Page 74 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

62 CHAPTER 3 Third Gen AI as Human Experience Based Expert Systems

Machine can enjoy 1000 copies which each explore with all possible different
boundary conditions that become collectively the missing experiences of a single
machine. Thus, MIT professor Marvin Minsky introduced the Rule-Based Expert
System (RBES) which in nowadays has become the Experience-Based Expert Sys-
tems (EBES) having the missing common sense of human behavior that “A rule is
made to break as intelligent behavior.” For example, a driverless car will stop at
different “glide lengths” near the traffic red light as RBES. However, EBES will
glide slowly through the intersection in the midnight when there is no detection
of any incoming car headlights on both sides.
The intrinsic assumption is the validity of the Wide Sense-Temporal Average
(WSTA) with the Wide Sense-Spatial Average (WSSA) in the MaxwelleBoltzmann
Probability ensemble, so we still know the time t and x of those cases which happen
in the midnight (in the desert).
We examine deeper into the deep learning technologies in three advances, which
are more than just architecture and softwares to be massively parallel and distributed
(MPD), but also sharing results at Cloud’s Big Data analysis (BDA). Concurrently
Werbos (“Beyond Regression: New Tools for Prediction and Analyses” Ph.D.
Harvard University 1974) and [9] McCelland and Rumelhart, PDP, MIT Press,
1986 developed backward error propagation. Recent breakthrough is due to the
persistent visionary of Geoffrey Hinton and his prote ´ge ´es: Andrew Ng, Yann LeCun,
Yoshua Bengio, George Dahl et al. (cf. Ref. [7] Deep Learning, Nature (2015))
applying cloud’s BDA to MPD hardware (GPU) and software (Tensor Flow) as
follows,
1. the hardware of graphic processor units (GPU) has 8 CPU per rack and
8 8 ¼ 64 racks per noisy aircooled room size at the total cost of millions
dollars. The massively parallel distributed (MPD) GPU has been miniaturized
as a back-plane chip.
2. The software of Backward Error Propagation has made MPD matching the
hardware over three decades, do away the inner do-loops followed with the
layer-to-layer forward propagation. For example, the Boltzmann machine took
a week sequential CPU running time, now like gloves matching hands, in an
hour. Thus, toward UDL, we program on mini-supercomputer and then program
on the GPU hardware and change the ANN software SDL to Biological Neural
Networks (BNN) “Wetware,” since the brain is a 3D carbon-computing, rather
2D Silicon computing; it involves more than 70% water substance.
Historically speaking, when Albert Einstein passed away in 1950, biologists
wonder what made him smart and kept his head for subsequent investigation for de-
cades. They were surprised to find that his head weighed about the same as an
average human’s, at 3 pounds, and by firing rate conductance measurement the
same number of neurons, about ten billion, as an average person. These facts sug-
gested the hunt remains for the “missing half of Einstein brain made of 10 billion
neurons [8].” Due to the advent of brain imaging (f-MRI based on hemodynamics
[based on oxygen utility red blood cells become ferromagnetic vs. diamagnetic