Why has evolution made neurons use spiking?

Question

I'm going to be forward and say that I'm not a biologist. I don't claim to fully understand the functionality of a neuron from an electrical/chemical perspective... I'm curiously gazing from the outside in as a computer programmer.

Now, the big new thing these days with artificial neural networks is simulating them with biological precision- that is, with spiking, actually placing the network in Euclidean space, and all sorts of other goodies (Check out the Human Brain Project; it's really cool. Our own brains are computable, given a powerful enough computer. Probably has a lot of fascinating and horrifying philosophical implications too.), but this isn't the only model that people have commonly used for neural networks.

In the past, multilayer perceptrons were the big thing; I'm not going to go deep into explaining them, but essentially, each neuron simply performs a mathematical function on its inputs and then outputs it. There's no spiking in the sense similar to our neurons. A neuron could be blasting its output at nearly 100% indefinitely and there isn't any objective difference between that and it being at 0% or anywhere else.

Now, I sort of went off on a tangent, but what I'm trying to understand is- why this behavior is not observed in the system of computation that evolution has produced over the many, many years? Why do we use spiking neurons instead? We have instead developed a strange system where spikes travel down chains of neurons in patterns we don't really understand yet to produce the behavior necessary for the creature to survive. My question is, why has this been favored over something similar to multilayer perceptrons? Is it more energy efficient to use a spiking model? The brain does use a disproportionate amount of energy. I wonder if it is more effective as a model for computation.

Answer 1

One of the many advantage of an all-or-none system is that resources can be conserved for timing events that require synchronized collaboration between many cells (like locomotion). Binary behavior may also partially be a side effect of speed and efficient long-distance information transfer (which is one of the great advantages of neurons as cells in the first place).

It should be noted, however, that neurons aren't universally binary. It is more accurate to say that neurons have binary properties that allow neurons to behave in a binary manner (but not all neurons do). On top of that, many continuous (i.e. non-binary) processes underlie the neuron's excitability threshold.

Answer 2

One thing I think you're missing: a perceptron model and a spiking model aren't mutually exclusive. In fact, in perceptrons, people usually use a sigmoid function. That's not a coincidence: it kind of simulates a spike. If you were to model an organism such as C. elegans, whose neurons aren't spiking (but use "graded action potentials"), you would use another type of activation function (by the way, if you like the Human Brain Project, you should also check the Openworm project, which is much more realistic and already gives some interesting outcomes).

So, back to the point. When you try to model learning and generation of memory, you often end up dealing with plasticity. Just a quick explanation: a neuron does fire an Action Potential (AP) when its transmembrane potential gets above a limit. The potential itself changes due to the inputs from other neurons. But all the inputs don't have the same weight. Let's say you look at neuron A, which gets synapses from neurons B, C, D... If B fires at the same time as C, it can be enough to create an AP from A. But, even in combination with B or C, D doesn't elicit enough transmembrane potential variation to create an AP; it has to fire exactly at the same time as E, F, G and H to get enough "firepower" to activate A.

However, if each time D fires, A gets an AP (because of the simultaneous activation of other incoming neurons), then you could say D is "useful" for A, and (depending on the type of neuron and other parameters) the weight of D can increase, so that, after a while, D is able to elicit an AP in A just by firing at the same time as any other incoming neuron. That's a plasticity phenomenon. Note that here, it is a really simple hypothetical example. And also that the idea that learning results from plasticity is more and more being challenged.

So, if you want to understand how learning and memory happen, you might be interested in the weights of the connections only, and how they evolve depending on the network's activity. In that case, a perceptron model is a good way to go.

Now, the big new thing these days with artificial neural networks is simulating them with biological precision- that is, with spiking

So, as you already mention in your question, this is a matter of choosing a model that you're going to use (as a researcher trying to understand the brain). This is not trying to give an accurate depiction of how the brain really is. If you're interested in the weights of the connections, you can use a neural networks approach (and this is not outdated), and you don't care about a single ion channel on the membrane of a neuron. But alternatively, you can also dedicate your life to understanding the structure of a single ion channel, and how simple details can determine the whole working of the brain.

Projects such as the Human Brain Project are trying to create a "biologically exact" model of the brain. So, it involves modeling the spikes, as well as the network organization, as well as single ion channels. So, in that case they don't need to use paradigms such as artificial neural networks, since the network's behavior should be a result of the single neurons' modeling: that would be an emergent response. However, they don't have anymore the simplifications brought by other models, so, they need huge computational power, and they also face new problems: for example they have to give values to all the individual parameters of their models, some of which are currently unknown, and could even not exist at all.

So, I think this gives answers your question, in the sense that it is not the brain which evolved to pick a model, but the scientists who try to modelize one particular aspect of the brain.

EDIT: I see in fact this question is quite old; sorry it was a suggestion on the right which brought me here. But I think my answer can still bring an aspect which wasn't already mentioned in the other answers. So I hope you will see it.

Answer 3

The use of spikes is a mixture of their computational advantages and the limitations of the biological substrate in which they are implemented:

1. They can travel long distances at high speeds because they can profit from the saltatory conduction which is much faster than diffusion and it can be maintained over long distances. A digital signal is less prone to noise which is also a reason why our computers operate digitally.
2. Spikes can be modulated using chemical synapses which are a) directional b) can amplify the signal c) can modulate the signal d) can even inhibit the signal. Analog voltages have to be transmitted through electrical junctions that do not allow modulation.
3. Frequency coding (signal lies in interspike interval) has a wider dynamic range than voltage coding.
4. The spike generation can serve as non-linearity in the transfer function
5. The spike can be used to code temporal signals more efficiently by reducing the redundancy (only signal change).

Btw. do not expect too much from the human brain project - it is a lot of talk and no results.

Answer 4

Neurons are binary, in that they either "fire" or "don't fire". There is no intermediate, because neuronal firing (biochemically) is an avalanche process. Excitatory triggers raise the membrane potential and inhibitory triggers lower it. Once the "threshold voltage" is reached, the neuron lights up and fires. This is how "decisions" are made by neurons: inputs from many inhibitory and excitatory neurons are summed, and if the sum is above the threshold voltage, the neuron in question fires. To use computational language, a neuron is a logic AND gate with hundreds of inputs.

Read the wikipedia article on "action potentials" for a more in depth discussion.