From IMS to the AI revolution
By: Aravin Sukumar
We are entering into a new era of technology that has only been depicted in sci-fi movies such as Ex Machina and A.I. (Artificial Intelligence). What is common in these movies are robotic characters that can simulate human behavior and intelligence. As we speak, the world is taking advantage of new advances in programming and computer hardware to create intelligent systems. Adam Santoro is one of these individuals. Adam, a Ph.D. graduate from the Institute of Medical Science (IMS), is a Research Scientist at DeepMind–an artificial intelligence (AI) company focused on creating intelligent systems that can learn to solve some of our world’s most complex problems. Capitalizing on the rise in AI technologies, which have recently been covered in Forbes, The New York Times, and The Economist, Adam followed his passion in the field of AI, where he has succeeded by applying the intangible skills acquired throughout his graduate studies. Evidently, AI has many potential applications for scientific disciplines within the IMS community, especially considering that the University of Toronto (UofT) is a globally-recognized leader in the fields of AI and machine learning.
Follow your passion
Adam completed his Bachelor’s degree at UofT with a double major in Biochemistry & Physiology before pursuing his Ph.D. in Neuroscience in the lab of Dr. Paul Frankland. Outside the lab, he loves being active and regularly goes to the gym, which helps him to clear his mind. “Many people have shower epiphanies; I tend to have gym epiphanies,” Adam mentions. In addition, he enjoys music composition and playing different instruments. In the lab, his research focused on ‘systems memory consolidation’; in other words, how our memories transform over time. His main findings, published in the Journal of Neuroscience, suggest that over time, our memories about specific events transform from being highly detailed to limited to only pertinent information. Adam explains, “If you went to a grocery store, you tend to remember details such as the store layout or employee names; yet, with time, these memories become broader, such as ‘the store sells fruit’.” Adam believes this process of memory transformation “enhances new learning in constantly changing environments; enabling us to forget specific details that are potentially irrelevant, and remembering more general knowledge that is more broadly applicable to the tasks we may face in the future.” It was at the mid-point of his Ph.D. that he became intrigued by the possibility of incorporating computational approaches to his study of the brain. This interest was sparked by Dr. David Marr, a highly influential neuroscientist, who introduced a framework for understanding intelligent systems by examining them from a computational level. This led to deeper theoretical questions: for instance, what qualities define an intelligent system. Adam was motivated to learn more about intelligent systems, and how research on the human brain can shape how we design an intelligent system. “Be a lifelong learner and keep learning and reading about anything that interests you,” Adam advises. After completing his Ph.D. in 2015, Adam began to work for DeepMind, which aims to design intelligent systems that can help us understand human behavior and intelligence.
Learning about deep learning at DeepMind
Based in London, UK, DeepMind was founded in 2010 and later acquired by Google in 2014. Employing over 500 scientists, engineers, and administrators, their overall mission is to create intelligent systems that can solve problems. Adam finds many similarities between working at DeepMind and his experience in graduate school, including the proposing, planning, and execution of experiments, analysis of data, and writing papers. DeepMind has received global attention for creating AlphaGo, the first computer program to defeat a professional player in the classical board game Go; in fact, AlphaGo defeated the game’s world champion, Lee Sedol. This game presented great challenges for AI programs to successfully outplay a human, underlying the magnitude of DeepMind’s achievement in developing the AlphaGo. The key feature of DeepMind’s technology is that their system is not programmed for specific tasks, but can learn from experience: for instance, allowing it to learn and play different games. This process is referred to as reinforcement learning. This process of learning falls under the concept of ‘machine learning,’ which describes the ability of programs to learn from their environment and improve their abilities at performing a task in the future. Each machine learning-based system consists of a neural network capable of converting an input (e.g. pixel values of an image) into an output (e.g. label as cat). This neural network is reminiscent of the human brain, which consists of layers of neurons that sequentially process data and produce a specific output. It is this layer-based machine learning, referred to as deep learning, that enables these artificial neural networks to be powerful problem-solvers by collecting and using data to train themselves into an intelligent system. Adam is currently working on relation reasoning, a key aspect of an intelligent system, which is the ability to make sense of the interactions between objects or entities. Published earlier this year, Adam’s research describes a relation networks algorithm, a specific deep learning algorithm that enables programs to assess the relations between objects. For example, the algorithm may compare the size of an object to that of a collection of other objects.
Notwithstanding the immense potential for AI in diverse applications such as self-driving cars, medical diagnostics, healthcare, and more, there are some who see potential pitfalls. Stephen Hawking proposes that programs superior to the human mind pose risks, from the loss of jobs due to automation to global security. Adam and many others in the field assure us that the best minds in the world are pondering these potential pitfalls. For example, in conjunction with Apple, Google, Amazon, Facebook, IBM, and Microsoft, DeepMind founded Partnership in AI, an organization devoted to ensuring that AI is used to benefit society. Acting in the public’s interest, this organization educates the public on the progress of AI and leads a dialogue on best practices in its implementation.
Across the abyss of scientific disciplines: AI
Transitioning from your current research discipline into another field can be daunting. Adam felt that his transition from academic neuroscience to the computer science industry presented a steep learning curve and required a lot of self-learning, especially in programming. “A good Ph.D. experience involves getting thrown into the deep end and not drowning,” Adam reflects on his experiences in graduate school. He believes that if you are really interested in something, you should pursue it. “Do not consider your Ph.D. as a path to becoming an expert on a random molecule that only three people in the world know about; your Ph.D. is a sandbox for scientific thinking, where you learn how to think about problems and how to answer them.” Adam and many others are convinced that the field of AI and machine learning is gradually becoming embedded into many disciplines of medical research, such as the diagnosis of rare diseases using a facial recognition AI software. Adam believes it is possible for anyone to build and train a neural network with a little bit of programming, linear algebra, and calculus. There is also open-source software, such as Google’s TensorFlow, that allows beginners to explore machine learning. His advice to the IMS community: “Don’t shy away from quantitative methods; machine learning and AI are already influencing analyses in many scientific fields, and their influence is only going to grow stronger.” As more people join this AI movement, we are one step closer to creating a world reminiscent of the sci-fi movies we have all watched in awe.