Search

An important problem in time-series analysis is modeling systems with time-varying dynamics. Probabilistic models with joint continuous and discrete latent states offer interpretable, efficient, and experimentally useful descriptions of such data. Commonly used models include autoregressive hidden Markov models (ARHMMs) and switching linear dynamical systems (SLDSs), each with its own advantages and disadvantages. ARHMMs permit exact inference and easy parameter estimation, but are parameter intensive when modeling long dependencies, and hence are prone to overfitting. In contrast, SLDSs can capture long-range dependencies in a parameter efficient way through Markovian latent dynamics, but present an intractable likelihood and a challenging parameter estimation task. In this paper, we propose switching autoregressive low-rank tensor (SALT) models, which retain the advantages of both approaches while ameliorating the weaknesses. SALT parameterizes the tensor of an ARHMM with a low-rank factorization to control the number of parameters and allow longer range dependencies without overfitting. We prove theoretical and discuss practical connections between SALT, linear dynamical systems, and SLDSs. We empirically demonstrate quantitative advantages of SALT models on a range of simulated and real prediction tasks, including behavioral and neural datasets. Furthermore, the learned low-rank tensor provides novel insights into temporal dependencies within each discrete state.

Building a shopping product collection has been primarily a human job. With the manual efforts of craftsmanship, experts collect related but diverse products with common shopping intent that are effective when displayed together, e.g., backpacks, laptop bags, and messenger bags for freshman bag gifts. Automatically constructing a collection requires an ML system to learn a complex relationship between the customer's intent and the product's attributes. However, there have been challenging points, such as 1) long and complicated intent sentences, 2) rich and diverse product attributes, and 3) a huge semantic gap between them, making the problem difficult. In this paper, we use a pretrained language model (PLM) that leverages textual attributes of web-scale products to make intent-based product collections. Specifically, we train a BERT with triplet loss by setting an intent sentence to an anchor and corresponding products to positive examples. Also, we improve the performance of the model by search-based negative sampling and category-wise positive pair augmentation. Our model significantly outperforms the search-based baseline model for intent-based product matching in offline evaluations. Furthermore, online experimental results on our e-commerce platform show that the PLM-based method can construct collections of products with increased CTR, CVR, and order-diversity compared to expert-crafted collections., Comment: Accepted to IEEE International Workshop on Data Mining for Service (DMS2021)

How can we effectively regularize BERT? Although BERT proves its effectiveness in various downstream natural language processing tasks, it often overfits when there are only a small number of training instances. A promising direction to regularize BERT is based on pruning its attention heads based on a proxy score for head importance. However, heuristic-based methods are usually suboptimal since they predetermine the order by which attention heads are pruned. In order to overcome such a limitation, we propose AUBER, an effective regularization method that leverages reinforcement learning to automatically prune attention heads from BERT. Instead of depending on heuristics or rule-based policies, AUBER learns a pruning policy that determines which attention heads should or should not be pruned for regularization. Experimental results show that AUBER outperforms existing pruning methods by achieving up to 10% better accuracy. In addition, our ablation study empirically demonstrates the effectiveness of our design choices for AUBER.

How can we efficiently compress Convolutional Neural Network (CNN) while retaining their accuracy on classification tasks? Depthwise Separable Convolution (DSConv), which replaces a standard convolution with a depthwise convolution and a pointwise convolution, has been used for building lightweight architectures. However, previous works based on depthwise separable convolution are limited when compressing a trained CNN model since 1) they are mostly heuristic approaches without a precise understanding of their relations to standard convolution, and 2) their accuracies do not match that of the standard convolution. In this paper, we propose FALCON, an accurate and lightweight method to compress CNN. FALCON uses GEP, our proposed mathematical formulation to approximate the standard convolution kernel, to interpret existing convolution methods based on depthwise separable convolution. By exploiting the knowledge of a trained standard model and carefully determining the order of depthwise separable convolution via GEP, FALCON achieves sufficient accuracy close to that of the trained standard model. Furthermore, this interpretation leads to developing a generalized version rank-k FALCON which performs k independent FALCON operations and sums up the result. Experiments show that FALCON 1) provides higher accuracy than existing methods based on depthwise separable convolution and tensor decomposition, and 2) reduces the number of parameters and FLOPs of standard convolution by up to a factor of 8 while ensuring similar accuracy. We also demonstrate that rank-k FALCON further improves the accuracy while sacrificing a bit of compression and computation reduction rates.