“Artificial Intelligence” was supposed to emulate Human Intelligence functioning in machines, and in line with neurons that make biological intelligence; AI uses notional “neural networks” for realizing deep learning. However, there are some glaring omissions / oversimplifications which could severely restrict the emulation.
- Physical neurons are made up of an “ionic channel tree” made up of morphology of several twigs and branches leading to output amplifier stage (the axon hillock “N0”). All Inputs ( synapses ) get connected through specific morphological linkage that characterises its “conductance” and “signal propagation delay”. The sequence and delays in which various inputs arrive at a twig also make a difference in the outcome. Notional neurons are just nodes where all inputs simply converge and all inputs are logically identical. It works by snapshot method.
- Inputs are certain ionic charge that flows through leaky ionic channels up to axon hillock with twig and branch logic similar to AND –OR-NAND-NOR gating and at axon hillock if the charge arriving in bursts cross a threshold than the neuron generates ON/OFF type of output but again as a pulse. Notional neurons treat all inputs as ON/OFF states and multiply with corresponding weightage factors … those weightage factors must be learnt and assigned externally.
Sigma ( I1*W1 + I2*W2+I3*W3 + …. In*Wn ) > Threshold Value .. for output
- Learning comes by way of physical morphological changes such as change in cross section area of channels with autonomic changing of the weightage factors, depending on extent of using or not using the specific pathway over a period of time. Notional neurons get evaluated through a standard computer, where learning is effected based on success / failure feedback. Learning is based on pre-deterministic algorithm, limited by the intelligence of creator.
- Some inputs are “excitory” while some are “inhibitory”, meaning that some weightage factors have positive and some have negative sign. Also some inputs may be self feedback (“autosynaptic”), which means they fetch the output from the same neuron after propagation delay. Notional neurons do not have a feedback control or feedback based stabilization at individual neuron level.
- Physical neurons have multiple logic statements while Notional neurons are single logic statement which is even lesser than a twig.
- While physical neurons in effect solve analogue non-linear and differential equations … Notional neurons only solve linear algebraic equation.
Physical neurons take care of time sequencing while Notional neural net functions purely on combinational logic, by itself.
Most courses on AI-ML today focus more on development tools such as Python and statistics rather than faithful modelling of neurons and human brain at large.
AVINASH KHARE