Consciousness and Computer Vision

Consciousness and Computer Vision: "

While recently abroad for two weeks in lovely Busan, South Korea, I worked my way through John Searls’s “Mind, Language and Society”. In this very accessible piece of work, the author provides a synthesis of his views regarding the lofty topics listed in his title. From a computer vision point of view his thoughts on consciousness may turn out to be particularly relevant. The book starts by asserting that the author espouses a rather Victorian view of philosophy, which insists that there is in fact a physical universe out there that persists regardless of what we do or do not think of it. This then corners us in to a declaration of whether consciousness is part of the supernatural or the material world. Since big bad Richard Dawkins has pretty much sealed the coffin on the former, we must accept the latter as our starting point for this discussion.

The author uses a number of interesting arguments to tease out what is and what is not consciousness. Most alarming is his conclusion that this phenomenon may not yield to reductionism, which is the mainstay of the modern scientific approach to things. Put another way, it may be the case that we simply cannot reduce consciousness into a set of constituent parts that are more easily understood. His arguments revolve around a number qualities than can be associated with human consciousness including its intrinsic subjective nature.

When presented with a black box variant of a computer vision system, I am always interested in how it fails, because this often provides me with potential insights into how it works. For example, a face model-fitting algorithm, which attempts to locate facial landmarks such as the eye’s nose and mouth, are often based on some sort of optimization process. The problem with this type of algorithm is that the search process can become trapped in what is known as a local maximum. This would be similar to you looking at your mother’s face and confusing her lips for her nose. By observing these local-maxima failure modes, one can come up with a pretty good guess as to what kind of mechanisms are being brought to bear. Recently, my colleague Jilin Tu (no genetic relation to yours truly) put together a new face-modeling algorithm based on linear programming. By approximating the face-modeling problem in linear formulation, the globally optimal solution can be found is a systematic fashion. While the algorithm can still fail, these failures are usually due to approximation errors. It turns out that such failures are not particularly informative. In a creepy way one starts to feel that the quest for human-like computer vision capabilities may hinge on achieving a certain level of inscrutability. As Marvin Minsky once commented, the veneer of intelligence seems to disappear once you know how its done…

"