The goal of query performance prediction (QPP) is to automatically estimate the effectiveness of a search result for any given query, without relevance judgements. Post-retrieval features have been shown to be more effective for this task while being more expensive to compute than pre-retrieval features. Combining multiple post-retrieval features is even more effective, but state-of-the-art QPP methods are impossible to interpret because of the black-box nature of the employed machine learning models. However, interpretation is useful for understanding the predictive model and providing more answers about its behavior. Moreover, combining many post-retrieval features is not applicable to real-world cases, since the query running time is of utter importance. In this paper, we investigate a new framework for feature selection in which the trained model explains well the prediction. We introduce a step-wise (forward and backward) model selection approach where different subsets of query features are used to fit different models from which the system selects the best one. We evaluate our approach on four TREC collections using standard QPP features. We also develop two QPP features to address the issue of query-drift in the query feedback setting. We found that: (1) our model based on a limited number of selected features is as good as more complex models for QPP and better than non-selective models; (2) our model is more efficient than complex models during inference time since it requires fewer features; (3) the predictive model is readable and understandable; and (4) one of our new QPP features is consistently selected across different collections, proving its usefulness.

Methods based on vector embeddings of knowledge graphs have been actively pursued as a promising approach to knowledge graph completion.However, embedding models generate storage-inefficient representations, particularly when the number of entities and relations, and the dimensionality of the real-valued embedding vectors are large. We present a binarized CANDECOMP/PARAFAC(CP) decomposition algorithm, which we refer to as B-CP, where real-valued parameters are replaced by binary values to reduce model size. Moreover, we show that a fast score computation technique can be developed with bitwise operations. We prove that B-CP is fully expressive by deriving a bound on the size of its embeddings. Experimental results on several benchmark datasets demonstrate that the proposed method successfully reduces model size by more than an order of magnitude while maintaining task performance at the same level as the real-valued CP model.

Prior work in visual dialog has focused on training deep neural models on the VisDial dataset in isolation, which has led to great progress, but is limiting and wasteful. In this work, following recent trends in representation learning for language, we introduce an approach to leverage pretraining on related large-scale vision-language datasets before transferring to visual dialog. Specifically, we adapt the recently proposed ViLBERT (Lu et al., 2019) model for multi-turn visually-grounded conversation sequences. Our model is pretrained on the Conceptual Captions and Visual Question Answering datasets, and finetuned on VisDial with a VisDial-specific input representation and the masked language modeling and next sentence prediction objectives (as in BERT). Our best single model achieves state-of-the-art on Visual Dialog, outperforming prior published work (including model ensembles) by more than 1% absolute on NDCG and MRR. Next, we carefully analyse our model and find that additional finetuning using 'dense' annotations i.e. relevance scores for all 100 answer options corresponding to each question on a subset of the training set, leads to even higher NDCG -- more than 10% over our base model -- but hurts MRR -- more than 17% below our base model! This highlights a stark trade-off between the two primary metrics for this task -- NDCG and MRR. We find that this is because dense annotations in the dataset do not correlate well with the original ground-truth answers to questions, often rewarding the model for generic responses (e.g. "can't tell").

We lower bound the complexity of finding $\epsilon$-stationary points (with gradient norm at most $\epsilon$) using stochastic first-order methods. In a well-studied model where algorithms access smooth, potentially non-convex functions through queries to an unbiased stochastic gradient oracle with bounded variance, we prove that (in the worst case) any algorithm requires at least $\epsilon^{-4}$ queries to find an $\epsilon$ stationary point. The lower bound is tight, and establishes that stochastic gradient descent is minimax optimal in this model. In a more restrictive model where the noisy gradient estimates satisfy a mean-squared smoothness property, we prove a lower bound of $\epsilon^{-3}$ queries, establishing the optimality of recently proposed variance reduction techniques.

Item response theory (IRT) is a non-linear generative probabilistic paradigm for using exams to identify, quantify, and compare latent traits of individuals, relative to their peers, within a population of interest. In pre-existing multidimensional IRT methods, one requires a factorization of the test items. For this task, linear exploratory factor analysis is used, making IRT a posthoc model. We propose skipping the initial factor analysis by using a sparsity-promoting horseshoe prior to perform factorization directly within the IRT model so that all training occurs in a single self-consistent step. Being a hierarchical Bayesian model, we adapt the WAIC to the problem of dimensionality selection. IRT models are analogous to probabilistic autoencoders. By binding the generative IRT model to a Bayesian neural network (forming a probabilistic autoencoder), one obtains a scoring algorithm consistent with the interpretable Bayesian model. In some IRT applications the black-box nature of a neural network scoring machine is desirable. In this manuscript, we demonstrate within-IRT factorization and comment on scoring approaches.

Natural gradient has been recently introduced to the field of boosting to enable the generic probabilistic predication capability. Natural gradient boosting shows promising performance improvements on small datasets due to better training dynamics, but it suffers from slow training speed overhead especially for large datasets. W e present a replication study of NGBoost ( Duan et al., 2019) training that carefully examines the impacts of key hyper-parameters under the circumstance of best-first decision tree learning. W e find that with the regularization of leaf number clipping, the performance of NGBoost can be largely improved via a better choice of hyperparameters. Experiments show that our approach significantly beats the state-of-the-art performance on various kinds of datasets from the UCI Machine Learning Repository while still has up to 4.85x speed up compared with the original approach of NGBoost.

We show that a variety of modern deep learning tasks exhibit a "double-descent" phenomenon where, as we increase model size, performance first gets worse and then gets better. Moreover, we show that double descent occurs not just as a function of model size, but also as a function of the number of training epochs. We unify the above phenomena by defining a new complexity measure we call the effective model complexity and conjecture a generalized double descent with respect to this measure. Furthermore, our notion of model complexity allows us to identify certain regimes where increasing (even quadrupling) the number of train samples actually hurts test performance. Right: Test error, shown for varying train epochs. All models trained using Adam for 4K epochs. The bias-variance tradeoff is a fundamental concept in classical statistical learning theory (e.g., Hastie et al. (2005)). The idea is that models of higher complexity have lower bias but higher variance. According to this theory, once model complexity passes a certain threshold, models "overfit" with the variance term dominating the test error, and hence from this point onward, increasing model complexity will only decrease performance (i.e., increase test error). Hence conventional wisdom in classical statistics is that, once we pass a certain threshold, "larger models are worse. Such networks have millions of parameters, more than enough to fit even random labels (Zhang et al. (2016)), and yet they perform much better on many tasks than smaller models. Indeed, conventional wisdom among practitioners is that "larger models are better' ' (Krizhevsky et al. (2012), Huang et al. (2018), Szegedy et al.

We place an Indian Buffet Process (IBP) prior over the neural structure of a Bayesian Neural Network (BNN), thus allowing the complexity of the BNN to increase and decrease automatically. We apply this methodology to the problem of resource allocation in continual learning, where new tasks occur and the network requires extra resources. Our BNN exploits online variational inference with relaxations to the Bernoulli and Beta distributions (which constitute the IBP prior), so allowing the use of the reparameterisation trick to learn variational posteriors via gradient-based methods. As we automatically learn the number of weights in the BNN, overfitting and underfitting problems are largely overcome. We show empirically that the method offers competitive results compared to Variational Continual Learning (VCL) in some settings.

Although convolutional neural networks (CNNs) are inspired by the mechanisms behind human visual systems, they diverge on many measures such as ambiguity or hardness. In this paper, we make a surprising discovery: there exists a (nearly) universal score function for CNNs whose correlation is statistically significant than the widely used model confidence with human visual hardness. We term this function as angular visual hardness (AVH) which is given by the normalized angular distance between a feature embedding and the classifier weights of the corresponding target category in a CNN. We conduct an in-depth scientific study. We observe that CNN models with the highest accuracy also have the best AVH scores. This agrees with an earlier finding that state-of-art models tend to improve on the classification of harder training examples. We find that AVH displays interesting dynamics during training: it quickly reaches a plateau even though the training loss keeps improving. This suggests the need for designing better loss functions that can target harder examples more effectively. Finally, we empirically show significant improvement in performance by using AVH as a measure of hardness in self-training methods for domain adaptation.

The ability to analyze and forecast stratospheric weather conditions is fundamental to addressing climate change. However, our capacity to collect data in the stratosphere is limited by sparsely deployed weather balloons. We propose a framework to collect stratospheric data by releasing a contrail of tiny sensor devices as a weather balloon ascends. The key machine learning challenges are determining when and how to deploy a finite collection of sensors to produce a useful data set. We decide when to release sensors by modeling the deviation of a forecast from actual stratospheric conditions as a Gaussian process. We then implement a novel hardware system that is capable of optimally releasing sensors from a rising weather balloon. We show that this data engineering framework is effective through real weather balloon flights, as well as simulations.