 |
| |
 |
| |
 |
| |
 |
| |
 |
The number of inputs, outputs, and empirical constraints are limited only by
hardware memory. Inputs and outputs may be integer, decimal, or
classification category names.
BIN is similar to an artificial neural network. BIN can be implemented using a
feed-forward neural network perceptron architecture with two hidden layers,
but the similarity stops there. BIN is not a back-propagation neural network,
because it does not back-propagate error. BIN solves for both the structure
and link weights, resulting in zero error for the training cases, so there is no
error to back-propagate. BIN can also perform data smoothing for statistical
data, with geometric convergence. BIN achieves superior generalization
with no training, no "local minimum" convergence problems, and no endless
experimentation with hidden nodes and layers. BIN results are consistent
and predictable, unlike the random "black-box" results of neural networks.
The compact support property BIN exhibits also results in superior
high-speed performance, as only a small percentage of the structure
requires evaluation for any given input. In other words, BIN is just like a
neural network, except it is better in every way.
BIN is similar to Fuzzy Logic, except that BIN does not have the weighted
average distortions of Fuzzy Logic.
BIN is similar to curve fitting techniques, such as Bezier curves, cubic
splines, B-Splines. However, BIN works in any number of dimensions,
without the requirement for regularly spaced data along each dimensional
axis. BIN does not require iteration to account for constraint errors, and
does not require derivatives to be specified at each constraint.
BIN is similar to numerical model shape functions used in Finite Element
modeling. However, BIN can automatically provide shape function values
and partial derivatives for any element shape with any given node
configuration, replacing current manual techniques.
BIN can empirically map entire databases for faster search or data
compression. BIN can determine empirical mappings for complex functions
of several variables, visual pattern recognition, data processing and
interpretation, classification, and feature extraction. Inversion techniques
allow modeling, simulation, and complex statistical data analysis.
Applications: Stock market, race-horse analysis, welding procedures and
parameter specification, visually controlled welding, musical analysis,
geologic modeling, model initialization, numerical model shape functions,
model initialization, evaluation of medication treatments, orthogonal
block-variance statistical analysis, experimental parameter optimization,
process control, automated vehicle control, electronic warfare signal
classification, seismic arrival picking, customer service prediction, speech
recognition, optical character recognition, robotic kinematics and dynamics,
micro-seismic source inversion, cross-hole lithology interpolation, weather
prediction, edge detection, vectorization, ...
