Measurement theory studies the concepts of measurement and scale. I suggest that in order to make further progress we need to develop a proper measurement theory of machine learning. This is clearly problematic, since if machine learning researchers are unclear about what exactly their experiments are telling them about their machine learning algorithms, then how can end-users trust systems deploying those algorithms? However, there is a range of areas where this understanding is still lacking, leading to ill-advised practices in classifier evaluation. Our understanding of performance evaluation measures for machine-learned classifiers has improved considerably over the last decades. Peter Flach, Professor of Artificial Intelligence at the University of Bristol The Highs and Lows of Performance Evaluation: Towards a Measurement Theory for Machine Learning Hung concludes the talk by discussing open opportunities regarding new directions of potential research on social experience monitoring and enhancement with respect to topics such as privacy, data labelling, personalisation, and multimodal experience enhancement. Through examples from her prior work, Hung demonstrated that this provides intriguing new opportunities for investigating unconventional approaches to multimodal data processing which opens up a new field re-exploring phenomena from a machine perspective that goes beyond the commonly understood modalities of sight, and hearing. This starts by first reconsidering traditional approaches to measuring human affective experience. In this talk, Hung argued that human experience has an inherently personal component that should be explored if we want to close the loop on enhancing the quality of human social experience. This approach strips away the possibility of measuring interaction quality, pushing the research focus more towards larger scale sociological studies that try to find generalisable patterns of human behaviour and its relation with their affective experience. However, when we step away from more restrictive social settings to cases where people are free to move around as they wish, most research (stemming from the ubiquitous and pervasive computing community ) have tended to apply proxies such as co-location as a measure of social interaction. Fortunately, with the rising popularity of wearable technologies, there is an opportunity to digitize momentary social experiences as they unfold in the real world. The drive for an interdisciplinary approach stems a lot from the idea that computational tools that could have the most impact for enhancing social experience must necessarily be embedded in people's everyday lives. While the text above may sound like the start of a social science presentation, in this talk, Hung argued that in order to enhance the quality of human social experience where it could have the greatest benefit, we need an inherently interdisciplinary approach combining both social science and computer science. Studying how social interactions unfold and how these can affect or enhance social relationships taps into human's instinctive perception of the experience of social interactions. In today's society, one can consider social bonding to be important in relationships with a romantic partner, friends, and family or with professional colleagues. Social bonding is a key component in human collaboration and with it, comes the possibility to achieve more as a group than as an individual. Humans interact with one another on a daily basis. She leads the Perceptive Computing Lab, which is part of the Pattern Recognition and Bioinformatics Group. Hayley Hung, Associate Professor at the Technical University of Delft. Unfortunately there is no video of this colloquium. I will explain the algorithmic improvements, and also show how we have applied these in the health domain. In this talk, I will focus on three algorithmic innovations we have made to improve the applicability of RL in the health domain: (1) sample efficient RL (2) safe RL with domain knowledge, and (3) explainable RL. Reinforcement Learning (RL) is a very natural fit, however it comes with some characteristics that do not fit the medical domain well. Such decision can be supported by AI-driven models. Just think of a doctor continuously modifying the ventilator of an intubated ICU patient or changing the dosage of fluids administered depending on how the patient is doing. In healthcare many of the decisions made are of a sequential nature. Mark Hoogendoorn, Full Professor of Artificial Intelligence, VU, Amsterdam You can leave our website to view this video. This video can not be shown because you did not accept cookies. Reinforcement Learning for Health and Wellbeing
0 Comments
Leave a Reply. |