Knowledge distillation (KD) is a successful approach for deep neural network acceleration, with which a compact network (student) is trained by mimicking the softmax output of a pre-trained high-capacity network (teacher). In tradition, KD usually relies on access to the training samples and the parameters of the white-box teacher to acquire the transferred knowledge. However, these prerequisites are not always realistic due to storage costs or privacy issues in real-world applications. Here we propose the concept of decision-based black-box (DB3) knowledge distillation, with which the student is trained by distilling the knowledge from a black-box teacher (parameters are not accessible) that only returns classes rather than softmax outputs. We start with the scenario when the training set is accessible. We represent a sample's robustness against other classes by computing its distances to the teacher's decision boundaries and use it to construct the soft label for each training sample. After that, the student can be trained via standard KD. We then extend this approach to a more challenging scenario in which even accessing the training data is not feasible. We propose to generate pseudo samples distinguished by the teacher's decision boundaries to the largest extent and construct soft labels for them, which are used as the transfer set. We evaluate our approaches on various benchmark networks and datasets and experiment results demonstrate their effectiveness. Codes are available at: https://github.com/zwang84/zsdb3kd.

We present SoftDICE, which achieves state-of-the-art performance for imitation learning. SoftDICE fixes several key problems in ValueDICE, an off-policy distribution matching approach for sample-efficient imitation learning. Specifically, the objective of ValueDICE contains logarithms and exponentials of expectations, for which the mini-batch gradient estimate is always biased. Second, ValueDICE regularizes the objective with replay buffer samples when expert demonstrations are limited in number, which however changes the original distribution matching problem. Third, the re-parametrization trick used to derive the off-policy objective relies on an implicit assumption that rarely holds in training. We leverage a novel formulation of distribution matching and consider an entropy-regularized off-policy objective, which yields a completely offline algorithm called SoftDICE. Our empirical results show that SoftDICE recovers the expert policy with only one demonstration trajectory and no further on-policy/off-policy samples. SoftDICE also stably outperforms ValueDICE and other baselines in terms of sample efficiency on Mujoco benchmark tasks.

Forecasting the particulate matter (PM) concentration in South Korea has become urgently necessary owing to its strong negative impact on human life. In most statistical or machine learning methods, independent and identically distributed data, for example, a Gaussian distribution, are assumed; however, time series such as air pollution and weather data do not meet this assumption. In this study, the maximum correntropy criterion for regression (MCCR) loss is used in an analysis of the statistical characteristics of air pollution and weather data. Rigorous seasonality adjustment of the air pollution and weather data was performed because of their complex seasonality patterns and the heavy-tailed distribution of data even after deseasonalization. The MCCR loss was applied to multiple models including conventional statistical models and state-of-the-art machine learning models. The results show that the MCCR loss is more appropriate than the conventional mean squared error loss for forecasting extreme values.

Gradient Boosting Machines (GBM) are hugely popular for solving tabular data problems. However, practitioners are not only interested in point predictions, but also in probabilistic predictions in order to quantify the uncertainty of the predictions. Creating such probabilistic predictions is difficult with existing GBM-based solutions: they either require training multiple models or they become too computationally expensive to be useful for large-scale settings. We propose Probabilistic Gradient Boosting Machines (PGBM), a method to create probabilistic predictions with a single ensemble of decision trees in a computationally efficient manner. PGBM approximates the leaf weights in a decision tree as a random variable, and approximates the mean and variance of each sample in a dataset via stochastic tree ensemble update equations. These learned moments allow us to subsequently sample from a specified distribution after training. We empirically demonstrate the advantages of PGBM compared to existing state-of-the-art methods: (i) PGBM enables probabilistic estimates without compromising on point performance in a single model, (ii) PGBM learns probabilistic estimates via a single model only (and without requiring multi-parameter boosting), and thereby offers a speedup of up to several orders of magnitude over existing state-of-the-art methods on large datasets, and (iii) PGBM achieves accurate probabilistic estimates in tasks with complex differentiable loss functions, such as hierarchical time series problems, where we observed up to 10% improvement in point forecasting performance and up to 300% improvement in probabilistic forecasting performance.

Humor is an important social phenomenon, serving complex social and psychological functions. However, despite being studied for millennia humor is computationally not well understood, often considered an AI-complete problem. In this work, we introduce a novel setting in humor mining: automatically detecting funny and unusual scientific papers. We are inspired by the Ig Nobel prize, a satirical prize awarded annually to celebrate funny scientific achievements (example past winner: "Are cows more likely to lie down the longer they stand?"). This challenging task has unique characteristics that make it particularly suitable for automatic learning. We construct a dataset containing thousands of funny papers and use it to learn classifiers, combining findings from psychology and linguistics with recent advances in NLP. We use our models to identify potentially funny papers in a large dataset of over 630,000 articles. The results demonstrate the potential of our methods, and more broadly the utility of integrating state-of-the-art NLP methods with insights from more traditional disciplines.

In this paper, we propose Domain Agnostic Meta Score-based Learning (DAMSL), a novel, versatile and highly effective solution that delivers significant out-performance over state-of-the-art methods for cross-domain few-shot learning. We identify key problems in previous meta-learning methods over-fitting to the source domain, and previous transfer-learning methods under-utilizing the structure of the support set. The core idea behind our method is that instead of directly using the scores from a fine-tuned feature encoder, we use these scores to create input coordinates for a domain agnostic metric space. A graph neural network is applied to learn an embedding and relation function over these coordinates to process all information contained in the score distribution of the support set. We test our model on both established CD-FSL benchmarks and new domains and show that our method overcomes the limitations of previous meta-learning and transfer-learning methods to deliver substantial improvements in accuracy across both smaller and larger domain shifts.

The missing data problem has been broadly studied in the last few decades and has various applications in different areas such as statistics or bioinformatics. Even though many methods have been developed to tackle this challenge, most of those are imputation techniques that require multiple iterations through the data before yielding convergence. In addition, such approaches may introduce extra biases and noises to the estimated parameters. In this work, we propose novel algorithms to find the maximum likelihood estimates (MLEs) for a one-class/multiple-class randomly missing data set under some mild assumptions. As the computation is direct without any imputation, our algorithms do not require multiple iterations through the data, thus promising to be less time-consuming than other methods while maintaining superior estimation performance. We validate these claims by empirical results on various data sets of different sizes and release all codes in a GitHub repository to contribute to the research community related to this problem.

In forming learning objectives, one oftentimes needs to aggregate a set of individual values to a single output. Such cases occur in the aggregate loss, which combines individual losses of a learning model over each training sample, and in the individual loss for multi-label learning, which combines prediction scores over all class labels. In this work, we introduce the sum of ranked range (SoRR) as a general approach to form learning objectives. A ranked range is a consecutive sequence of sorted values of a set of real numbers. The minimization of SoRR is solved with the difference of convex algorithm (DCA). We explore two applications in machine learning of the minimization of the SoRR framework, namely the AoRR aggregate loss for binary/multi-class classification at the sample level and the TKML individual loss for multi-label/multi-class classification at the label level. A combination loss of AoRR and TKML is proposed as a new learning objective for improving the robustness of multi-label learning in the face of outliers in sample and labels alike. Our empirical results highlight the effectiveness of the proposed optimization frameworks and demonstrate the applicability of proposed losses using synthetic and real data sets.

The shortcomings of maximum likelihood estimation in the context of model-based reinforcement learning have been highlighted by an increasing number of papers. When the model class is misspecified or has a limited representational capacity, model parameters with high likelihood might not necessarily result in high performance of the agent on a downstream control task. To alleviate this problem, we propose an end-to-end approach for model learning which directly optimizes the expected returns using implicit differentiation. We treat a value function that satisfies the Bellman optimality operator induced by the model as an implicit function of model parameters and show how to differentiate the function. We provide theoretical and empirical evidence highlighting the benefits of our approach in the model misspecification regime compared to likelihood-based methods.

Modern machine learning models often employ a huge number of parameters and are typically optimized to have zero training loss; yet surprisingly, they possess near-optimal prediction performance, contradicting classical learning theory. We examine how these benign overfitting phenomena occur in a two-layer neural network setting where sample covariates are corrupted with noise. We address the high dimensional regime, where the data dimension $d$ grows with the number $n$ of data points. Our analysis combines an upper bound on the bias with matching upper and lower bounds on the variance of the interpolator (an estimator that interpolates the data). These results indicate that the excess learning risk of the interpolator decays under mild conditions. We further show that it is possible for the two-layer ReLU network interpolator to achieve a near minimax-optimal learning rate, which to our knowledge is the first generalization result for such networks. Finally, our theory predicts that the excess learning risk starts to increase once the number of parameters $s$ grows beyond $O(n^2)$, matching recent empirical findings.