Thread by @IntuitMachine

Wonderful discussion with Paul Cisek at the Learning Salon. Paul proposes a refactoring of our taxonomy for understanding cognition. He argues that the structure should be driven by studying the history of evolution. crowdcast.io/e/learningsalo…

What I love about the Learning Salon that is hosted by @criticalneuro @neuro_data John Krakauer is that the hosts are all ready to tear apart the arguments of the speaker. Krakauer has an uncanny ability in conjuring up strong cases against the speaker.

I subscribe to Cisek's thesis in that to understand cognition, we should be informed by evolution. Cognition is a consequence of history (or the baggage) that lead us to our present state. Studying this information can lead to explanations of the peculiarities of human thinking.

Krakauer's strong argument against this is that it 'runs out of steam'. That is, it cannot explain the more complex cognitive mechanisms. I suspect the argument may relate 'virtualization' in that cognition is decoupled from substrate.

Functionalists will argue that they can discover the mechanisms of cognition independent of the unique path of evolution. This is related to convergent evolution where different paths lead to the same function. What is the issue with this stance?

No study is more anthropocentric than the study of the mind. Ironically, the mind is predominantly unconscious and below the surface of conscious introspection. Yet we have developed taxonomies that are based on arm-chair explorations.

At every level of biology leads to a different context that is solved in many ways. Our brains are a consequence of that billion-year history of solving problems that a majority of which we are unaware of. The study of the evolution of brains gives us insight on these problems.

In short, it informs us of the curriculum that evolution followed. However, the argument against this is that it does not reveal to us how it was done but rather only identifying the steps in the process.

But we already know from advances in deep learning that understanding why and how may be impenetrable. We predominantly work at the boundaries, timing just the branches and allowing the learning algorithms to grow the solution.

The problem though of learning algorithms that is (1) Goodhart's Law and (2) Deceptive objectives. The first implies that learning algos will seek the easiest solution to a problem and not the most general. 2nd implies that seeking immediate objects closes one out to solutions.

The evolutionary record reveals to us the stepping stone problems that were solved prior to achieving more advanced capabilities. It reveals the curriculum of biology. Evolution makes progress by solving immediate objectives and not long-range ones.

One may argue that studying vacuum tubes does not lead us to understand transistors. But, we should recognize that the invention process of evolution differs from that of humans. Biology is a differentiation process rather than an additive one.

In other words, the connection with the past is informative to how something works in the future. Think of understanding the processes in human bureaucracies. The warped logic can only be understood from the lens of how it evolved.