Shifting complexity from modeling to interpreting results

A typical AI problem involves quite a bit of tinkering before the model can start learning. Say you want to recognize objects; you need a fairly large number of pictures depicting the objects and associated labels (the names of the objects).
The pictures must be curated, to make sure that they are the same size, the objects stand out, etc. In this example, you would like the model to learn the association between an object and a label, so that it can tell which object appears in a picture it has not seen before. You might consider that the system has learned when it is able to recognize an object in a new picture 90% of the times.

I was intrigued by this paper End to End Learning for Self-Driving Cars because of the "end to end" approach. The system is continuously fed the frame of the road and the steering wheel turn associated with the frame, so that it is able to learn this association. That's it! Actually, it is a bit more complicated than that, it involves tinkering as usual, but that`s the gist of it.

The system learns from raw data, not "from explicit decomposition of the problem, such as lane marking detection, path planning, and control". This means that the system is not built to recognize "human-selected intermediate criteria, e.g., lane detection", but just learns what it needs, from its perspective, to be able to perform the association successfully. Also this means that the network can be smaller because it is optimized to learn only what it needs, not what engineers think it might need.

This is great, except that now we would like to know what the system learned by looking inside the network, for example, but it is a bit hard to draw any conclusions from that. We cut off a lot of complexity involved in devising a learning system, but it is fair to say that we now have to devise a method to understand what the system learned.

Also, the system learns depending on the type of training. This adds another layer of complexity when one tries to understand the rusults.

Biology can provide an inspiration. When you teach a child how to perform a certain task, you cannot just look inside his brain to make sure everything is right, you ask questions, you let him perform the task in different contexts.

Update: interesting critique of End to End learning: Prof. Amnon Shashua talk.

marco palombi

marco palombi

I am intrigued by stuff that does not have a codified description, yet. As a corollary, my field is not definable, otherwise I would not be interested.

Read More