarkers. Efforts have been dedicated to further improve the accuracy of TACOMA, a cutting-edge automatic scoring algorithm for TMA images. One major advance is due to deepTacoma, an algorithm that incorporates suitable deep representations of a group nature. Inspired by the recent advance in semi-supervised learning and deep learning, we propose mfTacoma to learn alternative deep representations in the context of TMA image scoring. In particular, mfTacoma learns the low-dimensional manifolds, a common latent structure in high dimensional data. Deep representation learning and manifold learning typically requires large data. By encoding deep representation of the manifolds as regularizing features, mfTacoma effectively leverages the manifold information that is potentially crude due to small data. Our experiments show that deep features by manifolds outperforms two alternatives -- deep features by linear manifolds with principal component analysis or by leveraging the group property.

Individual behavior and decisions are substantially influenced by their contexts, such as location, environment, and time. Changes along these dimensions can be readily observed in Multiplayer Online Battle Arena games (MOBA), where players face different in-game settings for each match and are subject to frequent game patches. Existing methods utilizing contextual information generalize the effect of a context over the entire population, but contextual information tailored to each individual can be more effective. To achieve this, we present the Neural Individualized Context-aware Embeddings (NICE) model for predicting user performance and game outcomes. Our proposed method identifies individual behavioral differences in different contexts by learning latent representations of users and contexts through non-negative tensor factorization. Using a dataset from the MOBA game League of Legends, we demonstrate that our model substantially improves the prediction of winning outcome, individual user performance, and user engagement.

Learning effective representations in image-based environments is crucial for sample efficient Reinforcement Learning (RL). Unfortunately, in RL, representation learning is confounded with the exploratory experience of the agent -- learning a useful representation requires diverse data, while effective exploration is only possible with coherent representations. Furthermore, we would like to learn representations that not only generalize across tasks but also accelerate downstream exploration for efficient task-specific training. To address these challenges we propose Proto-RL, a self-supervised framework that ties representation learning with exploration through prototypical representations. These prototypes simultaneously serve as a summarization of the exploratory experience of an agent as well as a basis for representing observations. We pre-train these task-agnostic representations and prototypes on environments without downstream task information. This enables state-of-the-art downstream policy learning on a set of difficult continuous control tasks.