The spectrum of human sensory inputs seems to fall within certain ranges suggesting normalization is built-in into biological NNs?

It also adapts to circumstantial conditions, e.g. people living in a city with certain factory smell eventually don't perceive the smell anymore, at least consciously (within working memory) / it adapts to a new baseline?


Yes, for many sensory inputs there is indeed something similar to normalization. But its not rally the same as in classical data analytics compared to what eg min/max normalization does or other technics.

Lets look on some examples and considerations:

  • mammals don't perceive heat or loudness in a linear way. This is because already many sensory receptors have chemical / physical limits. Double decibels will not perceived with double intensity. Inside your ear, the small hammer and abil will brace to protect you. --> its like normalization with logarithmic effects applied.

  • heat perception is more like a difference integration than a absolute temperature measurement. Its measured via H+ ions flow in mitochondria in the cell (if i recall correctly)

  • On the neuronal side gradual signals in the dendrites (analog signal) sum up gradually to later form an spike at the axon hill. where in turn a fire frequency is then encoded - the maximum frequency of this serves as a a natural maximum limit. I remember that grasshoppers increase axon fire frequency when objects started covering more ommatidial area on their "eye". The more of their "eyes" are covered by the shadow the more input on the neuron --> higher fire rate.

  • a lot of sensory input is post processed in higher cerebral areas. Eg. compared to what is expect able and heuristics are applied to compare a signal with former events.

  • when doing computational data analysis we may want go for accuracy and maximum comparability. Mostly on all data that could be available. --> eg. with respect to properties of a standard normal distribution. Hence we put some effort to be accurate and know the true parameters, remove outliers and so on --> big data comes into play here. Nature in contrast strives often for efficiency with the means of reaching the minimal required with minimal resources.

Summary: Compared to normalization in an analytical sense (eg. mean, min-max or other feature normalization techniques), nature is often only interested in the current difference between stimuli. And this only within some relevant range. Other information is not integrated. And normalization with the goal of making measurement points comparable only happens within this range of the mapping function provided by the sensor/neuron/receptor whatever!

So this should also answer your question about, why you are not smelling something in the city after a while any more. However, this for sure happens at higher cerebral regions (it might also be that your smell receptors saturate) its the same principle. Your consciousness just saves energy by not concentrating on something that is anyway not changing.

If you want to read more have a look here: https://en.wikipedia.org/wiki/Weber%E2%80%93Fechner_law

| improve this answer | |
  • 1
    $\begingroup$ But is that normalizing or squashing the values in a range? $\endgroup$ – DuttaA Jan 10 '19 at 15:20
  • 2
    $\begingroup$ @DuttaA, I think within the range of the mapping its comparable to normalisation. Outside of the range its just dropping/ squashing information. There is this law by weber and fechner: "en.wikipedia.org/wiki/Weber–Fechner_law". Just by intuitition you will be able to say that silent and loud are equidistnace to louod and louder but you can not judge on extreme silent or loud. $\endgroup$ – PlagTag Jan 10 '19 at 20:39
  • $\begingroup$ @LeoGallucci the link gives 404 $\endgroup$ – PlagTag Feb 18 '19 at 8:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.