When I was a little girl I tried to watch every episode of the American television series Knight Rider, not because of the male main character who fought injustice and crime, but because of Dr. Bonnie Barstow, the female chief technician of KITT. KITT was an artificially intelligent self-driving car that could understand and communicate in natural language, and one could remote-connect to it via a wrist watch.
Which 10 year old’s heart wouldn’t start beating faster by this? I wanted to be like Dr. Bonnie Barstow when I grew up, building technology and machines that were so intelligent, that humans could talk to them!
Eventually, I started my university education and focused on natural language processing or how to make machines understand and communicate in natural language. Back then (about 20 years ago) we studied linguistic theories in order to understand the underlying principles of how humans process natural language, and then developed so-called finite state methods (e.g. formal grammars, graphs and automata) that could be implemented in machines.
Since recently, everyone in natural language processing (or in any other research area of your choice) seem to use machine learning methods to solve problems: given a problem statement x and a machine learning algorithm y, apply y to x, and obtain very good and fast results. The only tiny remaining question is: why do we obtain these results?
Do machines today speak and reason like humans do due to the omnipresent machine learning methodology? No, they don’t. Did we make progress? Yes and no. We obtain faster and better results for specific problems, but those who want to understand some principles of human cognition, do not gain the scientific insight they seek.
Currently I’m working with colleagues at the Department of Computing Science, and three Pepper robots on developing dialogue management approaches that combine machine learning methods and finite-state methods. Machine learning algorithms serve us well for problems that involve mining patterns or correlations over large data sets, and finite-state methods give us the understanding that we, as scientists, naturally seek. I believe that the key to a new advancement in AI is combining methodological approaches, so that we get reliable and fast results, combined with insight into stated problems and their solutions.
I want to share an advice by Marvin Minsky (one of the pioneers of AI) who said ”If you just have a single problem to solve, then fine, go ahead and use a neural network. But if you want to do science and understand how to choose architectures, or how to go to a new problem, you have to understand what different architectures can and cannot do”.
Over the past years, I obtained my PhD, tried out various methodological approaches, worked with Pepper robots, and I can say, that I am Dr. Bonnie Barstow but my Peppers are not KITT. Not yet.