What Humans Need to Learn

John Langford wants to create a machine learning algorithm to solve problems at least as complicated as anything that a human can do. Today he explained the six characteristics of such a system that would be essential:

• Access to large data sets. This is clearly necessary because when you deprive a human (or a cat!) of sensory inputs he won't develop properly.
  
• Prediction making and immediate feedback on the efficacy of these predictions. Also known as "online" learning.
  
• An exchange between the learner and the verifier, also known as interactive learning. For humans, the verifier is reality and it is always trusted.
• A system that can be broken down into components and can be integrated back into the whole. One might imagine a learning system based on the evolutionary approach, but one basic research goal is a faster and more efficient design than randomness.
  
• A large input context that includes tons of information bits and allows for multiple ways to reach the same conclusion.
  
• Non-linear input representations can and often must be used.

Machine learning algorithms to automate the reconstruction of neural connectivity matrices following serial section transmission electron microscopy would be a great leap for neuroscience. Currently it would be technically possible for a human to do but a whole brain reconstruction would take 90,714,400 work hours, given 40 hours per mm (as given here) and an average brain width of 140 mm, length of 167 mm, and height of 93 mm. The only connectivity matrix that has currently been mapped is that of C. elegans, which took one intrepid neuroscientist 15 years to map, despite the fact that the worm contains only 302 neurons! The point I am trying to make is that a "reasonable" ML algorithm could change the world in at least one concrete way, so keep up the good fight.