Social robots are helpful companions, designed to interact and collaborate with people in an interpersonal way to help both succeed. As people of all ages and stages begin to live with AI in the home, we have the opportunity to develop personified technologies that promote human growth and flourishing. To support a diversity people as an empathetic and capable ally, smart technologies need to be socially and emotionally intelligent, too. Social robots are a particularly powerful example over other technologies given their validated ability to sustain engagement and support a holistic human experience. This gives rise to important challenges and opportunities for how to design, deploy and evaluate the impact of social robots in the real world to help people learn, thrive and connect. This requires a tightly interwoven approach at the intersection of design/aesthetics, understanding human social psychology, and innovations in AI and machine learning. The goal is to create successful real-world social robots that are enduring in their appeal and engagement, are helpful and competent, can personalize, and can build an emotionally uplifting relationship that provides meaningful value to people’s lives in an ethical way.

The idea of drawing inspiration from the brain to create artificial intelligence has a long history, and has become popular again with the resurgence of neural nets. For the past dozen years, my colleagues and I have been working in the opposite direction: applying artificial neural nets to solve the structure of the brain. Today my lab is applying convolutional nets to reconstruct the wiring diagram of the visual cortex from electron microscopic images. We are returning full circle, because convolutional nets were originally inspired by 1960s-era neuroscience discoveries about the visual cortex. We hope that wiring diagrams of the cortex, combined with data about neural activity, will give us insights into how the cortex learns. I will show some preliminary findings in this direction, and discuss their relation to theories of cortex-inspired unsupervised learning based on Hebbian and anti-Hebbian plasticity.

It is curious how machines can learn to translate from one language to another, while still having major difficulties in performing tasks humans complete with ease, such as document summarization and conversation. Equally curious is the fact that language models, trained only from natural language text, often generate such unnatural text that humans would rule implausible instantly.

In this talk I will discuss the missing representations in neural language models and how we might lift their limitations. First, I will argue that the current paradigm of “learning by reading” is fundamentally limiting for long-form text generation and introduce a new framework of “learning by writing” where machines can learn through practice. Second, I will discuss how commonsense knowledge and reasoning are critical for enabling machines to read between the lines, and how we might be able to model this latent commonsense through language.

For natural language understanding (NLU) technology to be maximally useful, it must be general: it must be able to process language in a way that is not exclusively tailored to any one specific task or dataset. In pursuit of this objective, this talk presents the General Language Understanding Evaluation benchmark (GLUE), a tool for evaluating and analyzing the performance of models across a diverse range of existing NLU tasks. GLUE is model-agnostic, but it incentivizes sharing knowledge across tasks because certain tasks have very limited training data. This talk discusses results from GLUE’s first few months, with a focus on the successes and limitations of representations learned through the task of language modeling.

Recommender systems play an important role in our online world. They influence the books we read, the movies we watch, and the music we listen to. Due to the ever increasingly connected world of the Internet of Things, and the maturation of AI technologies, their reach is now beginning to extend from the online world into the real-world in which we live. The next generation of recommender systems will look beyond our clicks and purchases to learn about our opinions and habits and serve us in a host of new application areas. In this session, Professor Smyth will look at how novel sources of recommendation knowledge are impacting recommender systems.
In particular, he will focus on two case studies:
1. The use of user-generated reviews in product recommendation.
2. The use of activity data from wearables in exercise recommendation.

As humans and robots increasingly co-exist in home, industrial and healthcare environments, it is crucial that robots and other smart devices incorporate into their decision making efficient learning algorithms for inferring and adapting to the user’s sensorimotor and cognitive profiles. In this talk, I will describe our learning architectures for enabling adaptive human robot interaction in joint tasks, and discuss the related computational problems, including user modelling, attention, perspective taking, prediction of forthcoming states, explanation generation, and personalised shared autonomy. I will give some examples from human robot collaboration experiments in robotic wheelchairs for disabled kids and adults, collaboration in edutainment tasks, activities of daily living (for example assisted dressing tasks), shared control for handheld robots, shared autonomy in driving, among others.

By integrating emotional intelligence with the existing artificial intelligence, AI is taking a crucial turn on its journey to becoming a transformational technology. However, even the most sophisticated AI technologies nowadays lack this essential factor and the ability to contextualize information as humans do. This is arguably the most important reason why AI is not able to adapt successfully to individual users. Companies able to effectively incorporate contextual understanding into their technologies will become the front-runners in this race to technological excellence. However, they must also be cautious of the social and personal implications of such a huge transition. This talk will describe ‘What is artificial emotional intelligence?’, how the integration of it with AI can be beneficial, and what the future of artificial emotional intelligence looks like.

Machine learning, and in particular deep learning, has received a great deal of attention in recent years as it has been able to solve notoriously difficult problems. Much of deep learning’s recent successes can be attributed to a virtuous cycle of advances in computing hardware, the availability of huge amounts of data, and the explosion of sophisticated models and algorithms. Efficient, on-device acceleration of AI workloads demand synergistic co-design of machine learning algorithms with the specialized hardware they run on. This talk provides an overview of our on-going research activities that take a full-stack approach from the algorithms all the way down to the circuitry in order to propose holistic solutions.