Page 72 - Artificial Intelligence in the Age of Neural Networks and Brain Computing

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

unless we are willing to change our mindset of current rule-based AI model toward
useful experience-based AI model, a safer and nuisance-less DAV is not going to
happen. For example, we need the Poincare and John Von Neumann Ergodicity
concept, since a DAV has a limited life cycle to experience human lifetime driving
experience, we shall duplicate by computer modeling and simulation that a hundred
1000 DAVs share their total occurrence events together through Cloud computing
(like Sophia did by a team of Sophia in Cloud to impress the United Arab Emirates
[UAM] to admit her as the first machine citizen of UAM). There are five more prin-
ciples; we need to relax our mindset as follows.
1. While the probability theory is defined by Komogorov [4,5] by a triplet: [a closed
set, a metric, and a measure] in a squared brackets closed set. A general
automation modeling to deal with human must be based on both the open set
FMF rounded brackets. Moreover, the fuzzy logic and the spatial ensemble
average Ergodicity Principle.
2. Analyticity: There is an analytic cost energy function of the landscape there
remains the set of N initial and boundary conditions that must correspond1-to-1
with the final set of N gradient descent results.
3. Causality: Artificial neural network (ANN) takes from the initial labeled or
unlabeled boundary conditions to reach a definite local minimum.
4. Deep learning (DL): Adapt Big Data the connection weight matrix [W j,i ] between
j-th and i-th processor elements (about millions per layer) in multiple layers
(about 10e100).
5. Neumann Poincre Ergodicity Theorem: Rather than a limited time duty cycle
using noise to explore environment variability to gain experience, we use
another Ergodic Principle to explore thousands to million copies of machines to
explore the boundaries of knowledge to determine the “landscape of knowl-
edge” within big databases in the Cloud for downlink and uplink.
6. Unsupervised deep learning (UDL): We consider unlabeled data should be based
on biological neural network (BNN) of both 10B neurons and 100B house-
keeping glial cells; the experience-based expert system can increase the trust-
worthiness and understanding.

Multiple Layer Known Deep Learning [6,7] (Fig. 3.3):
To increase the probability of detection of A and minimize the false alarm rate B,
we clearly need multiple cuts to separate Bs from As.
We wish to improve the supervised deep learning with least mean square (LMS)
cost to become self-organization “follow the leader” as unsupervised deep learning
with thermodynamic equilibrium at constant temperature T o at minimum free energy
(MFE).
. . . . . .
Supervised cost function is taken as min k A B k¼ min k A C þ C B k¼
. . . . .
min k A C kþ min k C B k ifft, the cost of search for C city is linear at
boundary.
We review thermodynamics to answer why we should keep our head blood tem-
perature constant at 37 C. This is necessary for the optimum elasticity of red blood