The concepts related to human decision making can get philosophical very quickly. But, this discussion will not be about philosophy, metaphysics or religion. Instead, lets explore how human decision making differs from current artificial intelligence technology, and how AI is more of an art from than a science.
Whether based on previous experiences and memories, or whether as a simple reaction to noxious stimulus, human decision making is ultimately learned from inputs of five senses. Lets take a residency training program choice as an example. One can only wonder what the outcome will be after thousands of medical school seniors make their choices in ranking training programs virtually, without visiting them. Do a smell of carpet, a certain temperature in the interview room, or pressure of a handshake contribute to the usual ranking process? Since up to two thirds of communication is considered to be non-verbal, these small nuances at least contribute to, if not dominate the overall subconscious impression about any given training program in an applicant’s mind.
A human “neural network” learns to make decisions based on dynamic, numerous, and diverse inputs. This is very different from artificial neural networks. Artificial neural networks are designed to output a simple program of weights, an equation of sorts. Any task becomes a problem with tens, hundreds, thousands, millions of unknowns arranged in matrix forms. AI solves equations. At the end, AI assigns probabilities in classifying images, text, time series, or whatever other input. As long as the input can be represented as a matrix, neural network will solve it. For example, a deep neural network will calculate a probability score after analyzing a set of pixels in an image. A 32x32x3 matrix of color pixels becomes a traffic signal with a probability of 99%. An audio waveform will represent a spoken “three” with a probability of 75%, or an “eight” with a probability of 80%. In a simplistic way, this whole process can be explained by a solution to the following problem.
What is x in the following sequence?
So, it is tempting to describe neural network as a fancy calculator to others. But, it is a little bit more than that. Someone had to come up with a structure of hidden units that modify and process inputs. The art part becomes evident as humans adjust architectures, change adjustment lambdas, and do all kinds of other clever things to get the network to improve its accuracy by one or two percent. Linear regression and back propagation can be understood with basic computing and calculus knowledge. But, linear regression gets accuracy only so far, and may not be appropriate for a given task. Humans have to try ReLUs, sigmoid functions, optimizers, larger GPUs, and patch methods. Self learning nets are now on the horizon. But, even those will only be as good as the ingredients that humans add to the computing salad.
Thousands of years ago, a human drew a buffalo hunt on a cave wall. The art of representing what is important to human in the form of a painting was born. Years later, lines, simple shapes and dots were used as symbols to represent sounds, words, and concepts. Fast forward to now, and we have a masterful combination of high and low bits that represent a multidimensional descriptions of objects, written and spoken concepts, or higher dimensional constructs of inputs that are important to humans. The paint pigments and canvases of today are partial differential equations and GPUs that allow new artists to paint representations of the world.