Nobody really has a good theory of how brains work. Yet we keep hearing people saying that artificial neural networks are not biologically plausible. Who anointed these folks to be the thought police of what is plausible?
A mistake that too many make about artificial neural networks is that they are implementation models (see: Marr's level of explanation). They are not! They are algorithmic models. When you realize this, the question of biological plausibility should be thrown out.
I think Pylyshyn's mapping of the semantic, syntactic and physical to Marr's computational, algorithmic and implementation is just wrong.
Richard Sutton has one of the clearest explanations of Marr's computational level. 
A computational theory must be explicit about the goals, the what's and why's of intelligence. It is surprising that very few researchers can explicitly express the computational theory of their research.
One can map different computational theories in two dimensions. One dimension is the functionalist versus enactivist. Another dimension is whether the theory is inspired by ecological or formal ideas.
A majority of research in Neuroscience happens at the Marr's implementation level. There are clearly severe limits in what can be attained at this level. So severe that Christof Koch declared:
The proper approach to understanding intelligence is by an explicit declaration of one's computational model, followed by an algorithmic model that can demonstrate intelligent behavior.
Neuroscience research is still useful, but one has to realize that it is *not* research about intelligence. It is research about how biological brains function. But it will take at least a century for a bottom-up approach like this to explain intelligence.
It's unfortunate that my conclusion may disappoint a lot of neuroscience researchers. Unfortunately, there are limits to how emergent complex phenomena can be derived from the details of neuro-circuitry.
To make a common analogy, no amount of reductionist research studying a tree will tell you the dynamics of the evolution of a forest. It's just at the wrong level of explanation.
Where I believe we can have a meaningful conversation about the nature of intelligence is in the comparison of different computational models. The resolution of which of these models is correct can only be determined through the implementation and execution of algorithmic models.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
