AI and Machine Learning their impact on research

AI and Machine Learning – their impact on research and knowledge Bernard Butler, WIT/TSSG Presentation at LIR Conference, Cork, 20191127

EMERGING NETWORKS LABORATORY ENL RESEARCH UNIT (WITHIN TSSG) ENL drives scientific excellence through postgraduate and postdoctoral training in key thematic research areas, in collaboration with nationally and internationally recognized academic and industrial partners. DR BERNARD BUTLER bbutler@tssg. org +353 (0)51 845695 ENL Unit Manager (Acting) www. tssg. org/about/people/dr-bernard-butler/ www. tssg. org/research/unit/enl 2

Who am I? • 20+ years research experience (lab, academia) • 7 years consultant (BI, data science, …) • Ongoing interest in – Cloud and fog computing – Identity and Access Management – ML and knowledge representation Dr Bernard Butler WIT West Campus Carriganore, Co. Waterford X 91 P 20 H Ireland bbutler@tssg. org +353 51 84 5695 3

http: //theconversation. com/understanding-the-four-types-of-ai-from-reactive-robots-to-self-aware-beings-67616 1. Purely Reactive 2. Limited Memory 3. Theory of Mind 4. Self Aware https: //futurism. com/images/types-of-ai-from-reactive-to-self-aware-infographic Four Levels of AI

Currently Realisable • Level 1 – Purely Reactive – Acts directly on sensed data with preprogrammed responses – No memory for experiences – No concept of wider world – Specialises in 1 area – Examples • IBM’s Deep Blue for Chess • Google’s Alpha. Go for Go • Level 2 – Limited Memory – Uses past info to update its programming, so less brittle – Limited memory for experiences – No concept of wider world – Specialises in 1 area – Examples • Self-driving vehicles • Chatbots, Alexa, siri, …

Futuristic, Sci-fi only at present • Level 3 –Theory of Mind – Understand thoughts, motives, intentions affecting human behaviour – Able to interact socially (pass the Turing test? ) – Examples: • Level 4 – Self-aware – – – Aware of their own states Can predict feelings of others Can infer from abstractions Creative; Sentient Examples:

Growing scope of AI Narrow (Weak) General (strong) Super (Singularity) • Do one thing, do it well • Current systems • • • Flexible • Human intelligence but better • Existential threat? Flexible Multi-functional Multi-domain Not there yet!

• Routine Cognitive Skills – routine tasks – rule-based operations – jobs where data is plentiful Nonroutine Cognitive Skills • Routine Manual Skills Nonroutine Manual Skills AI (and the automation it enables) is ideal for AI (hence automation) struggles with – creativity – variety • What about librarians and researchers? ? https: //www. businessnewsdaily. com/15273 -jobs-not-replaceable-by-robots. html Effect of AI and Automation on Jobs

AI is not the whole story • • Machine Learning is what we have now Foundation of – Purely Reactive and Limited Memory scope AI – Artificial Narrow Intelligence

Machine Learning has applications everywhere • Financial • – Fraud detection and AML – Identify Investment opportunities • Health care – Wearable devices – Epidemiology and Drug discovery • Government – Evidence-based policymaking Retail – Personalisation – Cross- and Up-selling – Supply planning • Industry 4. 0 – Process control – Automation and Robotics • Transportation – Route planning and logistics – On demand scheduling

Symbolists • Learning is the inverse of deduction, draw upon ideas in philosophy (ontologies), psychology and logic Connectionists • Reverse engineer the human brain, mathematical model is inspired by neuroscience and physics Evolutionists • Nature is the guide: simulate evolution on computer and draw on genetics and evolutionary biology Bayesians Analogizers • Learning is a form of probabilistic inference, with prior beliefs updated by evidence in the form of data • Extrapolate from similarity judgments and are influenced by psychology and mathematical optimization he Master Algorithm, Domingos, Pedro. Sept 2015. isbn 978046506570 -7 5 Tribes of Machine Learning

Machine Learning Algorithms

Myths of Machine Learning (various) • • • Just about summarizing data Just about correlations Cannot predict “black swans” More data -> spurious predictions Ignores pre-existing knowledge ML is complex and incomprehensible to humans Simple models are more accurate/correct Patterns with strong support represent “truth” Phenomena can always be modelled given enough data • • • First summarize, then predict Can derive rules and support Models can generalize More data enables richer models Many models update this knowledge True for many DL models, but not other types Only if they do not underfit the data; use validation checks No guarantee that they are the only such pattern Can derive a model with good training performance but poor test results

Early enthusiasm for Gartner’s “Citizen Data Scientist” proved misplaced…. • Data Science requires expertise • Automated ML projects like this and Google’s Cloud Auto. ML hold promise • Trade AI + computational resources for ML expert effort • Inefficient and overhyped for now

• • • Intentional fakery: Deep. Fakes (audio, video); BERT Brittleness: tweak input, get different result (adversarial) Generative systems: use DL to fool DL Lack of memory to learn across experiences Learn from less data – get better at extrapolating Nature 574, 163 -166 (2019) doi: 10. 1038/d 41586 -019 -03013 -5 Deep Trouble for Deep Learning

Knowledge Generation: Research • • Research is about designing and answering Research Questions 1. Basic Aid - web search etc. – Many research methods to answer them – Qualitative vs. Quantitative – ML consistent with Positivism – Mainly inductive, sometimes abductive 2. Where can Machine Learning help? 3. Data Preparation - Tagging - Finding relationships, e. g. , provenance, authorship, … - Visualisation (dimensionality reduction); enhanced stats Direct Prediction - Survival on the Titanic - House Prices 4. Intrinsic and Fundamental - Discovery of Higgs boson

• • Books, papers, tutorials, courses, competitions, … Open data and open source software, such as scipy and keras The machine learning firehose: arxiv. org (60 papers uploaded on 5 December!) Machine learning itself is an extremely active research topic – Concerns about validation and interpretability (doi: 10. 1145/3316774) https: //www. asimovinstitute. org/neural-network-zoo/ Machine learning resources

Conclusions • Machine learning has been at the top of the hype curve for several years • Libraries and researchers gain productivity • Whole areas of research depend completely on ML • Methodological concerns (ethics, bias, …) need to be addressed

Dr Bernard Butler Thank you WIT West Campus Carriganore, Co. Waterford X 91 P 20 H Ireland bbutler@tssg. org +353 51 84 5695