Learning with Scalability and Compactness

Artificial Intelligence has been thriving for decades since its birth. Traditional AI features heuristic search and planning, providing good strategy for tasks that are inherently search-based problems, such as games and GPS searching. In the meantime, machine learning, arguably the hottest subfield of AI, embraces data-driven methodology with great success in a wide range of applications such as computer vision and speech recognition. As a new trend, the applications of both learning and search have shifted toward mobile and embedded devices which entails not only scalability but also compactness of the models. Under this general paradigm, we propose a series of work to address the issues of scalability and compactness within machine learning and its applications on heuristic search. We first focus on the scalability issue of memory-based heuristic search which is recently ameliorated by Maximum Variance Unfolding (MVU), a manifold learning algorithm capable of learning state embeddings as effective heuristics to speed up $A^*$ search. Though achieving unprecedented online search performance with constraints on memory footprint, MVU is notoriously slow on offline training. To address this problem, we introduce Maximum Variance Correction (MVC), which finds large-scale feasible solutions to MVU by post-processing embeddings from any manifold learning algorithm. It increases the scale of MVU embeddings by several orders of magnitude and is naturally parallel. We further propose Goal-oriented Euclidean Heuristic (GOEH), a variant to MVU embeddings, which preferably optimizes the heuristics associated with goals in the embedding while maintaining their admissibility. We demonstrate unmatched reductions in search time across several non-trivial $A^*$ benchmark search problems. Through these work, we bridge the gap between the manifold learning literature and heuristic search which have been regarded as fundamentally different, leading to cross-fertilization for both fields. Deep learning has made a big splash in the machine learning community with its superior accuracy performance. However, it comes at a price of huge model size that might involves billions of parameters, which poses great challenges for its use on mobile and embedded devices. To achieve the compactness, we propose HashedNets, a general approach to compressing neural network models leveraging feature hashing. At its core, HashedNets randomly group parameters using a low-cost hash function, and share parameter value within the group. According to our empirical results, a neural network could be 32x smaller with little drop in accuracy performance. We further introduce Frequency-Sensitive Hashed Nets (FreshNets) to extend this hashing technique to convolutional neural network by compressing parameters in the frequency domain. Compared with many AI applications, neural networks seem not graining as much popularity as it should be in traditional data mining tasks. For these tasks, categorical features need to be first converted to numerical representation in advance in order for neural networks to process them. We show that a na\ {i}ve use of the classic one-hot encoding may result in gigantic weight matrices and therefore lead to prohibitively expensive memory cost in neural networks. Inspired by word embedding, we advocate a compellingly simple, yet effective neural network architecture with category embedding. It is capable of directly handling both numerical and categorical features as well as providing visual insights on feature similarities. At the end, we conduct comprehensive empirical evaluation which showcases the efficacy and practicality of our approach, and provides surprisingly good visualization and clustering for categorical features