A recent article (1) further reinforces what autonomous vehicle industry has been doing. Neural net systems with feedforward and feedback control architectures trained by historical data on specific surfaces and conditions. Remnants of 1960s technologies, ably assisted by zero cost computing, have been percolating across the autonomous landscape. This trajectory is problematic for many reasons.
First, a brain trained on historical data selected by a biased human is a disaster waiting to happen. The situation is no better with hand-coded heuristics as demonstrated by recent aircraft failures. What computer and data scientists have to understand first is that their own brains still remain to be vastly superior to code they write running even on a super-computer. Hence, blind attempts at removing the human from complex decision-making processes are likely to fail.
Second, hype and ignorance have propelled AI to the stratosphere without significant practical use cases. AI is a tool and it is not a panacea. AI still fails when it encounters the unexpected. This is important as it indicates conventional computing and Silicon based architectures, albeit great engineering innovations, have nothing to do with "intelligence." We have not advanced AI much from the 80s, when the "oldies," used to call it expert systems. Granted, simulated voices, believable human faces, and incredible jumping robots are great inventions, but unfortunately, these have nothing to do with AI.
And finally, high human resource intensity in model building often leads to costly failures. For practical AI, two important things need to come together - rapid and flexible prototyping with automation and considering AI to be augmenting the human, not replacing her.
(1) http://robotics.sciencemag.org/content/4/28/eaaw1975
First, a brain trained on historical data selected by a biased human is a disaster waiting to happen. The situation is no better with hand-coded heuristics as demonstrated by recent aircraft failures. What computer and data scientists have to understand first is that their own brains still remain to be vastly superior to code they write running even on a super-computer. Hence, blind attempts at removing the human from complex decision-making processes are likely to fail.
Second, hype and ignorance have propelled AI to the stratosphere without significant practical use cases. AI is a tool and it is not a panacea. AI still fails when it encounters the unexpected. This is important as it indicates conventional computing and Silicon based architectures, albeit great engineering innovations, have nothing to do with "intelligence." We have not advanced AI much from the 80s, when the "oldies," used to call it expert systems. Granted, simulated voices, believable human faces, and incredible jumping robots are great inventions, but unfortunately, these have nothing to do with AI.
And finally, high human resource intensity in model building often leads to costly failures. For practical AI, two important things need to come together - rapid and flexible prototyping with automation and considering AI to be augmenting the human, not replacing her.
(1) http://robotics.sciencemag.org/content/4/28/eaaw1975
No comments:
Post a Comment