Deep Learning Summit schedule

Embeddings in the Real World: Two Case Studies

We present two novel embedding mechanisms that are derived for two particular applications in search: spellchecker and queries. The two novel embeddings are designed to improve the quality of the underlying services with the constraint of not increasing the computational time necessary to process queries. We show how we tackled this problem along with some preliminary results on test datasets.

Fabrizio Silvestri is a Software Engineer at Facebook London in the Search Systems team. His interests are in web search in general and in particular his specialization is building systems to better interpret queries from search users. Prior to Facebook, Fabrizio was a principal scientist at Yahoo where he has worked on sponsored search and native ads within the Gemini project. Fabrizio holds a Ph.D. in Computer Science from the University of Pisa, Italy where he studied problems related to Web Information Retrieval with particular focus on Efficiency related problems like Caching, Collection Partitioning, and Distributed IR in general.

Probability & Uncertainty in Deep Learning

In this talk I will motivate the need for introducing probabilistic and Bayesian flavor to "traditional" deep learning approaches. For example, Bayesian treatment of neural network parameters is an elegant way of avoiding overfitting and "heuristics" in optimization, while providing a solid mathematical grounding. I will also highlight the deep Gaussian process family of approaches, which can be seen as non-parametric Bayesian neural networks. The Bayesian treatment of neural networks comes with mathematical intractabilities, therefore I will outline some of the approximate inference methods used to tackle these challenges.

I completed my PhD studies under Neil Lawrence in Sheffield, and subsequently pursued a post-doc in the intersection of machine learning and bio-inspired robotics. I've now moved to Amazon as a machine learning scientist, based in Cambridge, UK. My area of interest is machine learning, and more specifically: Bayesian non-parametrics (focusing on both data efficiency and scalability), representation and transfer learning, uncertainty quantification. In a recent strand of work I seek to bridge the gap between representation learning and decision-making, with applications in robotics and data science pipelines

Systems Challenges for Deep Learning

Training neural network models and deploying them in production poses a unique set of computing challenges. The ability to train large models fast, allows researchers to explore the model landscape quickly and push the boundaries of what is possible with Deep Learning. However a single training run often consumes several exaflops of compute and can take a month or more to finish. Similarly, some problem areas like speech synthesis and recognition have real time requirements which places a limit on how much time it can take to evaluate a model in production. In this presentation, I will talk about three systems challenges that need to be addressed so that we can continue to train and deploy rich neural network models.

Shubho is now working on AI Research at FAIR. He was previous a senior research scientist at Silicon Valley AI Lab (SVAIL) at Baidu Research.

I am an architect of the High Performance Computing inspired training platform that is used to train some of the largest recurrent neural network models in the world at SVAIL. I also spend a large part of my time exploring models for doing both speech recognition and speech synthesis and what it would take to train these model at scale and deploy them to hundreds of millions of our users. I am the primary author of the WarpCTC project that is used commonly for speech recognition. Before coming to the industry, I got my PhD in Computer Science from UCDavis focusing on parallel algorithms for GPU computing and subsequently went to Stanford for a Masters in Financial Math.

Considerations for Multi GPU Deep Learning Model Training

As the amount of available data grows, deep learning products and applications profit from the use of bigger and bigger models. To effectively train these models in a timely manner it is necessary to parallelize training over multiple GPUs and multiple machines. In this talk, we explore what is needed to efficiently achieve this kind of scaling, both from a hardware and library perspective, but also from the perspective of the end user. Methods of multi GPU parallelization are described and discussed, and best practices are presented. Furthermore, the choice of deep learning framework, and its impact on multi GPU training is discussed, along with the resulting tradeoff between flexibility and engineering effort. Finally, examples and benchmarking results are described and discussed, showing the possibility of near-linear scaling of training time both to multiple GPUs and multiple machines.

Jonas Lööf is a Deep Learning Solution Architect at NVIDIA, where he helps guide customer decision making on both hardware and software in their deep learning projects. Before joining NVIDIA, Jonas has worked in research and development, applying deep learning in the fields of speech recognition and natural language processing, both in a startup environment and the corporate world. Jonas holds a doctoral degree in computer science from RWTH Aachen University, where he worked on acoustic model adaptation for speech recognition.

Representations for Deep Learning

Current deep learning algorithms have achieved impressive results on a variety of tasks ranging from super-human image recognition to beating the world champion at the game of Go. Despite these successes deep learning algorithms still suffer from a variety of drawbacks: they require very large amounts of training data, they lack the ability to reason on an abstract level and their operation is largely opaque to humans. One way to overcome these problems is by creating models that form useful representations which exhibit beneficial properties such as disentanglement, the ability to generalise etc. This talk focusses on recent work that has addressed this issue for deep learning models.

Marta is a research scientist at DeepMind and also currently halfway through her PhD at Imperial College London under the supervision of Prof Murray Shanahan. Her research interests include deep generative models and reinforcement learning, in particular finding meaningful representations using the former to improve the latter.

Embeddings in the Real World: Two Case Studies

We present two novel embedding mechanisms that are derived for two particular applications in search: spellchecker and queries. The two novel embeddings are designed to improve the quality of the underlying services with the constraint of not increasing the computational time necessary to process queries. We show how we tackled this problem along with some preliminary results on test datasets.

Fabrizio Silvestri is a Software Engineer at Facebook London in the Search Systems team. His interests are in web search in general and in particular his specialization is building systems to better interpret queries from search users. Prior to Facebook, Fabrizio was a principal scientist at Yahoo where he has worked on sponsored search and native ads within the Gemini project. Fabrizio holds a Ph.D. in Computer Science from the University of Pisa, Italy where he studied problems related to Web Information Retrieval with particular focus on Efficiency related problems like Caching, Collection Partitioning, and Distributed IR in general.

Towards Teaching Machines to Read and Reason

We are getting better at teaching machines how to answer questions about content in natural language text. However, progress has been mostly restricted to extracting answers that are directly stated in text. In this talk, I will discuss our work towards teaching machines not only to read, but also to reason with what was read. We investigate two complementary approaches. In the first, we incorporate reasoning into neural models by enabling them to manipulate memory in a hierarchical sequence of steps—just like a virtual machine that executes a program. In the second, we take black box models (neural or not) and keep their structure as is. Instead, we develop rewards and penalties that encourage the model to behave as if it reasons. I will show the effectiveness of these approaches on various benchmark datasets, including relation reasoning and solving math word problems.

Sebastian is a Reader at University College London, leads the UCL Machine Reading group and is Co-Founder and Head of Research at Bloomsbury AI. He is an Allen Distinguished Investigator and received an $1M award from the Paul Allen Foundation to 'move the needle' towards answering broad scientific questions in AI. Sebastian is generally interested in the intersection of Natural Language Processing and Machine Learning, and particularly interested in teaching machines to read and to reason with what was read.

Labelling Topics Using Neural Networks

I am an Applied Scientist at Amazon, Cambridge, UK. Prior to that, I worked as a Research Associate at the Department of Computer Science at UCL and I completed a PhD in NLP at the Department of Computer Science of the University of Sheffield. My main research interests are in Natural Language Processing and Machine Learning. More specifically, I'm interested in applying statistical methods for detecting and interpreting the underlying topics in large volumes of text data. I also develop methods to analyse text and uncover patterns in data to solve problems in other scientific areas such as social and legal science.

Much of the information in large digital libraries is often stored in an unstructured way and is not organised using any automated system. That is usually overwhelming for users in a way that makes it difficult to find specific information or explore such data collections. A particular set of unsupervised statistical methods, namely topic models have been extensively used in Natural Language Processing and Information Retrieval for automatically analysing and organising document collections. Topics generated by topic models are typically presented as a list of terms. Automatic topic labelling is the task of generating a succinct label that summarises the theme or subject of a topic, with the intention of reducing the cognitive load of end-users when interpreting these topics. In this talk, I will present neural network approaches to labelling topics with text and images showcasing their effectiveness on providing meaningful representations of the topics.

Opinion Mining Using Heterogeneous Online Data

Understanding public opinion is important in many applications, such as improving company's product or service, marketing research, recommendation systems, decision and policy making and even predicting results of elections. Social media is a very powerful tool to transfer information and express emotions for users and a rich source of data that enables researchers to mine public opinion. In this talk I will describe different lines of work within our group, including rumour stance and veracity classification classification, predicting well-being based on heterogeneous user generated data and target-dependent sentiment recognition. False information circulating on social media presents many risks as social media is used as a source of news by many users. Detecting rumourous content is important to prevent the spread of false information which can affect important decisions and stock markets. Rumour stance classification is considered to be an important step towards rumour verification as claims that attract a lot of scepticism among users are more likely to be proven false later. Therefore performing stance classification well is expected to be useful in debunking false rumours. In our work we classify a set of Twitter posts from conversations discussing rumours as either supporting, denying, questioning or commenting on the underlying rumours. We propose a LSTM-based sequential model that achieves state-of-the-art results on this task through modelling the conversational structure of tweets. The task of automatically assessing well-being using smartphones and online social media is becoming of crucial importance, as an attempt to help individuals self-monitor their mental health state. In the current work, a multiple kernel learning approach is proposed as a mental health predictor, trained on heterogeneous (text and smartphone) user-generated data. The results reveal the efficiency of the proposed model and sequential approaches for time series modelling (i.e., LSTMs) are proposed for future work. Opinion mining is usually achieved by determining the overall sentiment expressed. However, inferring the sentiment towards specific targets is limited by such an approach since a Social Media posts may contain different types of sentiment expressed towards each of the targets mentioned. Our work on target-specific sentiment recognition goes beyond tweet-level or single-target approaches, and proposes a multi-target-specific sentiment classification model, which explores the context around a target as well as syntactic dependencies involving the target.

I am a Computer Science PhD student at the University of Warwick supervised by Dr. Maria Liakata and Prof. Rob Procter. Currently I am based at the Alan Turing Institute in London. My background is Applied mathematics and Complexity science. I work in the area of Natural Language Processing. My research is focused on Rumour Stance and Veracity Classification in Twitter conversations. I am studying the benefits of utilising the conversation structure in supervised learning models.

Applications of Deep Learning: Anomaly Detection, Sentiment Classification, and Over-sampling

The talk will present my latest work in deep learning for advancing three strands: I) anomaly detection using stacked denoising auto-encoders. II) Sentiment analysis of variable size datasets using a combined LDA and deep learning approach III) deep over-sampling to overcome class bias in highly imbalance datasets. The methods are demonstrated through real-life problems including: spam filtering, sentiment analysis, and Brain-Computer Interfaces.

Dr. Al Moubayed is an Assistant Professor at the school of computer science in Durham University. Her main research interest is in unsupervised deep learning, natural language processing, and optimisation. Dr. Almoubayed obtained her PhD from the Robert Gordon University, followed by post-doctoral positions in the University of Glasgow and Durham University. She developed machine learning and deep learning solutions in the areas of social signal processing, cyber-security, and Brain-Computer Interfaces. All of which involve high dimensional, noisy and imbalance data challenges.

Cross-Modal Machine Translation

The advances on neural machine translation across natural language have opened new venues in the field of cross-modal analysis. Given the unified the machine learning framework broadly adopted by the language and vision communities, novel opportunities have arisen by using deep learning framework to transform across modalities. This talk will provide an overview of the state of the art on cross-modal translation and (eg. lipreading, facial animation, sign language) and present our work in speaker visualization from speech.

Xavier Giro-i-Nieto is a learning enthusiast working as an associate professor at the Universitat Politecnica de Catalunya (UPC), in Barcelona, and a certified instructor at the NVIDIA Deep Learning Institute. He has been a visiting scholar at Columbia University and works regularly with Dublin City University, the Barcelona Supercomputing Center and Vilynx. His research interests focus on deep learning for computer vision, speech and natural language processing. His current service includes associate editor of the IEEE Transactions in Multimedia. Xavier Giro-i-Nieto is a learning enthusiast working as an associate professor at the Universitat Politecnica de Catalunya (UPC), in Barcelona, and a certified instructor at the NVIDIA Deep Learning Institute. He has been a visiting scholar at Columbia University and works regularly with Dublin City University, the Barcelona Supercomputing Center and Vilynx. His research interests focus on deep learning for computer vision and natural language processing applied to large scale image retrieval, affective computing, lifelogging from wearables and visual saliency prediction. His current service includes associate editor of the IEEE Transactions in Multimedia and ACM SIGMM Records.

Iterative Approach To Improving Sample Generation

Generative modelling allows us to synthesise new data samples, whether the samples be used for imagining new concepts, augmenting datasets or better understanding our datasets. When synthesising samples, generative models don’t always get it right first time - the samples may not be sharp, may have artefacts or may be nonsensical. I will present recent research that shows how we can improve these samples, by applying an iterative procedure to get slightly better samples with each step.

Antonia is a PhD candidate at Imperial College London, in the Bio-Inspired Computer Vision Group. Her research focuses on unsupervised learning and generative models. She received her masters in Biomedical Engineering from Imperial College London with an exchange year at the University of California, Davis. Antonia has interned at DeepMind, Twitter (Magic Pony), Cortexica and UNMADE.

Discovering and synthesizing novel concepts with minimal supervision

Scaling up visual category recognition to large numbers of classes remains challenging. A promising research direction is zero-shot learning, which does not require any training data to recognize new classes, but rather relies on some form of auxiliary information describing new classes. We propose and compare different class embeddings learned automatically from unlabeled text corpora, expert annotated attributes and detailed visual descriptions. Furthermore, we use detailed visual descriptions to generate images from scratch and to generate visual explanations which justify a classification decision. Finally, we explore humans' natural ability to determine distinguishing properties of unknown objects through gaze fixations.

Zeynep Akata is an Assistant Professor at the University of Amsterdam and a Senior Researcher at the Max Planck Institute for Informatics. She received a MSc degree in 2011 from RWTH Aachen and a PhD degree in 2014 from INRIA Grenoble. Her research interests include machine learning with applications to computer vision, such as zero-shot learning and multimodal deep learning with generative models that combine vision and language. She received Lise-Meitner Award for Excellent Women in Computer Science from Max Planck Society in 2014 and a DARPA grant Explainable Artificial Intelligence in 2017 in collaboration with UC Berkeley.

Cooperative AI for Critical Application

Deep Learning or AI excel at prediction accuracy, but hard to interprete, which becomes a key barrier for many impactful critical applications, such like surgery and finance. Dr Fangde Liu from DSI will talk about their research and idea to make the machine learning system interpretative, controllable, data efficient to train and with fault detection functionality. In his talk , he will talk about the Deep Poincare Map algorithm that combine the dynamic theory with deep learning to perform medical image analysis, and also the autonomous vigiliant management system for one line AI applications, the TensorDB framework.

Dr Fangde Liu, UK National Exceptional Talent, Senior Associate of Royal Society of Medicine, Member of British Machine Vision Society, is an expert in Data Science, Robotics and Surgery. At the Data Science Institute, he manages the engineering development of several big data healthcare projects of global impact. The Optimise-MS, an international joint effort for cloud based autonomous system for effectiveness, performance and post market pharmacovigilance, has ben deployed at about 20 leading clinical institutes across the UK, Belgium, Poland and Finland. As it becomes the technology foundation for neurology registry of several European nations, in the coming decade, it will deployed and monitor over $10 Billion neurology drugs usage yearly across 90 countries. His earlier works on real-time image guided surgery navigation system and robotic operating system on GPU are key technologies of several EU flagship surgical robot projects for endovasc ular surgery, neurosurgery, and orthopaedics, including the EDEN2020, the biggest ever surgical robot project funded in EU. His AI research focus on cooperative AI system aiming breaking down the barriers for critical applications. He focuses on the idea combining domain knowledge and deep learning to make AI system interpretable, controllable and data efficient, he also works on the vigilance management of AI for critical applications. Dr Liu is very experienced and has great insight applying technology in clinical medicine. He works closely with world top clinicians, industry partners and leading academics, such as Johnson&Johnson, Biogen, GE, Renishaw, St Mary Hospital, Mayo Clinic. As an official consultant of Imperial College, he has also provided consulting service on clinical applications of AI to technology giants such as IBM Watson, Huawei Research, NEC research, Samsung Electronics HNA group and AliCloud Europe.

Artificial Intelligence for Space-Based Technology

Understanding change is essential to addressing our most pressing global challenges. As populations grow, infrastructure must expand. As globalization spreads, businesses must react. As the environment changes, humans must adapt.

To make sense of these developments, we need actionable insight and informed collaboration now more than ever. Today, the ability to observe, analyze and monitor our planet is unprecedented.

DigitalGlobe owns and operates the most sophisticated satellite imaging constellation. This technology has enabled an image library spanning over 25 years and amounting to over 100 Petabytes of high resolution imagery. To be able to process and make sense of such as vast amount of data, at scale, DigitalGlobe is pioneering the use of deep learning and artificial intelligence. From establishing economic indicators, to automatedly detect objects and make predictions about future events, artificial intelligence applied to geospatial information is helping reveal insights that would otherwise be hidden.

Luca has over 5 years experience in developing applications that rely on space-based radar and optical data. Luca currently serves as a Technical Engineer at DigitalGlobe where he is responsible for creating solutions based on satellite imagery, geospatial big data and deep learning algorithms. Academically, Luca holds an MSc in signal processing and remote sensing.

Attendees and speakers from the Deep Learning Summit & AI Assistant Summit will come together for an evening of discussions, collaborations and merriment after the first day of presentations. Network with your peers and make new contacts. Diners will include CTOs, CEOs, Data Scientists, Founders, Professors and Engineers.

Enjoy a champagne reception with views over the River Thames followed by dinner. Attendees will enjoy extended networking over 3 courses of the finest British cuisine from Chef Alan with wine to match and compliment the evening.

This dinner is only open to attendees of the Deep Learning Summit and AI Assistant Summit. For enquiries about the event please contact [email protected]