from the shallow end to deep graph neural networksjaven/talk/ml01_what_is_machine_learning.… ·...
TRANSCRIPT
Javen Qinfeng Shi Associate Professor, The University of Adelaide (UoA)
Director and Founder, Probabilistic Graphical Model Group, UoA
Director of Advanced Reasoning and Learning, Australian Institute of Machine Learning (AIML), UoA
WHAT IS MACHINE LEARNING? FROM THE SHALLOW END TO
DEEP GRAPH NEURAL NETWORKS
WHAT IS MACHINE LEARNING?
Using data to uncover a unknown underlying process Yaser S. Abu-Mostafa
Gives “computers the ability to learn without being explicitly programmed” Arthur Samuel, 1959
Q:
• Thousands of machine learning algorithms out there. How can we possibly study they all?
• Many algorithms come out every year, how do we keep up with them?
A:
• Learning theory analyses sets of algorithms’ behaviour • Many algorithms can be formulated in a unified
framework called Empirical Risk Minimisation (ERM).
WHAT IF YOU DON’T HAVE ‘BIG’ DATA? • Startups usually do not have ‘Big’ data • Traditional companies/businesses’ data are often not collected
or organised in a way that AI systems want • Without big data, can you use AI? • Yes you can! • Bonus: it will get you more data once start using.
• How do you guarantee performance without big data?
DOMAIN KNOWLEDGE • It works (prior to the era of AI) • Uncertainty (happens in some probability, rarely 100%) • Medical • Industry 4.0, Fish-bone diagram
• Long time to train a person (think of how many years to train someone to be a doctor)
• Extremely long time to accumulate (human has accumulated knowledge in thousands of years)
WHAT CLIENTS WANT? • AI to absorb all human knowledge and experience (even the
wrong ones) • AI can improve as the quantity and quality of the data improve • New AI system can incorporate the previous AI systems (or IT
systems) • With new AI system, predication and decision making should be
better than before • Interpretable
HOW TO ALLOW AI TO ABSORB KNOWLEDGE
• Using knowledge and rules to generate (simulate) data, and use AI system to train on them (Silly way)
• Directly using knowledge and rules to build ‘System of Experts’ (old way, problematic)
• Build a super machine brain that can
• Perform both probabilistic reasoning and logic reasoning
• Use both deep neural networks and graphical models
• Absorb any knowledge and rules
• Can correct wrong knowledge and rules • Can incorporate and inherit all previous systems
• Can discover new knowledge
• Interpretable