How is the biological accuracy of ANNs typically measured?

Question

I am referring to the computational neuroscience side of neural network research that focuses on biological accuracy. I've read references to improving biological realism (using say spiking neurons or binding neurons of various types) by emulating known properties of real neurons (action potentials and what not). However, I haven't been able to find any detail on how the effectiveness of such attempts is measured empirically.

For example, cultured neuronal networks (such as a neurochip) could be used as a benchmark: Create an ANN and a BNN with the same number and arrangement of neurons, give them the same inputs, and compare the results... Has anything like that been attempted? How is the biological accuracy of ANNs typically measured?

Answer 1

One way the biological plausibility of an artificial neural network could be assessed is to look at how much a neural network abstracts away from the behavior of real neurons.

For instance, it is common in psychology and machine learning to use a sigmoidal activation function to determine the output of a node. If biological plausibility is a concern, one might prefer to use the Hodgkin-Huxley model. (See also: other neuronal models).

To many practitioners, biological plausibility is not a primary concern. In most ANNs, each node is not really believed to be implemented by a single neuron in the brain, and thus using a more realistic activation function does not necessarily make the network more biologically realistic.

Personally, I don't think looking at the behavioral accuracy of a biological neural network (compared to an artificial neural network) says anything about the network's biological plausibility. Using more abstractions might produce more similar behavior due to less overfitting. Another concern is that the ANN could produce the same behavior as a BNN even if the actual computations involved are quite different.

Answer 2

Due to my newness to the field, I can only talk about comparisons of biological plausibility when discussing the Neural Engineering Framework (NEF) and functional modeling. What is missing from this answer is a purely bottom-up modelling perspective in the same vein as the Blue-Brain project, but I'll leave that to another user.

One of the claims driving the push towards cognitive modelers using the NEF is the fact that it is more biologically plausible than back-prop ANNs. What is typically discussed is the neural mechanism for encoding, decoding and learning of the NEF has more scientific evidence claiming their plausibility than back-prop ANNs.

From a more systematic biologically plausible perspective, the NEF is also used to build neural networks that use Semantic Pointer Architecture, which is also claimed to be more biologically plausible for a number of reasons. First of all, the number of neurons required for most models in significantly less than other proposed models (see Eric Crawford's work on knowledge representation). Secondly, modifications to the network mimicking the effects of neurological disorders can be shown to have the same effect of those disorders (see Dan Rasmussen's work on general intelligence).

In terms of the bottom up perspective, the NEF could help here as well. The NEF is neuron-model agnostic, so it can be used with any of the neuron models you mentioned in the question. Hypothetically, it could be possible to create a similar network to one grown on a Neurochip and compare the various inputs and outputs as you suggested, but I've never heard of such a study being done.