Several research groups have recently announced successes in modeling ‘real neurons’ in impressive-sized quantities - each with the aim of seeing whether their model has explanatory power for actual biological brains.

A secondary aim of all of these models is to explore which features are important, and which can be safely left out.

Modeling rainfall

Suppose one wanted to determine the shape of puddles of water after a shower of rain.

One could choose to model :

  • the dynamics of the drops of water forming in the cloud

  • the wind pattern’s effect on collections of droplets

  • the precise location on which each drop falls

  • the dynamics of water-flow as it drains downhill

  • etc…

However, the only important factors are the quantity of water, and the profile of the land on which the puddle is forming. The puddle’s shape is determined not by the behaviour of the water, but by shape of the ground for a given depth of puddle.

Building the most detailed model possible - and subtracting

Biological neurons are complex to model. It may be that :

  • some of the types of hysteresis found in neural output is significant to their operations

  • be that the ‘3D’ substrate between neurons plays a role in learning (by defining chemical gradients between correlated signals, for instance)

  • etc…

There are innumerable features and subtle distinctions that could separate a workable simulation (i.e. one that could, at scale, lead to brain-like behaviour) from one that does not have the right ‘spark’.

Right now, it is extremely difficult to know at which point a model move into ‘complete overkill’ territory. But, once a comprehensive simulation is available, it’s possible that it could be stripped back to its core - revealing the ‘essence’ of what makes neurons work effectively in the brain.

Building a simplistic model - and adding

Another (more mathematically appealing) approach is to try and guess at the important features now - and simplify everything else away.

In some respects, this approach is more research-friendly. Each different research team can pick a different set of features to incorporate into a model, do lots of experiments, and make recommendations about next steps. A selection of papers can be generated from each iteration of the process. However, if there are twenty embeddable features, and each researcher chooses ten, it’s a method that comes with a lifetime employment guarantee…

The right abstractions will be obvious (in hind-sight)

Whichever way is used to acheive the goal of a workable brain-scale network, once the correct palette of parameters and abstractions is identified, later research will look back in amusement at the range of dead-ends and blind-alleys were explored.

But it’s a hard slog to get to that point.

Martin Andrews

{Finance, Software, AI} entrepreneur, living in Singapore with my family.

blog comments powered by Disqus