This review article is a good start, but can be made more complete by including other approaches to implementing models of single neurons in silicon. Also, there should be a discussion on the merits of implementing models of neurons in silicon, which are not obvious to most readers. That discussion should also explain the choice of implementing models in analog, digital, and mixed-signal circuit technologies.
A first merit of a silicon neuron is as a neural prosthesis, interfacing with biological neural tissue. There has been some work on implantable silicon neurons by the Diorio group at University of Washington, among others.
A second merit is as a modeling tool. For that purpose, a purely digital implementation on an FPGA is perfectly adequate, as shown by the Butera group at Georgia Tech. Analog conductance-based models are also useful for modeling, but I would include more recent approaches than the seminal 1991 Mahowald and Douglas implementation. Since then, there has been much progress in modeling the important details of membrane dynamics and channel kinetics in the Hodgkin-Huxley model. See for instance the paper "A Bio-Physically Inspired Silicon Neuron" by Farquhar and Hasler in IEEE Transactions on Circuits and Systems I, 52(3), 477-488, 2005. Besides the HH conductance-based model, there are several silicon implementations of other models, such as Morris-Lecar and Fitz-Nagumo.
A third merit of a silicon neuron is as a cell in a larger network for neural computation. The larger fabric of the synaptic network would be part of the (future) "VLSI Implementations of Neural Networks" page (which you link to), but it is important to mention here how silicon neurons interface to implement synaptic connections, since that affects their design. You could mention up front that this article concerns spiking neural models only and does not consider mean-rate neural models as used in some models of large-scale neural computation (with a link to, for instance, the Hopfield model). The spiking model is attractive for scalable implementation of biologically realistic neuronal networks, because each spike represents an event in time that can be efficiently routed using the address-event communication scheme (references). A silicon neuron can thus be taken as an address-event transceiver, where spiking events generated by the neuron are routed through an address table for synaptic input to other neurons in a silicon array.
The term "hybrid model" suggests that it refers to a combination of both integrate-and-fire and conductance-based aspects in the neural modeling. The example shown is strictly an integrate-and-fire model with additions, so "extended model" would be a more appropriate term. However, it would be good to discuss models that actually combine the simplicitity of an integrate-and-fire mechanism (for event generation as neural transceiver) with the biological realism of conducance-based membrane dynamics (in the synaptic input). See for instance the work by Vogelstein et al. (IEEE Transactions on Neural Networks, 18(1), 253-265, 2007) and references therein.
Biological neurons are distributed elements, and it would be good to discuss implementation of more advanced models with multi-compartmental dynamics and axonal/dendritic propagation in silicon. See for instance the work on dendritic trees by the Elias group at University of Delaware. Before the advent of silicon CMOS integrated circuits there was also pioneering work on implementing axons as traveling wave amplifiers.
The font in some of the figures is very hard to read, making it difficult to follow the text.
Comments by another reviewer
Fonts are hard to read in some figures, making it difficult to follow the text.
Since no one is selling any silicon neurons there is no market basis on which to compare their relative merits, so a more complete listing of alternative approaches being explored would be useful for future developers.
Didn't the Boahen lab also develop some membrane models that capture the sodium inactivation or variation of sodium time-constant behavior?
It might be good to discuss the problems of matching and precision, at least to mention them as concerns for developers and interesting for biologists who wonder how brains deal with or utilize heterogeneity.