Can processing effort for sub-tasks in neural networks be measured?

Question

I often heard statements like:

80% of your brain processing is computing the effect of gravity

or, similarily:

You only use 20% of your brain power

My question isn't about the truth of these statements but if they are well-formed: can we measure 'processing' in the brain (or a sufficiently complicated artificial neural net) to quantify how much of the processing resources are allocated towards specific sub-routines?

As an example of when this is doable in another domain: consider a computer with a standard von Neumann architecture. In this case we can tell how much processing goes into sub-routine by simple counting the number of steps each subroutine holds the processor. However, this simple counting is only possible because of two features:

1. There is a clear central processor, and data is not transformed by anything except it.
2. There is a clear duality between software and hardware allowing us to identify when a specific sub-routine (software) is being allowed access to the processor (hardware).

Both features are missing in a computer with a neural architecture. How can we measure the resources that a computer with neural architecture (example: the brain) devotes to subtasks?

---

Notes

I am interested in settings where there isn't an obvious localization of sub-tasks. So an answer like:

we can measure the amount of resources that go towards recognizing faces by looking at the metabolic uptake of the fusiform face area.

Is not that interesting to me, since it is only valid for sub-tasks that happen to have specially-devoted areas. Of course, if there is an argument that any reasonable (or more restricted: human-like) neural-architecture has to compartmentalize its computations then I would like to hear that. However, the answerer would be left with the task of explaining how to identify the group of neurons that should be associated with a specific sub-task.

Answer 1

Well firstly, what are those statements really saying? How do you measure "processing" or "power" as it relates to the brain? For an electrical engineer working in communicatons, it's easy: you just take the square of the amplitude of the signal (signal power) or count the number of instructions per second (processing power). But that's the discipline where processing power is defined; making an analogy to a brain brings a lot of caveats. What's an instruction? How many little instructions can you break one instruction down into? Where do you want to draw the line in a color spectrum that divided orange from yellow? Or do you want to draw more lines and have dark and light yellow? Or yet more lines...?

Hans Moravec, a machine intelligence researcher, is famous for his comparisons of technology to organismal processing power. Here he explains a little bit about how he comes up with his numbers, but they're really not satisfactory to me. He seems to take one example (the retina) and judge it by what it can do empirically with an image, then extrapolate from there to the rest of the brain. So this only considers a single network topology in a system that has millions more variations in topology.

The major problem that is a typical complaint by biologists when people use the computer science analogy for brains is that metabolism and functions are too closely coupled to easily separate. Functional mechanisms can occur as negotiations between molecular networks (including genetic transcription) in addition to the electrical signals that electrical engineers and computer scientists generally think of as "processing".

This is especially outlined in the glutamine, gaba, and adenosine molecular networks (i.e. the transmitters are closely coupled to metabolism through ATP and the Krebb cycle). A lot of recent evidence has also put a larger emphasis on astrocytes. This paper is about epilepsy, but takes some time in the introduction to develop ideas about astrocytes and brain metabolism complete with references. The brain-metabolism material is typical in any neurobiology text book; astrocyte participation is a more recent line of research. There's also still a lot of research coming up in molecular mechanisms for memory.

There have been all kinds of measurements on processing power, including at the genetic level, but all these measurements make an assumption about what a bit is by breaking biological subsystems into two states (this, 1 and 0). But then later down the line, somebody will find that there was actually a third significant functional state. So it's all about where people want to draw lines.

Another problem (more at the macro level now) is that the same task will have different processing requirements depending on how robustly the organisms internal model is approximating and predicting the external world. So a task that you know what to expect for will take considerably less "processing power" than one where your expectations are not met. Karl Friston wrote an interesting review suggesting a modeling framework that might help to explain this.