Disclaimer: Working on neuromorphic chips.

Several options available but followings are superior to others in terms of capturing the real neuron/synapse behavior with a simple and compact VLSI implementation.

For neurons: Exponential adaptive integrate and fire models @ real-time.

For synapses: STD/STP + LTP/LTD circuits @ real-time.

Yes thats what i mean. IF models are approximations. I thought the HBP was implementing compartmental simulations. Also the models we have for LTP are at best speculative

Every model is an approximation :) I don't work at HBP, therefore can't speak for their implementations. These are the SoTA of nm engineering. What do you mean by compartmental simulation?

Multi compartmental models of neurons simulate the local voltage dynamics of membrane in detail , by considering the neuron as a graph of connected cylinders , and uses the cable equation along with hodgkin-huxley-like models to model the ionic channels of the membrane. It's the most detailed and closest to reality way to simulate neuronal dynamics. Most neuromorphic chips simulate much more abstracted, integrate and fire neurons. In both situations, plasticity is the big unknown because we know so little about it (and it's so complex).

https://en.wikipedia.org/wiki/Multi-compartment_model

> It's the most detailed and closest to reality way to simulate neuronal dynamics

Nope, you can keep going with the same logic: HH is just an approximatiin of the molecular kinetics and what you really need are the FEM models in 3D with all the protein pathways etc etc.

But this leads to an entirely intractable project. The science lies not in reproducing the exact neurons bug for bug, but keeping only the necessary details, within technical possibility, for explaining some observed phenomenon. LIF and AdEx neurons seem like a good compromise for neuromorphic hardware.

the HH model is good enough to reproduce almost everything that is recorded from neurons (Thats why H & H got the Nobel prize after all). Even dendritic regenerative spikes are reproduced by such models. I think it's generally acceptable that one does not need to do molecular dynamics to recreate membrane voltages. LIF and AdEx are very crude approximations though (i.e. no dendritic spikes, no plateaus, impossible to use them to approximate compartmentalized calcium levels that induce plasticity). And if you go that route, you have to justify why they are a better choice than e.g. Izhikevich neurons or indeed just sigmoid units.

> I think it's generally acceptable that one does not need to do molecular dynamics to recreate membrane voltages

everyone draws the line somewhere, this is yours. Even we take this statement to be true, you still have metabolic networks changing transmitter concentrations, dendritic arbors evolving in entirely unidentifiable ways, etc. The goal of modeling is not to say that every detail is there, but ones relevant to account for specific feature of data.

While there is more detail that can be simualted, there is a vast very successful literature using compartmental models. Markram's work was the simulation of a cortical column with compartmental models anyway.

The problem with this of course is the necessary level of detail depends on the phenomenon: simulating seizure propagation is probably different from simulating emotions. For more complex tasks we have no clue what that level is.

Somewhat off topic... but do you happen to know if the dendritic tree and its functional subunits in multi-compartment models can be treated mathematically as a multilayer network of simpler neurons?

I've been wondering if dendritic arborization means that current deep learning ANNs are hopelessly far away from biological reality, or if perhaps with deep networks simple ANNs could indeed learn to compute in a similar way to complex biological neurons, just over many layers of artificial neurons.

Absolutely, dendritic trees are generally considered active and can elicit dendritic spikes. This 2-layer model is well studied in hippocampal CA1 neurons for example[1]. ANNs are far from reality but they do validate connectionism as probably the correct abstract model of learning. The thing that's harder to crack is plasticity i.e. the learning rules. Plasticity in real neurons is a very complex process and there is no indication that anything like backpropagation takes place.

1 https://www.sciencedirect.com/science/article/pii/S089662730...

2 https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4279447/

3 https://www.ncbi.nlm.nih.gov/pubmed/18270515