Monday, September 15, 2014

Redefining AI

Artificial Intelligence (AI), an artifact of the 80s, has been directionless. This is partly due to the overuse of the term and assigning “intelligence” to such common place activities as search, rules based logic and machine learning. Recent news that researchers at North Carolina State University has been able to use “AI” to predict the goals of a player in a video game using machine learning, highlights the idea that the term AI is poorly understood. It may be time to redefine it more precisely so that claims of progress in this area could be tempered.

If AI is about intelligence – human intelligence – then most contemporary attempts at replicating it has failed. If AI is about naive search of large data spaces for patterns or the use of classification, clustering or rules based logic on “big data”, then AI will continue to flourish with no innovation in knowledge or software. In this vein, all AI needs is raw computing power. The current leader, Watson, is a case in point. Packing silicon ever closer together and massively parallel processing set logic channels, is not AI – even though it may be able to find the answer to any vexing question asked in trivial games. Machine learning – the latest fad discovered by business brains, without understanding that it has been happening for many decades – has nothing to do with AI - it is just raw application of mathematics, afforded by cheap memory and cheaper computers. What the AI crowd seems to be missing is that, none of these – ability to create models from data, ability to guess answers to trivial questions, ability to predict goals – is about intelligence. It is about the inevitable marriage of computing power with established mathematics.

Human intelligence, however, is not mathematical, even though every scientist and engineer would like it to be. This is why the preeminent engineering schools in the world – in the East, West and in the middle, cannot make any progress in this area. Soccer playing robots and self driving cars, unfortunately are not intelligent. They are unlikely to imagine string theory or appreciate art. Considering intelligence to be mechanical and mathematical, is the first problem. Lately, it has been shown that the hardware itself, the human brain, is a quantum computer. Feeble attempts at replicating this hardware phenomenon is not going to get humans any further in AI because fundamental issues remain in understanding the operating system and applications that run on it.

Artificial Intelligence is meant to represent complete replication of human intelligence. It is not parroting answers in Jeopardy or predicting behavior based on historical data. Humans may be giving themselves less credit by assuming that the crude machines they build, are in fact intelligent.