Chpt 4: What Neural Network Models Can and Cannot Explain about Cognition Flashcards
McCulloch & Pitts thought of neurons as computing elements, or building blocks for networks that can solve cognitive problems, at a time when the basic functioning of neurons had already been described although the membrane mechanisms underlying action potential generation were still poorly understood. What did their claim hold regarding these simulated networks?
McCulloch and Pitts ’ s claim holds that networks of simplified model neurons are capable of universal computation: they can, with appropriate connectivity, solve the same computational problems that an ordinary digital computer can address as well. They can do so in a distributed way, meaning that there is no central neuron or computational unit that acts like the central processor in a personal computer. The computing is done by the action of many interconnected neurons together.
How did McCulloch and Pitts not follow the traditional, behaviourist idea of nervous activity?
They did not follow the traditional, behaviourist idea of nervous activity as just a relay between sensory inputs and motor
outputs, but as a way to compute ideas . They allocated these ideas not to some abstract, Platonic realm or Cartesian mind space but stated that ideas were immanent in nervous activity, implying they are inseparable from it.
How did MC&P design their model?
McCulloch and Pitts designed a model based on the physiology of the neuron as known at that time. Chapter 2 presented a scheme of neurons with dendrites as elements that receive synaptic inputs, somata as elements converting inputs into all-or-none output (spikes), and axons and their terminals as means to relay the cell ’ s output to target neurons. Crucially, McCulloch and Pitts ignored a wealth of details, such as the complex spatial geometry of dendrites, and produced a reduced version of the neuron that would capture its basic functional operation
Describe a MC&P neuron scheme
In this model, the neuron is reduced to its computational core elements, comprising a dendritic compartment that summates ( Σ) incoming synaptic inputs from presynaptic neurons 1, 2, and 3 and a somatic compartment that performs a threshold operation ( ϴi ) and emits spike trains as outputs to target neurons. The subscript i denotes the identity of the postsynaptic neuron, which is usually part of a larger array in a network. The synaptic weight from presynaptic cell 1 to the postsynaptic neuron is indicated by w i,1.
Describe this modelled neuron in terms of the biology it mimics
The dendrites are anatomically continuous with the soma, giving rise to an axon connecting to target neurons (not shown). Assuming the synaptic inputs are excitatory and the cell is at rest (i.e., non-spiking) when inputs are absent, the activation of a single synapse will depolarise the neuron, mathematically captured by a positive-valued input. The amount of depolarisation depends on two factors. The first is the strength of the synapse, which is a function of the quantity of transmitter molecules released, the number of postsynaptic receptor channels, the extent to which the channels open up, and so on. The second factor is the intensity of the presynaptic cell ’ s activity. If the synapse is very potent but the presynaptic cell is inactive, the net impact on the receiving neuron is zero. The impact of the presynaptic cell on its target cell can thus be captured by the product of presynaptic activity ( a j ) and synaptic strength (or weight, w ij ), where the subscript “ j ” denotes one of the complete array of presynaptic cells and “ i ” denotes the postsynaptic cell:
presynaptic cell’s effect = ajwij
where the Σ sign indicates that all of the presynaptic inputs are summed to compute the total depolarising effect on the target cell.
