Page 75 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

2. Third Gen AI 63

when combined with oxygen], CT based on microcalcification of dead cells, PET
based on radioactive positron agents), the neurobiology discovered the missing
half of Einstein brain to be the nonconducting glial cells (in Greek: glue cells
made mostly of fatty acids) that are smaller in size, about 1/10th of a neuron, but
doing all the work except communication with ion firing rates. Now we know a brain
takes two to tango: “billions neurons (gray matter) and hundreds billions glia (fatty
acid-insulated white matter).” Instead, the traditional approach of SDL is solely
based on Neurons as processor elements (PE) of ANN overlooking the name recog-
nition. Instead of SDL training cost function the Lease Mean Squares garbage-in and
garbage-out, using LMS error energy,

2

.

E ¼ desired Output S pairs actural Output S pairs ðtÞ (3.3)
b

1. Sensory Inputs: “While agreed, the signal; disagreed, the noises”
. .
X pairs ðtÞ¼ ½A ij Š S pairs ðtÞ (3.4)
.
2. Power of Pairs: The agreed signals become the vector pair time series X pairs ðtÞ
with the internal representation of degree of uniformity of neuron firing rate
.
S pairs ðtÞ described with Ludwig Boltzmann entropy with unknown space-
variant impulse response functions mixing matrix [A ij ] and the inverse by
learning synaptic weight matrix.

2.2 THE INVERSE IS CONVOLUTION NEURAL NETWORKS

.
S pairs ðtÞ¼½W ji ðtފX pairs ðtÞ (3.5)
b
The unknown environmental mixing matrix is denoted [A ij ]. The inverse is the
Convolution Neural Network weight matrix [W ji ] that generates the internal knowl-
edge representation.
Our unique and the only assumption, which is similar to early Hinton’s Boltzmann
Machine, is the measure of degree of uniformity, known as the entropy, introduced first
by Ludwig Boltzmann in MaxwelleBoltzmann phase space volume probability W MB .
Since there are numerous neuron firing rates, the scalar entropy becomes vector
entropy for the internal representation of vector clusters of firing rates.
.
fS j g/ S (3.6)
We apply biological NI and unsupervised deep learning (UDL) on BNN which is
derived from the first principle, isothermal brain at minimum Helmholtz free energy
(MFE). Then from convergence theorem, and D.O. Hebb learning rule, we derive for
the first time the mathematical definition of what historians called the “missing half
of Einstein brain” namely glial cells as MFE glue forces. In other words, rather than