In this talk we'll describe the problem of detecting duplicate items in the media catalog and how it powers many features on the Amazon website. I'll give a brief introduction to the problem, the data and how we use machine learning to solve it. Because the data is complex and unstructured our ability to solve the problem is constrained by representation more than model capacity. I'll then present how we can use deep neural networks to improve both model capacity and richness of representation. I'll demonstrate how this model leads to better results on two benchmark problems of similar form, and conclude by showing what the model actually learns.

Bayesian optimization has been successfully used to optimize complex black-box functions whose evaluations are expensive. In many applications, like in deep learning and predictive analytics, the optimization domain is itself complex and structured. In this work, we focus on use cases where this domain exhibits a known dependency structure. The benefit of leveraging this structure is twofold: we explore the search space more efficiently and posterior inference scales more favorably with the number of observations than Gaussian Process-based approaches published in the literature. We introduce a novel surrogate model for Bayesian optimization which combines independent Gaussian Processes with a linear model that encodes a tree-based dependency structure and can transfer information between overlapping decision sequences. We also design a specialized two-step acquisition function that explores the search space more effectively. Our experiments on synthetic tree-structured functions and the tuning of feedforward neural networks trained on a range of binary classification data sets show that our method compares favorably with competing approaches.

Sean Murphy will introduce you to the world of real-time online advertising. Amazon's approach to online advertising of catalogue items will be described how items are selected and how bids are calculated and with a focus on remaining machine learning challenges and future research directions.

Academic research on Machine Translation is naturally focusing on domains where public data is available. In this talk I will give an overview of the variety of non-standard use cases for Machine Translation at Amazon, and focus on the particular use case of product localization to illustrate the specific challenges and research questions we faced with Statistical Machine Translation in this domain. Secondly, I will discuss how Neural Machine Translation (NMT) may fit into this environment and give some details on implementing an NMT system in MXNet.

A key enabler for optimizing business processes is accurately estimating the probability distribution of a time series future given its past. Such probabilistic forecasts are crucial for example for reducing excess inventory in supply chains. In this talk, we present some of the forecasting challenges at Amazon and then introduce DeepAR, a novel methodology for producing accurate probabilistic forecasts, based on training an auto-regressive recurrent network model on a large number of related time series. We show through extensive empirical evaluation on several real-word forecasting data sets that our methodology is more accurate than state-of-the-art models, without requiring any manual feature engineering.

In this talk I will give an overview of our work in speech technology including our current approaches to modelling speech acoustics, work on multilingual and cross-lingual approaches, and automatic adaptation to the speaker and to the domain, with a focus on broadcast and media applications.

Billions of lines of source code have been written, many of which are freely available on the Internet. This code contains a wealth of implicit knowledge about how to write software that is easy to read, avoids common bugs, and uses popular libraries effectively.

We want to extract this implicit knowledge by analyzing source code text. To do this, we employ the same tools from machine learning and natural language processing that have been applied successfully to natural language text. After all, source code is also a means of human communication.

We present three new software engineering tools inspired by this insight:

Naturalize, a system that learns local coding conventions. It proposes revisions to names and to formatting so as to make code more consistent. A version that uses word embeddings has shown promise toward naming methods and classes.

Data mining methods have been widely applied to summarize the patterns about how programmers invoke libraries and APIs. We present a new method for mining market basket data, based on a simple generative probabilistic model, that resolves fundamental statistical pathologies that lurk in popular current data mining techniques.

HAGGIS, a system that learns local recurring syntactic patterns, which we call idioms. HAGGIS accomplishes this using a nonparametric Bayesian tree substitution grammar, and is delicious with whisky sauce.

Social media is becoming a hub for most of internet users to communicate, share their thoughts, report news, and express themselves. Large amount of research started to utilize the huge amount of data from social media in many different applications. In this talk, three example fields of research with social media data are presented. Initially, the task of Information retrieval is discussed, where the main objective is to retrieve the most relevant information needed from social media effectively and efficiently.

Secondly, data mining applications from social media is presented and demoed for applications such as text classification and sentiment analysis.

Thirdly, using social media for computational social science studies is presented with some focus on political issues. Few examples are presented, including: studying the antecedent of ISIS support on social media, public response towards Muslims after Paris attacks 2016, and Finally, the nature and dynamics of social media during the US Presidential Election 2016. Most of the work in this presentation was featured in news articles in popular press, such as CNN, Washington Post, the Independent, Aljazeera, the Daily Mail, and Mirror.

Movie analysis is an umbrella term for many tasks aiming to automatically interpret, extract, and summarize the content of a movie. Potential applications include generating shorter versions of scripts to help with the decision making process in a production company, enhancing movie recommendation engines by abstracting over specific keywords to more general concepts (e.g., thrillers with psychopaths), and notably generating movie previews. In this talk I will illustrate how NLP-based models together with video analysis can be used to facilitate various steps in the movie production pipeline. I will formalize the process of generating a shorter version of a movie as the task of finding an optimal chain of scenes and present a graph-based model that selects a chain by jointly optimizing its logical progression, diversity, and importance. I will then apply this framework to screenplay summarization, a task which could enhance script browsing and speed up reading time. I will also show that by aligning the screenplay to the movie, the model can generate movie previews with minimal modification. Finally, I will discuss how the computational analysis of movies can lead to tools that automatically create movie "profiles" which give a first impression of the movie by describing its plot, mood, location, or style.

In this talk I will give an overview of our research into neural machine translation which allowed us to conclusively demonstrate that neural machine translation can outperform previous state-of-the art phrase-based methods across a wide range of language pairs. I will show how applying subword units allow us to overcome limitations in vocabulary size and how using synthetic data allows us to train using in-domain target language data. I will also present research showing that modelling syntax on the target-side improves translation quality, especially for sentences with complex grammatical dependencies like questions and co-ordination. I will also describe two NMT toolkits that are being developed in Edinburgh.