Stephen Thaler has created a "creativity machine" that can do things like compose songs, make up new words, invent new toothbrushes, design warheads and power a robot army. Neat, huh? He's founded a company called Imagination Engines to monetize the technology. The technology has been called Artificial Intelligence's "best bet" by NASA.
The technology works by using a neural network in a unique way. Neural networks have been around a long time. They are basically a connected set of nodes that perform a simple math calculation. Inputs feed into each node from other nodes or the world outside the neural net, and computed outputs are produced for other nodes or as outputs to the outside world. Each node weights the inputs it is receiving. The network is trained by feeding a set of data with inputs and expected outputs. The training (setting the weights and other parameters) is where the hard part is, but once done, the networks will perform some fairly amazing tasks, like face recognition. Feed the network a previously unseen set of inputs, and it will dutifully compute an output.
Thaler has managed to take that one significant step further. He takes a trained network that is say trained to recognize songs, then "takes it to the brink of death" by adjusting the weights and starving the network of inputs. What the network then produces as output is a "creative" piece that is an amalgamation of what it has learned before. It doesn't do this randomly, however. The network knows from it's training what notes sound good together, so it tends to maintain that, even as it is recombining notes in previously unheard ways.
Thaler's company has a product that provides prediction of the next half hour of stock price movement based purely on technical indicators. I can't tell if this software is one of his creative neural nets or just a plain neural net. It doesn't seem like the creative aspect would be of much use in the pure prediction game. There are other neural nets that attempt to predict stock movement, include a free online one. In essence, any investor is a predictor, be it 30 minutes in advance or 30 years, and your bankroll growth is your score of how well you are doing. CASTrader will be no different. In the final analysis, systems like neural nets, CASTrader, and direct reinforcement learning, and even the human brain are prediction machines.
Thaler's company claims it's brain is "vastly more powerful" than genetic algorithms, but I'm not so sure. Maybe the secret is in the sauce. I'm sticking with the genetic approach for predicting the markets at least. We'll see how it goes. I can always throw in neural networks at a later date.
Comments