Is noise bad for neural computation?
Imagine a CPU of a computer receiving exactly the same input repetitively. The CPU gives the same output every time. If it does not, obviously it would be a problem. Right? But see Fig. 1. It shows super-imposed traces of excitatory post-synaptic potentials evoked repetitively by the same stimulation. As you see, the response of the neuron is different every time. This is of course because of the stochastic mechanism of synaptic transmission. This means that the neural system is fundamentally stochastic, at the level of synaptic transmission and also ion channel kinetics.
A question: Is this a problem?
Fig. 1 EPSPs (excitatory post-synaptic potentials) evoked in the post-synaptic (2nd) neuron by activating action potentials in the pre-synaptic (1st) neuron. While the size of the presynaptic action potentials (inputs) is constant,, the size of the EPSPs (responses) varies, illustrating the stochastic nature of synaptic transmission. (From pyramidal neurons in layer 2/3 of mouse visual cortex. Unpublished data)
Fig. 2 Every time you give the exact same input to a computer CPU, the exact same output is expected to be given. In neural networks, which we consider biological computational systems, it does not work in this way. Fundamental mechanisms underlying signal processing in neural networks are probabilistic (synaptic transmission and ion channel kinetics) and hence the output is stochastic: every time it gives a different result. Is this a problem?
At a grance, it may sound problematic: if the output of a computer is different every time the same input is given, do you trust the computer?
To me, this suggests that biological computation is conducted based on principles that are very different from computers. The question is how?
To address the issue, I am conducting a project right now that is based on the following working hypothesis.
The stochasticity of neural mechanisms helps neural signal processing.
The reason is that sensory input signals neural systems receive in the everyday life are noisy, ambiguous, and underrepresented. Hence, it is beneficial if the signal processing mechanism is NOT deterministic but rather responds stochastically, leaving multiple options available. Only by dynamic interactions between different brain areas detecting different features of sensory inputs, the final interpretation, the perception, is established. That is,
Being non-deterministic actually helps the neural system to estimate probabilistically the most appropriate responses to the sensory inputs.
In other words, the worldview we have is the result of "negotiations" between different properties of the sensory signals detected in different neural circuits distributed in the neural system. Hence, our perception of the world is, by nature, a most likely estimation (with the best of the ability of the neural system) of what is out there.
In this project, I investigate the roles of the stochastic mechanisms, i.e. the intrinsic noise, in decision-making processes. To do so, I implement biologically defined, computer-simulated noises while analysing the neural competition dynamics explained on the other page.
The biologically defined noises are categorized as follows.
1.noise in evoked synaptic signal (EVK)
2.synaptic noise (background noise) (BKG)
3.noise in ion channels (ICH)
4.noise in incoming spike trains (SPT)
5.fluctuation of (intracellular and extracellular) chemical/physical environment (ENV)
By using the hybrid neural circuit of neural competition described on the other page, and by implementing modelled noise defined above to the dynamic clamp system, we investigate the possible positive roles of biological noise in signal processing.
Fig. 3 Biologically defined intrinsic neural noise of 5 different origins