Artificial retinas see well enough to balance a pencil
Sunday, 23 January 2011
A team of researchers has built a neural information system that is good enough and fast enough to balance a pencil in real time. If you think it's an easy task, try it!
Building a machine to balance a pencil on its point used to be a high-level research topic - if you aren't into AI I'd better add that this isn't a joke! Think for a moment about the task. You have to move the point of balance around so as to bring the point back under the center of gravity of the pencil. The problem is also dynamic because you have to take account of draughts and vibrations that disturb the equilibrium. As a classical control problem it is also difficult because the equilibrium is unstable.
In most cases the task of balancing a small object such as a pencil is too difficult even for a human and the problem is scaled up to a larger pole balancing task. The response time needed goes down as the pole gets bigger
I've tackled the pole balancing problem a number of times from different points of view - fuzzy set theory, reinforcement learning and so on but never using an AI vision system for the feedback sensor. A team at the Institute of Neuroinformatics, ETH / University Zurich have used what look like video cameras to do the job but in fact they are analog silicon retinas. They work so fast that even with fairly basic hardware they can balance a pencil.
See it in action:
The whole performance is remarkable but all the more so when you know that it is all biologically inspired. It isn't simply using the position of the pencil outline against a background and it does the job better than a human can.
A new course on TensorFlow starts this month and sets out to introduce you to the state-of-the-art approach to building artificial intelligence algorithms. However, as it is on Kadenze, you can e [ ... ]