The key word here is "fast". Electrical and fast chemical transmission at synapses are presumably responsible for perception, the immediate reactions of an organism to it's environment, thought, etc. Neuromodulators, etc, work on a slightly longer time scale, and are presumed responsible for homeostasis, learning, changing the system state, etc. (By the way, it is quite untrue that the electrical side is only capable of summation and negation. Single neurons are capable of multiplication, for example (http://arjournals.annualreviews.org/doi/abs/10.1146/annurev.neuro.28.061604.135703)).

*Cleanly* separate, no. But separate to a degree that allows us to make progress based on approximations, yes.

I think the problem is that the original article was very sketchy, and understates the degree to which the chip development is influenced by biology (within the very considerable constraints of VLSI technology). The main advantage of the analog hardware approach compared to pure software simulations such as Blue Brain (which is an excellent project, I agree), is that it runs very fast, which will allow us to study slow processes such as long-term plasticity, the interactions between neuromodulation and spike-timing-based learning rules, etc, in a reasonable time frame, as well as to do exhaustive scans of parameter space.

I am one of the researchers involved in this project. You are right, of course, that we are only simulating 0.1% or less of the complexity of the brain, so even if we simulate 100% of the number of neurons in the brain, we are still orders of magnitude of complexity away from reproducing a brain, let alone understanding it.

However, we have to start somewhere and, in the words of Henry Markram (Blue Brain Project) "If we don't start now, when do we start?". The neuron models in the chip ignore spatial processing in the dendrites, but they do reproduce the variety of firing patterns found in real cortical neurons. The models of the chemical synapses incorporate have both short-term (adaptation, etc) and long-term (learning) plasticity, based on experimental data. Neuromodulation (by dopamine, etc) could be simulated by modifying synaptic and neuronal parameters, using the digital logic on the chips, although we haven't really thought about this yet.

The FACETS project involves experimental neurobiologists, theoreticians, modellers, and solid-state physicists (who are developing the chips). We are very aware of the necessary simplifications we are making, but we are also confident that we are making progress both in understanding brain function and in developing new approaches to highly-parallel, fault-tolerant computing.

This is a good point, but on the other hand, the cortex is not that far from being 2D - it is only about 2mm thick, and has a laminar structure. Most of the volume of the brain comes from (i) the intricate folding of the cortex, to pack as much surface as possible into the volume, (ii) the medium- and long-range interconnections between different brain regions.