The modern computational theory of cognition began after Alan Turing (1936) published his mathematical theory of computation in terms of what are now known as Turing machines. Contrary to a popular misconception, however, it wasn’t Turing who turned his machines into a model of cognition. That step was taken by Warren McCulloch and Walter Pitts in 1943.
McCulloch and Pitts reasoned that the recently discovered all-or-none law of nervous activity, such that neurons fire only when excitation reaches a certain threshold and does not fire at all otherwise, makes neurons into digital (on-off) devices. With a few other assumptions in place, that made networks of neurons into digital computing devices. McCulloch and Pitts argued, roughly, that neural networks are Turing machines.
(Contrary to another popular misconception, McCulloch and Pitts did not thereby propose the first mathematical theory of neural networks. The mathematical theory of neural networks was started by Nicolas Rashevsky and his biophysics group, of which Walter Pitts was a member. McCulloch and Pitts’s innovation was to propose a kind of digital neural network, and to connect that to Turing machines and computation. Technically, McCulloch and Pitt networks are finite state automata, or Turing machines without tapes; this does not affect our discussion.)
McCulloch and Pitts’s idea had enormous ramifications. It contributed to the origin of computational psychology (computational models of cognition), connectionism (artificial neural network models of cognition), contemporary computational neuroscience (models of neural computation), artificial intelligence, and more.
McCulloch and Pitts’s idea also led to overreach and misconceptions. It gave people unwarranted confidence that they understood the basic computational structure of nervous systems (in terms of digital computation). It contributed to the mistaken idea that the computational theory of cognition could be justified a priori. It was used to support the autonomy of computational psychology without adequate justification. It was even used to defend pancomputationalism, the idea that every physical system is computational.
I argue that, once we rule out these misconceptions, the accurate kernel that remains is that multilevel neurocognitive mechanisms have the function to manipulate multiply realizable variables in accordance with rules. I argue on independent grounds that physical computation is the functional manipulation of multiply realizable variables. (That’s not exactly how I used to put it, but the underlying idea is the same.) Therefore, neurocognitive activity is a kind of computation. This is the basis of mainstream cognitive neuroscience to this day.
It remains to be seen what type of computation neurocognitive systems engage in. Is it digital, analog, or sui generis? Does it manipulate representations? I’ll discuss these questions in my next post.