What are the current limits of artificial neural networks?

Hi, Jonathon. You seem to have lots of good answers, but let me just add my own brief list to the mix.

1. They implement a massively simplified, 1960s-era model of neural functioning that doesn’t begin to reflect how brains actually work (this one has a half dozen sub-points I won’t elaborate here).
2. They assume a uniform, layered matrix of neurons, completely ignoring the specialized structures, functions and architectures of real brains.
3. They rely on an algorithm called back propagation to get learning to work, an algorithm that has no real counterpart in real brains.
4. They make little or no use of inhibition, which is critical to brain functioning.
5. They ignore the large-scale effects of neuromodulators, a cortical broadcast system the brain uses to alter its function in response to current needs (including the role of dopamine in triggering the learning process).
6. They rely almost exclusively on the process of recognition, which is one of the most basic brain functions and doesn’t require the sort of multi-step processing required for language and cognition (though it’s pretty amazing just how much they can do working with just recognition).
7. They employ a wide variety of optimization algorithms to direct the learning, and the selection of the algorithm can have profound effects on the result that aren’t yet well understood.
8. They require truly massive resources to implement, including having to divert full power from Hoover Dam during their runs (okay, maybe not quite that much),

I’m sure there are more limitations I could add to this list, but that’s enough for the moment. I should mention that many if not most of these limitations reflect a generation of ANNs (artificial neural networks) that is still in its infancy, and many will be overcome as the field matures. The current emergence of dedicated neural-network hardware that can simulate millions of neurons operating in parallel should provide an interesting opportunity to go beyond some of these limitations. Still, it is worth noting that, despite the truly amazing things that ANNs are capable of doing, they are still a long way from the goal of creating true artificial intelligence.