Building A Useful Chatbot: Beyond the ML and NLP

About two years ago, chatbots seemed to be the next big thing since mobile apps. In the meantime, things have cooled down a lot, with chatbots failing to deliver on the expectations. However, conversational AI is still moving forward in great strides. So, how can companies avoid the chatbot bubble and still achieve impact with the latest conversational technology?

Deep learning for automated language teaching and assessment

Automated language teaching and assessment offers the opportunity to increase language learning efficiency, and open access to learning worldwide. To develop effective automated models, we need to emulate human behaviour in language instruction and how value judgements about someone's language proficiency are made. In this talk, I will describe recent developments in deep learning for the automated assessment of text produced by non-native learners of English. I will discuss how we can overcome some of the challenges we face with learner data and develop models that provide immediate and detailed feedback -- a fundamental part of language instruction. I will then conclude with future directions in the field.

Data-driven automatic sign language processing

Sign languages are fully natural languages with their own grammars and vocabularies. Automatic sign language processing, a sub-field of natural language processing (NLP), comprises tasks such as automatic sign language translation, sign language recognition, and sign language synthesis. The present talk introduces each of these NLP applications, presenting the state of the art and the remaining challenges, with a focus on data-driven methods. At the end, the potential of combining the individual applications into an overall information and communication technology (ICT) solution for deaf sign language users will be discussed.

Using AI in Women’s Health

Data science can empower women by providing powerful health insights throughout their entire reproductive lives. Ava has developed a combination of machine learning and wearable technologies to help couples conceive. By measuring seven physiological parameters while sleeping, the effects of hormone changes throughout the menstrual cycle can be tracked to detect a woman’s fertile days. As the user database has grown from non-existent to “big data”, suitable detection algorithms have evolved from expert systems to deep learning.

Artificial Intelligence approaches for personalized medicine

In recent years, deep learning has become one of most active fields in machine learning with astounding performances in a broad area of applications such as computer vision, speech recognition and natural language processing. In computational biology, the recent availability of large amounts of data generated by word-wide consortia together with technical developments facilitating the implementation and training of more performant models have made possible the broad application of deep learning to a vast set of problems. In this talk, I will present current activities at the Computational Systems Biology group in IBM Research, Zurich, that illustrate the application of AI approaches to integrate disparate data types. Specifically, I will explain how a multi-modal neural network can be trained to ingest disparate data types, such as compound molecular structure, transcriptomic data and prior molecular knowledge, and predict drug sensitivity in cancer cell lines.

Combining remote sensor data capture with advanced signal processing and machine learning to innovate clinical research for neurodegenerative and psychiatric diseases

Remote patient monitoring in clinical trials of neurodegenerative and psychiatric diseases using smartphones, wearables and other sensors can provide rich information on disease status and progression that might not be captured during in-frequent clinical visits. While patients perform dedicated daily ‘active’ tests on their smartphone tailored to specific disease pathophysiology or just go about their daily activities (‘passive monitoring’), we are collecting large data sets from acceleration, gyroscope, magnetometer and other sensors embedded into wearable devices. Using dedicated signal processing algorithms in combination with statistical methods and machine learning we extract information about disease symptom severity as well as the influence of the disease on the daily life of a patient. Data collected in different studies and diseases show strong agreement between remotely collected sensor signals and clinical assessments. Furthermore, examples from Multiple Sclerosis, Parkinson’s Disease and Schizophrenia demonstrate that remote patient monitoring can augment and extend our understanding of these severe diseases which pose a high burden on patients, families and global health at large.

Algorithm Selection and Configuration with Monte-Carlo Tree Search

The AutoML task consists of selecting the proper algorithm in a machine learning portfolio, and its hyperparameter values, in order to deliver the best performance on the dataset at hand. This task is key to the knowledge transfer from research labs to industry. A Monte-Carlo tree search (MCTS)-based approach is presented to handle the AutoML hybrid structural and parametric expensive black-box optimization problem. Extensive empirical studies are conducted to independently assess and compare: i) the optimization processes based on Bayesian optimization or MCTS; ii) its warm-start initialization; iii) the ensembling of the solutions gathered along the search. The proposed approach is assessed on the OpenML 100 benchmark and the Scikit-learn portfolio, with statistically significant gains over AutoSklearn, winner of former international AutoML challenges.

Causality and robust machine learning

Deep neural networks have achieved outstanding performance on prediction tasks like visual object recognition. These current algorithms excel at discovering and exploiting dependencies in the training data for prediction. However, when the distribution encountered at test time differs in some respects from the training distribution, predictive performance often degrades considerably. Such “domain shifts” can be caused by changing conditions such as color, background or location changes. In this talk, I will discuss these distribution shifts from a causal viewpoint and present so-called “conditional variance penalties” which increase the robustness of estimators under domain shifts.

Bayesian Overlap Clustering for Distance Data

We present a Probabilistic model for Overlapping Clusters on Distance data (POCD) which enables the modeling of overlapping clusters where objects are only available as pairwise distances. Examples of such distance data are genomic string alignments, protein contact maps or pairwise patient similarities. Even if it is possible to embed the distance data into a vector space, it is preferable to work directly with the distance matrix to avoid unnecessary bias and variance which can be caused by embeddings. Currently, there are no probabilistic methods that infer overlapping clusters for distance data and POCD aims to fill this gap.

The Power of Networks

Forget about your tables and SQL queries for this talk. Think of your data as a network - represented as nodes and their relationships (edges). This provides different opportunities for data scientists to explore data and extract knowledge using graph algorithms and graph visualization. In this talk, we give two examples of how representing, querying, and visualizing data as networks empowers users to effectively and efficiently understand data and make decisions.

Automatic Understanding of the Visual World

One of the central problems of artificial intelligence is machine perception, i.e., the ability to understand the visual world based on input from sensors such as cameras. In this talk, I will present recent progress of my team in this direction. I will start with presenting results on how to generate additional training data using weak annotations, motion information and synthetic data. Next, I will discuss our results for action recognition in videos, where human tubelets have shown to be successful. Our tubelet approach moves away from state-of-the-art frame based approaches and improve classification and localization by relying on joint information from several frames. We show how to extend this type of method to weakly supervised learning of actions, which allows us to scale to large amounts of data with sparse manual annotation. Finally, I briefly present recent work on grasping with a robot arm based on learning long-horizon manipulations with a hierarchy of RL and imitation-based skills.

Data Driven Experiences that Matter

Every organization in the world—large or small, new or mature, private or public—needs to look outward to close the massive gaps between the experiences organizations think they are delivering and what is really happening. Experience matters. Experiences are most meaningful when the data is accurate, relevant, trustworthy and personal. Hear the latest approach to how organizations can leverage data and insights to create meaningful, exceptional experiences that transform customers into loyal brand advocates.

Data driven technologies: challenges in industrial applications

The last few years saw a tremendous interest in the use of advanced technologies to extract insights from data and help improve business operations. However, a number of challenges still exist such as data acquisition from different types of sources, establishing links among different data types using both structure and content, scalability for analytics and query processing, just to name a few. In this presentation we will use industrial applications to motivate some of the advanced technologies, point out the practical and technical challenges, and highlight the business benefit from these applications.

Key Success Factors in Building Up an Analytics and Data Science Function

Despite the significant investments made in the space of Big Data, Analytics and Data Science, the majority of programs and initiatives in this space fail to live up to its promises and expectations beyond the PoC stage. In this talk, Tanvi Singh outlines the key success factors behind building and scaling a successful Data Science team and managing a broad portfolio of projects with 50+ use cases. She also provides insights into how she and her team continuously innovate and work on new ideas in a Sandbox setup.

Predicting aesthetic appreciation of images

Image aesthetics has become an important criterion for visual content curation on social media sites and media content repositories. Previous work on aesthetic prediction models in the computer vision community has focused on aesthetic score prediction or binary image labeling. However, raw aesthetic annotations are in the form of score histograms and provide richer and more precise information than binary labels or mean scores. In this talk I will present recent work at Naver Labs Europe on the rarely-studied problem of predicting aesthetic score distributions. The talk will cover the large-scale dataset we collected for this problem, called AVA, and will describe the novel deep architecture and training procedure for our score distribution model. Our model achieves state-of-the-art results on AVA for three tasks: (i) aesthetic quality classification; (ii) aesthetic score regression; and (iii) aesthetic score distribution prediction, all while using one model trained only for the distribution prediction task. I will also discuss our proposed method for modifying an image such that its predicted aesthetics changes, and describe how this modification can be used to gain insight into our model.

On the usage of Data Science in Non technical Environment

In this talk, we present some strategies and methodologies we developed at Warner Bros to implement a data-driven culture. We explain how we got the cinema experts to appropriate themselves new tools such as predictive modelling, data-driven decision tools and the problems we encountered along the way.

MRI image analysis

Deep learning has become the predominant choice for medical image analysis tasks. The methods have demonstrated high clinical value for various imaging modalities and diagnostic tasks, such as image segmentation and disease diagnosis and prognosis. In this talk, I will present a few projects from DeepMind Health Research that apply deep learning in clinical domains, including mammography, ophthalmology, and head and neck cancer. Furthermore, I will shortly present the methods and findings of my dissertation work on the breast cancer diagnosis based on dynamic MRIs.

Data Science for Improved Stroke Care

Acute Ischemic Strokes are the leading cause of acquired disabilities in adults in the developed world and the second cause of death overall. Although good therapies and drugs are available, diagnostic time and time-to-treatment are essential to improve the chances of full recovery in a patient. In this talk we introduce SCOPE, an AI and data science engine that has allowed to optimise the time-to-treatment of emergency patients in over 1,500 hospitals as part of the Angels Initiative.

Learning to recommend to help your clients with their paradox of choice

All the individuals are unique thanks to their experiences, tastes, goals and many other factors. Personalization has been the key differentiator for client advisors and sales people since ages by making their clients feel understood, empowered, and known. Today, considering the vast amount of items and digital footprint of users, powerful techniques of recommender systems generate personalized offers to retain customer satisfaction in the digital world. This talk will give practical ideas on how to build recommender systems for a company which does not have it integrated yet in their product offering. It will focus on an example use case for a real-life recommender system in financial services, emphasize the requirements and challenges, and discuss how they can be evaluated.

Engineering challenges in Data Science: It’s a Team Sport

While the core competence of Data Scientists used to be applying sophisticated statistical methods to one-shot data sets in order to produce analytical insights, more and more companies require data scientists to also participate in the development and deployment of data products which leverage Machine Learning algorithms. This requires a shift from a one-woman show to a team sport as well as a shift from pure analytics to engineering. In this talk, I will highlight the skills needed, both interpersonal as well as technical, to fulfill this challenging role successfully and contribute to a real-world impact of data science products.

The role of innovative data collection methods in advancing understanding, and the importance of considering the biases within

Making use of openly available online, often crowdsourced data allow for new insights into people’s experiences and subjective perceptions. This can inform business and policy decisions, or simply further research and understanding into various domains, for example perception of place, or experiences with crime victimisation. For example, people reporting environmental antisocial behaviour using civic technology platforms can give insight into what people consider ‘signal disorders’ in their environment, and other platforms such as safecity.in allow victims of crime to share their experiences anonymously, removing many barriers that prevent official reporting of sensitive crimes. Such data can complement traditional sources and offer another angle to identify spatial and temporal correlates of crimes and fear of crime. However there are various sampling and other issues associated with such data which may affect the conclusions that we can and cannot draw from them. It is important to highlight these issues, as well as explore ways to tackle them, to make the best use of such forms of data in our work.

An invitation to thinking about Machine Learning

Machine learning technologies are being used in various every-day life systems with a growing influence on science, industry and society. This makes it increasingly important for users and society to understand when and whether we can put trust in their predictions. In this talk I will outline which role machine learning theory has played traditionally in providing correctness guarantees for learning methods. However, the immense success of deep learning methods requires new tools for understanding their reliability and brought new challenges. I will argue why many of these new challenges, such as privacy or fairness issues, inherently require theoretical treatment.