What was so interesting about Poggio’s proposal for revised “Levels of Understanding” framework in place of the old Marr & Poggio’s tri-level framework? Why was there a need for revision?

Here’s what I found interesting from Poggio’s proposal:

One

Poggio, argues in this revised proposal, that it is now necessary to both, emphasize on the connections between the different levels as well as extending the range of levels, if we are to move ahead in understanding of the brain. The points he gives to support this argument of his are:

• He mentions that David replaced the behaviour level with the level of computations in work of Marr and Poggio in 1997. However, he points to the fact that in his original work with Reichardt in 1996, it was stressed how insights and theory at higher levels helps performing the experiments on lower levels in the right way and nervous system needs to be studied at all levels simultaneously to explain a behaviour.

• This fact pointed out by him, wasn’t a part of the revised work by David. But, since this seconds the general opinion that experiments and theory should go hand in hand, he believes that bringing the connections between levels into perspective is of great significance.

Two

Poggio believes that besides the computations and representations in the brain, it is also important to investigate how an organism performs learning and then evolves from what they learn. He pushes forward his argument that learning is in fact, at the core of the problem of intelligence and understanding of the brain, and hence, should be included to the list of levels, above computational level. The points he gives to support this argument of his are:

• During my research on Convolutional Neural Networks, I have often wondered, how they have been vastly inspired from the hierarchical structures of visual processing in the visual cortex. However, even though we are achieving state-of-the-art performance in object detection, in several datasets (ex CIFAR), deep neural networks, still remain more of a black box. So, in a way, we have been able to successfully simulate (or rather replicate from humans) vision tasks like object detection in machines (which is a difficult problem since it involves both regression and classification), just by knowing how to learn to do some computations, without any explicit algorithmic understanding. I think this is the point that Poggio reciprocates, in this work, implying that “understanding at the level of learning is sufficiently powerful to solve a problem and therefore perfectly adequate as an explanation all by itself”.

• Poggio stresses that he wishes the capture the understanding of intelligence. With respect to the statement mentioned by me in the previous point, about deep neural nets being essentially a black box, a nice quote by Marr (1975) grabbed my attention: “the primary unresolved issue is what functions you want implemented, and why. In the absence of this knowledge, a neural net theory, unless it is closely tied to the known anatomy and physiology of some part of the brain and makes some unexpected and testable predictions, is of no value.” This implies that, connections and synergy between levels, is important for understanding of the brain.

• Also, as an interesting argument, he questioned how statistical learning models, compare with learning in brains. One challenge that we face now is, that deep learning models, require huge datasets and energy to learn, which is very unlike in the brain.

Three

Poggio also argues that Evolution should be added to the list of levels. The points he gives to support this argument of his are:

• We have gathered so far from the previous points that understanding how an organism learns is important. But it is also important how those learning algorithms evolve in themselves. So, for example, if a learning strategy performs poorly in helping an animal forage for food, it is bound to be evolved over time, for the organism to avoid starving. If this evolution of learning is understood, a machine could learn without knowledge of learning algorithm and domain knowledge.

• Poggio argues that the effectiveness of humans in learning, is also supported by huge prior knowledge, or rather ‘constraints’ in Marr’s words, which are discovered through evolution (for ex, associative or Pavlov conditioning), which makes it possible for the human to evolve intelligence or the ability to learn, and hence evolution should be placed at the top of the levels of understanding.