Deep Learning Summit schedule

Deep Learning for Embedded Devices: The Next Step in Privacy-Preserving High-Precision Mobile Health and Wellbeing Tools

State-of-the-art models that, for example, recognize a face, track emotions, or monitor activity are increasingly based on deep learning principles. But bleeding-edge health tools, like smartphone apps and wearables, that require such user information must rely on less reliable learning methods to locally process data because of the excessive device resources demanded by deep models. In this talk, I will describe our research that drives towards a complete rethinking of how existing forms of deep learning executes at inference time on embedded health platforms. Not only does this cause radically lower energy, computation and memory requirements; it also significantly increases the utilization of commodity processors (e.g., GPUs, CPUs) -- and even emerging purpose-built hardware, when available.

Nic Lane is a Principal Scientist at Bell Labs where he is a member of the Internet of Things research group. Before joining Bell Labs, he spent four years as a Lead Researcher at Microsoft Research based in Beijing. Nic received his Ph.D. from Dartmouth College (2011), his dissertation pioneered community-guided techniques for learning models of human behavior. These algorithms enable mobile sensing systems to better cope with diverse user populations and conditions routinely encountered in the real-world. More broadly, Nic's research interests revolve around the systems and modeling challenges that arise when computers collect and reason about people-centric sensor data. At heart, he is an experimentalist who likes to build prototype sensing systems based on well-founded computational models.

Input Tailored Concepts Maps - A Way to Facilitate AI for Navigating Scientific Knowledge

Iris AI uses non-semantic models to apply text understanding techniques to a scientific body of knowledge. The current algorithm is a mixture of neural topic models, keyword extraction and heuristics functions to form an input tailored concept map filled in with scientific articles. The current approach facilitates both existing unsupervised as well as supervised techniques for AI training. To verify our current progress we use state-of-the-art metrics, but moreover we conduct real life experiments comparing our tool to existing tools in the field through sci-thons. The results show great potential for new techniques in text understanding, and will change the way people navigate scientific knowledge.

Victor Botev is the CTO of Iris AI. Before joinin Iris he was an Artificial Intelligence Research Engineer at Chalmers University, in Gothenburg, Sweden. He has conducted research on clustering and predictive neural networks models, as well as usage of signal processing techniques in studying Big Data. As a Masters Thesis Student at CPAC Systems AB he has worked in the development of an autonomous compactor for pavement. Previously he was a senior software developer at Pinexo, a tech lead at Skrill and a web developer at Seedburger AG. He also has a second Masters (Artificial Intelligence) and BSc. (Software Development) degrees from Sofia University.

3D Printing with Autonomous Robotics and AI

The presentation will showcase the recent works of Ai Build, a London based company developing Artificial Intelligence and Additive Manufacturing technologies for the built environment. Current methods of construction are wasteful, labour intensive, unsafe and time consuming. They are also incapable of creating complex forms. Ai Build’s large scale robotic 3D printing technology is combining computer vision and machine learning techniques to automate low volume manufacturing process to enable mass customization of designs by architects, designers and engineers.

Daghan Cam is the co-founder and CEO of Ai Build, a London based startup developing Artificial Intelligence and Additive Manufacturing technologies for the built environment. He is also a visiting lecturer at University College London doing research on robotic fabrication, large scale 3d printing and parallel algorithms with GPU computing. His work focuses on developing intelligence for automating complex tasks in design and manufacturing by using computer vision and machine learning techniques. Before starting Ai Build, he worked at Zaha Hadid Architects and ran his own architectural design practice. He holds a masters degree with distinction from the Architectural Association.

Thanks to deep learning, AI algorithms can now surpass human performance in image classification. However, behind these results lie tens of thousands of man hours spent annotating images. This significantly prohibits commercial applications where cost and time to market are key. At Tractable, our solution centers on creating a feedback loop from learning algorithm to human, turning the latter into a “teacher” rather than a blind labeler. Dimensionality reduction, information retrieval and transfer learning are some of our core proprietary techniques. We will demonstrate a 15x labeling cost reduction on the expert task of estimating from images the cost to repair a damaged vehicle – an important application for the insurance industry.

Alex is Co-founder & CEO of Tractable, a young London-based startup bringing recent breakthroughs in AI to industry. Tractable's current focus is on automating visual recognition tasks. Its long term vision is to expand into natural language, robot control, and spread disruptive AI throughout industry. Tractable was founded in 2014 and is backed by $2M of venture capital from Silicon Valley investors, led by Zetta Venture Partners. Alex has a degree in econometrics & mathematical economics from the LSE, and a postgraduate degree in computer science from Imperial College London. Alex's experience within Deep Learning investing is on the receiving side, particularly on how to attract US venture capital into Europe as early as the seed stage.

Layer Cake - AI in Food

Deep Learning demonstrates impressive results these days and NLP is definitely among the areas which can benefit significantly. Whisk redesigned their core engine for recipe understanding and product matching with Deep Bidirectional LSTM networks on top of TensorFlow library. This solution not only immediately outperformed previous methods with context-free grammars, but also it is way easier to maintain. The new system allows Whisk to operate on bigger scale of recipes with significantly reduced need for manual intervention and keep improving as Whisk expands it’s training set.

As CTO of Whisk.com, Viktor is now leading an agile team of 10 people following Lean practices when developing products. With extensive implementational and architectural experience with distributed technologies and compute engines, Viktor has been one of the technology leaders behind Whisk for 5 years. Recently he has been building a data science team who have been implementing modern Machine Learning approaches and techniques to scale parts of Whisk.

Building AI to Better Understand & Respond to Your Customers

Businesses are talking to their customers more and more. Every customer conversation is an opportunity to directly engage with your customer, to recommend them products, to better understand them and the user's experience. But supporting these conversations and extracting insights from them is expensive and difficult. re:infer provides a solution powered by the latest advances in AI that does the heavy lifting to help businesses better understand and interact with their customers. Using Machine Learning to understand customer conversations is hard. More often than not, customers use informal, poorly spelt and ambiguous language. Because of this, traditional computer science techniques that use hand coded rules fail. And since most businesses have insufficient training data, traditional machine learning methods do not work in this context either. To solve this problem we take a different approach. In this talk I'll describe the business problem we're solving, the engineering constraints it imposes and how we've built a novel, deep learning, natural language processing system to solve it.

AI and machine learning has dominated my entire professional career — I’ve been working in this space for over 10 years. I gained my PhD and MSc in AI from the research groups at UCL and Edinburgh. My research has been published in the leading AI journals and conferences including NIPs, AISTATs, Neural Computation, the Journal of Machine Learning Research and NeuroImage. Outside of academia and before re:infer, I helped to build production AI technologies for problems in search, ad-tech, consumer intent modelling and finance.

Source Code Abstracts Classification Using CNN

Convolutional neural networks (CNN) are becoming the standard approach for many machine learning related problems. Usually, those problems are related to images, audio or natural language data. At source{d} we are trying to apply the common and novel deep learning patterns to the problems with software developers and projects as the input which is something very different. We are standing at the beginning of our fascinating journey, but already have something to share. In this particular talk I am going to present the bits of our SourceNN deep neural network that enable classification of short source code fragments (50 lines) taken randomly from several projects. The input features are extracted by a syntax highlighter and look similar to minimaps in source code editors.

Eiso Kant is the co-founder & CEO of source{d}.

Diabetic retinopathy is when there is retinal damage in the eye due to diabetes, potentially leading to loss of vision and even blindness. In his talk Jeffrey will reflect on his experience of trying to build a model, using convolutional neural networks, to grade the severity of diabetic retinopathy in high-resolution fundus images (images of the back of the eye). He did this work in the context of the Kaggle Diabetic Retinopathy Detection competition where he finished fifth.

Jeffrey De Fauw studied pure mathematics at Ghent University before becoming more interested in machine learning problems through Kaggle competitions. Soon after he was introduced to (convolutional) neural networks and has since spent most of his time working with them. Besides always looking for challenging problems to work on, he has also become very interested in trying to find more algebraic structure in methods of representation learning.

AI is Changing the Drug Discovery Paradigm

Drug discovery is a challenging business, despite huge societal and commercial benefits in the discovery of new drugs it is incredibly challenging to develop discover and develop new therapies, with typically around 30 new drugs developed per year from the entire worldwide pharma and biotech R&D budget. The reasons for this are complex, but the bottom line is that the vast majority of started projects do not successfully finish, there is huge attrition from the idea of a scientist through discovery and clinical development stages. We are developing powerful real world evidence-based artificial intelligence solutions to address drug discovery. Key to recent progress is the availability of large quantities of data, high performance computing and developments in deep-learning approaches to mine for hypotheses that can be rationally scored and prioritised for success.

John studied Chemistry at Bath, graduating in 1987. He then studied for a PhD at Birkbeck College, on protein modelling, followed by a postdoc at ICRF (now CRUK). John then joined Pfizer, eventually leading a multidisciplinary group combining rational drug design, informatics and structural biology. In 2000 he moved to a start-up biotech company, Inpharmatica, where he developed the drug discovery database StARLite. In 2008 John moved to the EMBL-EBI, where the successor resource is known as ChEMBL. Most recently John joined Benevolent.ai, where he continues his research as director of bioinformatics. In this role, John is involved in integrating deep learning and other AI approaches to drug target validation and drug optimisation

How Artificial Intelligence will Change the Face of Healthcare

Everyone in the world is facing differing degrees of the same issue - the accessibility and affordability of healthcare. For some, the problem is convenience, cost or speed. For others, the issue is more serious with almost 50% of the world having little access to quality healthcare. Yet, four unstoppable trends are coming together to see the creative reconstruction of medicine within the next decade. The result will be a service that is more accessible, effective and democratic, irrespective of where people live. Today, everyone has near equal access to everything that is digital. The same may soon be happening to healthcare, and these trends show why: 1 - Diagnostics is improving at double the rate of Moore’s Law 2 - Information is free and getting smarter 3 - Smartphones and “the internet of everything” will create a global channel of healthcare delivery 4 - Intervention will be unrecognisable

These four trends are melting all that is solid in medicine into air. How it will develop no one can know, but one thing is for sure; a very different model of healthcare delivery is unfolding, and it should make the future of healthcare significantly smarter and better value for everyone.

Ali is an engineer and healthcare entrepreneur, and the founder and CEO of babylon, the UK’s leading digital healthcare service. Its purpose is to democratise healthcare by putting an accessible and affordable health service into the hands of every person on earth. In order to achieve this the company is bringing together one of the largest teams of scientists, clinicians, mathematicians and engineers to focus on combining the ever-growing computing power of machines with the best medical expertise of humans to create a comprehensive, immediate and personalised health service and make it universally available. Launched in February 2015, the service now has over 600,000 registered users globally.. babylon’s home grown success has seen the company expand into Europe and Africa in 2016, with a second head office located in Rwanda. Around 120 businesses, including Citigroup, BNY Mellon, LinkedIn and leading employee benefits and health insurance providers have partnered with babylon to offer its services to UK employees. Further, the company has partnered with the NHS to make its services available to the broader UK population. Prior to Babylon, Ali founded Circle and built it within a few years to become Europe's largest partnership of clinicians, with some £200m of revenue, near 3,000 employees and a successful IPO. Earlier, Ali was the recipient of the Royal Award for the Young Entrepreneur of the year for founding his first business, V&G, and the Healthcare Entrepreneurial Achievement Award for establishing Circle. Ali was named by The Times among the 100 global people to watch, and by HSJ among the 50 most influential people in healthcare. Ali is the UK Cabinet Office Ambassador for Mutuals and has a PhD in Engineering Physics.

Deep Learning for Embedded Devices: The Next Step in Privacy-Preserving High-Precision Mobile Health and Wellbeing Tools

State-of-the-art models that, for example, recognize a face, track emotions, or monitor activity are increasingly based on deep learning principles. But bleeding-edge health tools, like smartphone apps and wearables, that require such user information must rely on less reliable learning methods to locally process data because of the excessive device resources demanded by deep models. In this talk, I will describe our research that drives towards a complete rethinking of how existing forms of deep learning executes at inference time on embedded health platforms. Not only does this cause radically lower energy, computation and memory requirements; it also significantly increases the utilization of commodity processors (e.g., GPUs, CPUs) -- and even emerging purpose-built hardware, when available.

Nic Lane is a Principal Scientist at Bell Labs where he is a member of the Internet of Things research group. Before joining Bell Labs, he spent four years as a Lead Researcher at Microsoft Research based in Beijing. Nic received his Ph.D. from Dartmouth College (2011), his dissertation pioneered community-guided techniques for learning models of human behavior. These algorithms enable mobile sensing systems to better cope with diverse user populations and conditions routinely encountered in the real-world. More broadly, Nic's research interests revolve around the systems and modeling challenges that arise when computers collect and reason about people-centric sensor data. At heart, he is an experimentalist who likes to build prototype sensing systems based on well-founded computational models.

Armando Vieira is a Physicist turned into a data scientist. He started working on machine learning after his PhD in Physics in 1997. From the beginning he is an aficionado of Artificial Neural Networks, and recently he is focused on Deep Neural Networks, especially for unsupervised and semi-supervised learning problems. He has more than 50 publications and writing a book on business applications of Deep Learning. He works as a Data Scientist consultant on several companies and startups.

Ingmar is an Associate Professor in Engineering Science at the University of Oxford specialising in applied machine learning solutions for robot perception and decision making. He is a long-standing member of the Mobile Robotics Group (now the Oxford Robotics Institute) where he leads research in machine perception and planning. His research is guided by his vision to create machines which constantly improve through use in their dedicated workspace by implicitly leveraging expert demonstrations in a manner entirely transparent to the user. Highlights of his work include state-of-the-art approaches to deep and shallow object detection, semantic segmentation, tracking and inverse reinforcement learning. Ingmar has coauthored over 40 research publications and is the recipient of a number of best paper awards at international robotics conferences such as ISER and ICAPS. He serves on the board of IJRR, the premier international robotics research journal and has repeatedly served as area chair and programme committee member for reputed conferences in robotics and machine learning. Recently Ingmar led a team to develop and demonstrate the first autonomous urban concept vehicle on a purpose built slow-speed racetrack at Shell’s Make the Future London event. In 2014 Ingmar also co-founded Oxbotica, a leading provider of mobile autonomy software solutions including the Selenium autonomy stack, which underpins a variety of public and commercial autonomous vehicle programmes such as the LUTZ and GATEWay projects in Milton Keynes and Greenwich. In 2015 Oxbotica was singled out by the Wall Street Journal as one of the top ten EMEA technology startups.

Conditional Convolutional Neural Network for Modality-aware Face Recognition

We propose a conditional Convolutional Neural Network, named as c-CNN, to handle multimodal face recognition. Different from traditional CNN that adopts fixed convolution kernels, samples in c-CNN are processed with dynamically activated sets of kernels. The activations of convolution kernels in a certain layer are conditioned on its present intermediate representation and the activation status in the lower layers. The activated kernels across layers define the sample-specific adaptive routes that reveal the distribution of underlying modalities. The proposed method is implemented via incorporating the binary decision tree, which is evaluated for multimodality face recognition problems.

T-K Kim is Associate Professor and the leader of Computer Vision and Learning Lab at Imperial College London, UK. He obtained his PhD and Junior Research Fellowship, from Univ. of Cambridge, in 2008 and for 2007-2010 respectively. His research interests primarily lie in tree-structure classifiers for articulated hand pose estimation, face recognition by image sets and videos, and 6D object pose estimation. He has co-authored over 40 papers in top-tier conferences and journals, his co-authored algorithm for face image retrieval is MPEG-7 ISO/IEC standard. He is co-recipient of the KUKA best paper award at ICRA14, and general co-chair of CVPR15/16 workshop on HANDS and ICCV15 workshop on Object Pose.

Autoencoding Blade Runner

Generative models for images have come a long way in recent years. This talk recalls the creative exploration of an unconventional use of a convolutional autoencoder trained with a learned similarity metric. By using all the frames from Blade Runner as the training dataset for this model, we have a generative model that has learned the distribution of scenes from Blade Runner. We use this model to reinterpret the film, and to reinterpret other films based on its understanding of Blade Runner.

Terence is an artist and research engineer at Goldsmiths, where he recently completed his Masters in Creative Computing. His focus has been on the generative capabilities and creative applications of deep learning, developing an interactive topological visualisation of a convolutional neural network, and remaking the film Blade Runner with a convolutional autoencoder.

The breakout session is an opportunity to advertise job openings in your company to 400+ attendees at the event; or to raise your profile if you are looking for a new role. Participants include:

Tyler Capital: What we “do” is trade the markets. What’s really cool is how we do this. We are at the forefront of adaptive learning and cutting edge technology. We are building deeply automated systems, using some of the most sophisticated machine learning techniques in the world. Some people in our industry think we are crazy for this approach; we think they are redundant. We want to surround ourselves with people who share our values, and our love for machine learning. We are as friendly as we are fierce in our determination to compete in the markets in this particular way.

COSMONiO® designs cutting-edge computer-vision and machine-learning systems that automate the process of extracting visual information from images even under the most challenging conditions. What sets COSMONiO apart from the competition is the universal nature of its software. Our projects cover a wide range of applications from single-cell micro-organism analysis to outer space exploration. COSMONiO are seeking an enthusiastic Deep Neural Network Engineer to join a small R&D team.

HAL24K is a Data Intelligence Research Lab based in Amsterdam, San Francisco and London. We help smart cities, the companies they outsource to, and transport infrastructure companies, to make objective and data-driven decisions by leveraging machine learning algorithms and artificial intelligence techniques.

Logikk is a boutique recruitment business with a focus on Data and Analytics. We are passionate about delivering service over sales, and provide our clients with bespoke solutions across Contract, Interim, Permanent and Executive Search.

Oxford Learning Solutions (OLS) have developed a new way to learn to read, based on an approach they call the Trainertext Method. Trainertext uses visual images to help the child decode the sounds in each word and they learn to read fluently through short daily decoding sessions. The OLS team have moved into AIEd with adaptive algorithms in a project supported by Innovate UK. Their aim in 2017 is to extend this with predictive AI to optimise the learner journey for each student. OLS are keen to talk to potential partners or new colleagues, who have experience in predictive AI.

At AimBrain, we are revolutionising the way we do authentication. We are developing a state-of-the-art biometric authentication platform to allow easy access to robust biometric authentication in any mobile or IoT device. Join our highly technical founding team as we make biometrics the main authentication method of the future. AimBrain is looking for a Machine Learning Software Engineer (Biometric Authentication).

source{d} has created a recruitment service built by developers for developers. We find developers by looking at real work samples, analyzing over 900 million code contributions of more than 6 million developers in the open source community through deep learning algorithms to find the best matches between companies and developers.

Textkernel is developing cutting edge semantic search and matching software for a fast growing number of international clients. We are looking for people who are as passionate about this as we are, and who combine an interest in advanced computing and R&D with the ambition to deliver outstanding products. The positions we are currently looking for include: NLP research engineer (Deep Learning is a bonus), Java engineer, search engine backend, Software Engineer Python & Big Data.

Tractable is a deep learning startup specialising in computer vision to solve specific high impact problems in industry. Our proprietary technology enables a 100x cost reduction in labelling – the current bottleneck in the development process. Our commercial approach is to identify impactful applications, acquire strategic datasets and integrate within the data flow. Our current focus is insurance, where our AI product produces expert review of auto claims at scale. We are equally experiencing strong demand in utilities, oil & gas and medical imaging. Tractable are looking to fill 3 roles: Deep Learning Researcher, Deep Learning Engineer and Full Stack Software Engineer.