Phew, I just submitted my first PhD application, with many more to go… I thought I would share my covering letter as it does a decent job of covering my interests.  

“Back in 2012, I was surprised by Daniel Kahneman’s and Amos Tversky’s results showing cognitive bias and the conclusion that: You are not so smart. I wondered; why are we flawed, is there a reason? how we are flawed, can I fix my flaws? what is the optimal way to think, can machines be better?

Slowly, and painfully, I have managed to narrow (or focus, depending on the time of day) my interests to three manageable topics; the principles of neural design, automated science and artificial general intelligence.
My background is in engineering, mathematics and computer science, and I think they provide THE tools for understanding these topics (though it’s open for discussion). I especially like the tools; understanding via construction, reduction to simple understandable cases, counting how models will scale with various resources, proving necessary and sufficient conditions, fitting models to data.

Principles of neural design: I want to deeply understand the brain and its connection to our thoughts. I especially like the idea of understanding what advantage is provided by a neural structure and the computational and/or physical constraints on that structure. As an example, I would be interested in constructing biologically plausible and computationally inspired models of;

  • language and reasoning (especially w.r.t model-based learning), 
  • executive control and its relationship to addiction,
  • learning and memory systems (spatial, working, episodic, short-long, goals/values, …).

The models above would explain the evolution, development and function of our faculties.

Automated science: I see science as society’s learning algorithm: the faster we can learn, the faster we can solve our problems. But, the problems that science is attempting to solve are getting more complex (as we solve the simpler puzzles). We (/scientists) need better tools to help us deal with the complexity. At the very least, I want these tools, because I don’t feel smart enough to keep up, let alone answer my own questions about intelligence and the universe. Tools that I would like are;

  • NLP that can parse academic papers (the arguments, theories and evidence presented) into structured data. This could allow us (or learned algorithms) to reason about; which evidence supports a theory. And which experiments would be useful given leading theories.
  • NLP that can summarise, communicate and compress explanations for existing research. Organising the essential facts/evidence and presenting theories in an intuitive, human understandable, format.
  • Efficient exploration (and optimisation); of semi-conductors, of anti-cancer drugs, of computers, of batteries, … the unknown!
  • Automated (and efficient) learning of causal and interpretable models from complex systems (that are partially observable and have limits on interventions). Another way of framing this is: better pattern finding tools for; bioinformatics, physics, economics…

Artificial general intelligence: Finally, I love the idea of building the ultimate tool, the smartest thing possible, and making it practical for others to use. At the moment this might translate to simply scaling deep learning to bigger and harder problems. Some key challenges that interest me are;

  • How can computers learn to generalise efficiently? How can computers learn so that their knowledge is easily transferable? (Transfer, lifelong and meta learning)
  • Smarter and more efficient reinforcement learning. Current reinforcement learning algorithms are profoundly dumb, implementing something close to random search with poor memory. How can we make them smarter? For example; creativity, curiosity, one-shot learning, …?
  • What are the trade-offs of designing learners? I want to understand the relationship between; trading accuracy for computational complexity, or trading sample complexity for robustness…  
  • Efficient learning in multi-agent settings. How can agents collaborate to achieve more? How do individual/local goals combine to generate global optima?

I have listed some projects that I can think of, but I have found that most fields of research can be interesting if you approach the problems from the right perspective, so I am open to other(s) ideas. But if I had to concisely explain my interests I might say; I want to understand intelligence from a computational perspective (cognitive science via computer science).”