Artificial intelligence (AI) continued to be a major driver of digital transformation in 2017, with the rapidly advancing technology affecting business strategy and operations, customer interactions and the workforce itself. While these are all general and broad impacts of AI, they will continue to be important for businesses trying to keep up with rapid technological advancements in 2018. Embedded deep learning will become the focus of software product teams in the coming year, as buyers will begin to inquire about the machine learning as a service capabilities of the tools they are purchasing. Many vendors are already including machine learning in products to enhance and automate certain functionalities, and build marketing campaigns around those AI enhancements. As embedded AI becomes more standard in solutions, there will be less of an emphasis around the glitz and glamour of machine learning and more of a focus on how the embedded AI is contributing to a business' overall digital transformation.

We propose approaches based on deep learning to localize objects in images when only a small training dataset is available and the images have low quality. That applies to many problems in medical image processing, and in particular to the analysis of fluoroscopic (low-dose) X-ray images, where the images have low contrast. We solve the problem by incorporating high-level information about the objects, which could be a simple geometrical model, like a circular outline, or a more complex statistical model. A simple geometrical representation can sufficiently describe some objects and only requires minimal labeling. Statistical shape models can be used to represent more complex objects. We propose computationally efficient two-stage approaches, which we call deep morphing, for both representations by fitting the representation to the output of a deep segmentation network.

Policy optimization is a core component of reinforcement learning (RL), and most existing RL methods directly optimize parameters of a policy based on maximizing the expected total reward, or its surrogate. Though often achieving encouraging empirical success, its underlying mathematical principle on {\em policy-distribution} optimization is unclear. We place policy optimization into the space of probability measures, and interpret it as Wasserstein gradient flows. On the probability-measure space, under specified circumstances, policy optimization becomes a convex problem in terms of distribution optimization. To make optimization feasible, we develop efficient algorithms by numerically solving the corresponding discrete gradient flows. Our technique is applicable to several RL settings, and is related to many state-of-the-art policy-optimization algorithms. Empirical results verify the effectiveness of our framework, often obtaining better performance compared to related algorithms.

In this paper we investigate the supervised backpropagation training of multilayer neural networks from a dynamical systems point of view. We discuss some links with the qualitative theory of differential equations and introduce the overfly algorithm to tackle the local minima problem. Our approach is based on the existence of first integrals of the generalised gradient system with build-in dissipation.

Transportation modes prediction is a fundamental task for decision making in smart cities and traffic management systems. Traffic policies designed based on trajectory mining can save money and time for authorities and the public. It may reduce the fuel consumption and commute time and moreover, may provide more pleasant moments for residents and tourists. Since the number of features that may be used to predict a user transportation mode can be substantial, finding a subset of features that maximizes a performance measure is worth investigating. In this work, we explore wrapper and information retrieval methods to find the best subset of trajectory features. After finding the best classifier and the best feature subset, our results were compared with two related papers that applied deep learning methods and the results showed that our framework achieved better performance. Furthermore, two types of cross-validation approaches were investigated, and the performance results show that the random cross-validation method provides optimistic results.

LSTMs and other RNN variants have shown strong performance on character-level language modeling. These models are typically trained using truncated backpropagation through time, and it is common to assume that their success stems from their ability to remember long-term contexts. In this paper, we show that a deep (64-layer) transformer model with fixed context outperforms RNN variants by a large margin, achieving state of the art on two popular benchmarks- 1.13 bits per character on text8 and 1.06 on enwik8. To get good results at this depth, we show that it is important to add auxiliary losses, both at intermediate network layers and intermediate sequence positions.

Because of their effectiveness in broad practical applications, LSTM networks have received a wealth of coverage in scientific journals, technical blogs, and implementation guides. However, in most articles, the inference formulas for the LSTM network and its parent, RNN, are stated axiomatically, while the training formulas are omitted altogether. In addition, the technique of "unrolling" an RNN is routinely presented without justification throughout the literature. The goal of this paper is to explain the essential RNN and LSTM fundamentals in a single document. Drawing from concepts in signal processing, we formally derive the canonical RNN formulation from differential equations. We then propose and prove a precise statement, which yields the RNN unrolling technique. We also review the difficulties with training the standard RNN and address them by transforming the RNN into the "Vanilla LSTM" network through a series of logical arguments. We provide all equations pertaining to the LSTM system together with detailed descriptions of its constituent entities. Albeit unconventional, our choice of notation and the method for presenting the LSTM system emphasizes ease of understanding. As part of the analysis, we identify new opportunities to enrich the LSTM system and incorporate these extensions into the Vanilla LSTM network, producing the most general LSTM variant to date. The target reader has already been exposed to RNNs and LSTM networks through numerous available resources and is open to an alternative pedagogical approach. A Machine Learning practitioner seeking guidance for implementing our new augmented LSTM model in software for experimentation and research will find the insights and derivations in this tutorial valuable as well.

Working memory is an essential component of reasoning -- the capacity to answer a new question by manipulating acquired knowledge. Current memory-augmented neural networks offer a differentiable method to realize limited reasoning with support of a working memory module. Memory modules are often implemented as a set of memory slots without explicit relational exchange of content. This does not naturally match multi-relational domains in which data is structured. We design a new model dubbed Relational Dynamic Memory Network (RDMN) to fill this gap. The memory can have a single or multiple components, each of which realizes a multi-relational graph of memory slots. The memory is dynamically updated in the reasoning process controlled by the central controller. We evaluate the capability of RDMN on several important application domains: software vulnerability, molecular bioactivity and chemical reaction. Results demonstrate the efficacy of the proposed model.

Vision is a common source of inspiration for poetry. The objects and the sentimental imprints that one perceives from an image may lead to various feelings depending on the reader. In this paper, we present a system of poetry generation from images to mimic the process. Given an image, we first extract a few keywords representing objects and sentiments perceived from the image. These keywords are then expanded to related ones based on their associations in human written poems. Finally, verses are generated gradually from the keywords using recurrent neural networks trained on existing poems. Our approach is evaluated by human assessors and compared to other generation baselines. The results show that our method can generate poems that are more artistic than the baseline methods. This is one of the few attempts to generate poetry from images. By deploying our proposed approach, XiaoIce has already generated more than 12 million poems for users since its release in July 2017. A book of its poems has been published by Cheers Publishing, which claimed that the book is the first-ever poetry collection written by an AI in human history.

Ticket assignment/dispatch is a crucial part of service delivery business with lot of scope for automation and optimization. In this paper, we present an end-to-end automated helpdesk email ticket assignment system, which is also offered as a service. The objective of the system is to determine the nature of the problem mentioned in an incoming email ticket and then automatically dispatch it to an appropriate resolver group (or team) for resolution. The proposed system uses an ensemble classifier augmented with a configurable rule engine. While design of classifier that is accurate is one of the main challenges, we also need to address the need of designing a system that is robust and adaptive to changing business needs. We discuss some of the main design challenges associated with email ticket assignment automation and how we solve them. The design decisions for our system are driven by high accuracy, coverage, business continuity, scalability and optimal usage of computational resources. Our system has been deployed in production of three major service providers and currently assigning over 40,000 emails per month, on an average, with an accuracy close to 90% and covering at least 90% of email tickets.