Using machine learning in the realm of wireless sensor networks, I have been able to improve sensor node localization and provide event detection. Since I found the concepts of machine learning interesting, my adviser provided me with a book to get an overview: Machine Learning, written by Tom Mitchell of Carnegie Mellon. Most machine learning techniques are equivalent to function approximation and most also require a significant amount of human intervention to work properly. A machine learning model can be trained to take a set of inputs (such as sensor readings) and provide an output (a tank is hauling butt towards my base). However, the model has to be trained to know what the correct outputs are (tank is present or not present), so a human must provide a limited set of training data where the correct output is known.
Given the state of the art in machine learning, there is no way for an AI to learn on its own. There is reinforcement learning, but even in this case a human must decide the conditions and the amount of a reward or penalty for each AI decision or output. This is why AI in games is terrible: when there are a large number of non-deterministic game states and a large number of non-deterministic actions to take, it is almost impossible to determine the correct action to take at every decision point. This means that it is difficult or impossible to provide the AI with labeled ground truth or a reward for training. More to the point, labeling each output with the correct value would be a real headache. Instead, game developers resort to a rule-based system that still has trouble covering every possible scenario. As a result, NPC characters still wind up doing something weird, like running into walls.
I’ve been working with this stuff for awhile, realizing its capabilities and especially its limitations. Then, this weekend I see a headline reading: “Scientists Worry Machines May Outsmart Man.” Overblown media hype at its best. The article concerns a conference on machine learning attended by the aforementioned Tom Mitchell and futurist Ray Kurzweil. From what I gather, the conference dealt with mostly philosophical issues with respect to advancing technology and its integration with everyday life. There wasn’t much about strong AI taking over the world, yet it was plastered all over the news that a Skynet-esque entity would rise from the Internet and doom us all. In some ways, simpler systems have already taken over our lives: GPS tells us where to drive, automated tools read MRI scans and provide diagnoses, and viruses wreak havoc on our personal computers. However, strong AI has quite a ways to go, with most machine learning research peaking decades ago. As one Slashdot commenter on the NYT article writes:
Any computer scientist who is worried about AI taking over no longer deserves to be referred to as a computer scientist. The state of “artifiical intelligence” can be best described as “a pipe dream.”
All of this comes on the heels of a TED talk on the development of a brain simulator. The speaker indicates his current brain “implementation” is running on a 10,000 core Blue Gene system. The article gives few details, but it sounds like a large scale artificial neural network, which still needs supervised training data to learn properly. 10,000 nodes is still way too small, since the average human brain has 100 billion neurons with 7,000 connections each. Maybe in ten years the requisite computing horsepower will be in place, but I’m guessing the algorithms and the intelligence will not.
In a similar light, a team of scientists recently used DNA computing to solve the NP-Complete Hamiltonian Path problem. Instead of using some artificial construct or model, billions of DNA sequences, each representing a possible path, were randomly constructed such that those having a correct solution would glow a different color. While massive parallelism makes this a relatively fast solution to an NP-Complete problem, this approach really isn’t a doomsday AI either.
While specific solutions continue to be discovered for our technological problems, development of strong AI (and the development of Skynet) will sit on the back burner. Until then (and it’ll be awhile), everyone can take off their tinfoil hats.
Recent Comments