A SIZABLE amount of real estate in the media is dedicated to artificial intelligence (AI) these days - and especially so in Singapore. From news of the government's commitment to AI research, policies delineating its ethical use, to announcements of the next big global firm setting up AI facilities in the Red Dot, the ubiquity of the technology foretells the central role that it is set to play in every aspect of society. At the enterprise level, adoption within day-to-day operations has been steadily ramping up - especially in sectors such as supply chain, where AI-led tech such as Internet of Things (IOT) analytics have shown a clear business case in its ability to substantially boost efficiency. Meanwhile, consumers are increasingly being exposed to AI as well - think, for example, of Google Pixel's AI-powered night sight, which enables photography in the dark. That said, and while it is already an aphorism that businesses need to adopt AI to maintain a competitive edge, it just as important that businesses refrain from regarding AI as one big, shiny, silver bullet.

Despite its huge empirical success, deep learning still preserves many features of alchemy [Rahimi, 2017]: progress in this field is obtained mainly by trial and error, and our intuition about how do neural networks actually work often misleads us. Alchemy, in order to become usual chemistry, needs a theoretical ground. For now, a solid theoretical ground for deep learning is lacking, however, fortunately, many pieces of theory appeared from different directions during several past years. The purpose of this essay is not to provide a comprehensive review, but to draw connections between some works on this topic. The list of works mentioned here is by no means representative, or, all the more so, complete. Since the theory of deep learning is lacking, some features of neural networks learning seem "mysterious". We emphasize two mysteries of deep learning: 1. Generalization mystery. It is very common for contemporary neural networks to have many more parameters than the number of training examples at hand.

The underlying objective of food authentication studies is to determine whether unknown food samples have been correctly labelled. In this paper we study three near infrared (NIR) spectroscopic datasets from food samples of different types: meat samples (labelled by species), olive oil samples (labelled by their geographic origin) and honey samples (labelled as pure or adulterated by different adulterants). We apply and compare a large number of classification, dimension reduction and variable selection approaches to these datasets. NIR data pose specific challenges to classification and variable selection: the datasets are high - dimensional where the number of cases ($n$) $<<$ number of features ($p$) and the recorded features are highly serially correlated. In this paper we carry out comparative analysis of different approaches and find that partial least squares, a classic tool employed for these types of data, outperforms all the other approaches considered.

Aggregating responses from crowd workers is a fundamental task in the process of crowdsourcing. In cases where a few experts are overwhelmed by a large number of non-experts, most answer aggregation algorithms such as the majority voting fail to identify the correct answers. Therefore, it is crucial to extract reliable experts from the crowd workers. In this study, we introduce the notion of "expert core", which is a set of workers that is very unlikely to contain a non-expert. We design a graph-mining-based efficient algorithm that exactly computes the expert core. To answer the aggregation task, we propose two types of algorithms. The first one incorporates the expert core into existing answer aggregation algorithms such as the majority voting, whereas the second one utilizes information provided by the expert core extraction algorithm pertaining to the reliability of workers. We then give a theoretical justification for the first type of algorithm. Computational experiments using synthetic and real-world datasets demonstrate that our proposed answer aggregation algorithms outperform state-of-the-art algorithms.

Much research work in computer vision is being spent on optimizing existing network architectures to obtain a few more percentage points on benchmarks. Recent AutoML approaches promise to relieve us from this effort. However, they are mainly designed for comparatively small-scale classification tasks. In this work, we show how to use and extend existing AutoML techniques to efficiently optimize large-scale U-Net-like encoder-decoder architectures. In particular, we leverage gradient-based neural architecture search and Bayesian optimization for hyperparameter search. The resulting optimization does not require a large company-scale compute cluster. We show results on disparity estimation that clearly outperform the manually optimized baseline and reach state-of-the-art performance.

We describe a new approach for mitigating risk in the Reinforcement Learning paradigm. Instead of reasoning about expected utility, we use second-order stochastic dominance (SSD) to directly compare the inherent risk of random returns induced by different actions. We frame the RL optimization within the space of probability measures to accommodate the SSD relation, treating Bellman's equation as a potential energy functional. This brings us to Wasserstein gradient flows, for which the optimality and convergence are well understood. We propose a discrete-measure approximation algorithm called the Dominant Particle Agent (DPA), and we demonstrate how safety and performance are better balanced with DPA than with existing baselines.

Sequence classification is an important data mining task in many real world applications. Over the past few decades, many sequence classification methods have been proposed from different aspects. In particular, the pattern-based method is one of the most important and widely studied sequence classification methods in the literature. In this paper, we present a reference-based sequence classification framework, which can unify existing pattern-based sequence classification methods under the same umbrella. More importantly, this framework can be used as a general platform for developing new sequence classification algorithms. By utilizing this framework as a tool, we propose new sequence classification algorithms that are quite different from existing solutions. Experimental results show that new methods developed under the proposed framework are capable of achieving comparable classification accuracy to those state-of-the-art sequence classification algorithms.

Shaping in humans and animals has been shown to be a powerful tool for learning complex tasks as compared to learning in a randomized fashion. This makes the problem less complex and enables one to solve the easier sub task at hand first. Generating a curriculum for such guided learning involves subjecting the agent to easier goals first, and then gradually increasing their difficulty. This paper takes a similar direction and proposes a dual curriculum scheme for solving robotic manipulation tasks with sparse rewards, called MaMiC. It includes a macro curriculum scheme which divides the task into multiple sub-tasks followed by a micro curriculum scheme which enables the agent to learn between such discovered sub-tasks. We show how combining macro and micro curriculum strategies help in overcoming major exploratory constraints considered in robot manipulation tasks without having to engineer any complex rewards. We also illustrate the meaning of the individual curricula and how they can be used independently based on the task. The performance of such a dual curriculum scheme is analyzed on the Fetch environments.

Due to the ease of modern data collection, applied statisticians often have access to a large set of covariates that they wish to relate to some observed outcome. Generalized linear models (GLMs) offer a particularly interpretable framework for such an analysis. In these high-dimensional problems, the number of covariates is often large relative to the number of observations, so we face non-trivial inferential uncertainty; a Bayesian approach allows coherent quantification of this uncertainty. Unfortunately, existing methods for Bayesian inference in GLMs require running times roughly cubic in parameter dimension, and so are limited to settings with at most tens of thousand parameters. We propose to reduce time and memory costs with a low-rank approximation of the data in an approach we call LR-GLM. When used with the Laplace approximation or Markov chain Monte Carlo, LR-GLM provides a full Bayesian posterior approximation and admits running times reduced by a full factor of the parameter dimension. We rigorously establish the quality of our approximation and show how the choice of rank allows a tunable computational-statistical trade-off. Experiments support our theory and demonstrate the efficacy of LR-GLM on real large-scale datasets.

Record companies invest billions of dollars in new talent around the globe each year. Gaining insight into what actually makes a hit song would provide tremendous benefits for the music industry. In this research we tackle this question by focussing on the dance hit song classification problem. A database of dance hit songs from 1985 until 2013 is built, including basic musical features, as well as more advanced features that capture a temporal aspect. A number of different classifiers are used to build and test dance hit prediction models. The resulting best model has a good performance when predicting whether a song is a "top 10" dance hit versus a lower listed position.