HALICON 2017

Healthcare costs are predicted to become unsustainable, and yet we as a human society aspire to have perfect healthcare for everyone. Can artificial intelligence come to the rescue? Can computer systems make accurate predictions of human health, and suggest preventative and therapeutic measures to manage the burden of disease in a reliable and cheap manner? How would an artificially intelligent system that can make meaningful decisions, like say the personalized health plan for an individual, be designed? Do the internet of things and machine learning techniques have any role in creating this AI? 

The world today has machine learning systems and artificially intelligent systems taking over the burden of making some critical decisions. Self-driving cars; algorithms that automatically classify digital text information; automatically categorize and tag photographs; identify fraud in banking software; invest financial assets automatically; etc. are all examples of somewhat intelligent systems aiding humans in making decisions that would otherwise take manual examination of a large amount of data. But can algorithms be trained to see even more complex patterns in data, that typically only humans have been able to? Can we train algorithms that take on real risk, like prescribing a treatment plan to a person? 

The most cutting-edge work in AI today is happening in a field called representation learning. The genesis of this sub-field of AI and machine learning work, has been inspired from cognitive neuroscience, and how the human brain manages to make many complex decisions quite effortlessly, and with only about 20 watts of power. A datacenter running machine learning algorithms consumes several kilowatts of power. 

This talk will give a background in representation learning. Why learning representations of raw data could boost the performance of traditional neural networks in accomplishing complex pattern-matching and decision-making tasks. We will talk about how such representations could be generated from raw-data streams and how they would enable semi-supervised and reinforcement-learning (learning from experience), in neural networks. Finally, we will look at how such techniques could be applied to processing data common in the healthcare industry and what types of outcomes could we hope to achieve.