I was doing my daily blog rundown, and this video caught my eye. Go watch it. It’s pretty cool.
Of course, a recorded demo doesn’t say that much about the actual technology accomplishments, so I went ahead to check out the lab’s web site. Their technology is named SOINN–Self-Organizing Incremental Neural Network, and their page is here. It seems that SOINN is essentially an online variation of the Self-Organizing Map (SOM), using pretty standard Hebbian learning, and also something called Growing Neural Gas (GNG), which I haven’t heard of before. The SOINN algorithm itself is really interesting, but doesn’t seem as revolutionary as HuffPo and Engadget are making it out to be. It’s another online unsupervised learning algorithm.
I’m more interested to see how the real-world knowledge was stored in the robot in the demo. They never mentioned how the robot knew it was handling a cup, and how a cup and water were supposed to be manipulated. Did it undergo reinforcement learning to learn the laws of physics and reality?