Current Research

Our brains are so good at processing visual information to build a robust and general-purpose representation of 3D space that early researchers were tricked into thinking it would be somewhat easy to teach a computer to do the same. Fifty years later and we are still trying to get a handle on the right way to approach this problem – it turned out to be far more difficult than initially thought. My opinion is that we are missing something fundamental and that it has to do with finding a formalism for building highly structured representations of scenes that also support some form of statistical inference. The main thrust of my research is in searching for a unification between symbolic computation and neural network models in order to tackle this problem. Vector Symbolic Algebras (also known as High-dimensional computing) suggest a way to make some progress along these lines and my main interest is in using this framework to build flexible and robust scene representations.

My other interest is in signal processing that leverages what we know about real brains. This includes theories of neural coding (in particular, sparse coding) that may be useful in engineering systems heavily constrained by computational, storage, or energy consumption requirements.