From Perceptrons to Tesla Vision
Part 4. Neural Plasticity and ANN Training
This is the fourth, and last, in my series of posts about biological and artificial neural network (ANN) solutions to driving an automobile.
Part 1 summarized some of the early history of attempts to use Formal Neurons to build models called Perceptrons.
Part 2 discussed issues related to the architecture that is needed to make an ANN operate in ways that are somewhat analogous to biological neural networks.
Part 3 Described how a “properly constructed” ANN based on dynamic attractor states has some advantages over rule based models.
In this post I will discuss how an ANN can be properly constructed.
As stated in my previous posts, any problem that can be solved algorithmically, can be solved with either a rule based approach or with a properly constructed ANN. This is true even for very large and complicated problems such as how to apply self-driving to automobiles. Ideally, a self-driving system would make an optimal decision about what action to take (accelerate or decelerate, turn steering wheel left or right) for any conceivable situation the automobile might find itself in. In a rule based system, this could take the form of a set of specific rules about what should be done under each specific condition the car might potentially encounter:
C₁, C₂, C₃, … Cₙ, where C is the condition and n is a potentially HUGE number,
in which case the number of rules would have to also be HUGE, making a rule based approach impractical.
An ANN could, in principle, solve this problem more efficiently by using vectors pointing to dynamic attractors, and then arranging that conditions similar to each attractor have vectors that point to positions near that attractor (I gave a tinker toy example of how this works in my Part 3 post). The hard part in ANN design is going to be figuring out how to get the (potentially HUGE) number of vectors for similar situations to point to the proper positions in multidimensional vector space.
One answer is to follow along in some of the same steps that biology has taken over evolutionary time to design brain neural circuits that can accomplish similar kinds of objectives. Considered broadly, brains accomplish this in two steps.
STEP 1 — DESIGN THE BASIC ARCHITECTURE
The basic architecture of the human brain is designed around groups of nuclei and cortical areas that communicate with one another. For example, the eyes connect to the thalamus in the brain which in turn connects with the visual cortex. This basic architecture was designed through trial and error over millions of years of evolutionary time. The information needed to set up this architecture is specified in our genes and the instructions are carried out during prenatal development.
For an ANN the basic architecture will have to be designed and constructed by computer scientists and engineers.
The basic architecture of biological brains allows some behavioral responses to specific situations to be built-in, hard wired. These are called fixed action patterns or instincts. However, most behaviors of humans are not instinctive. Instead, we respond in ways that our brains must learn how to do through experience and learning.
The connections that form the basic architecture of our brains constitute only a tiny percentage of the total connections. The vast majority of the connections in an adult brain were not specified by our genes. There is a fundamental reason for this. There are something in excess of 5 X 10¹⁴ synaptic connections in an adult brain, a number that is orders of magnitude too large to allow each of these connections to have been specified by the genes. So, Where does the information about how to make those connections come from? This is sometimes referred to as the Developmental Problem.
STEP 2 — USE EXPERIENCE IN THE POSTNATAL ENVIRONMENT TO FIGURE OUT HOW THE REST OF THE CONNECTIONS SHOULD GET MADE
The vast majority of the connections in an adult brain were not present at birth. They were formed during an extended postnatal period of developmental neural plasticity. The process involves making an overabundance of new connections and then pruning back some of those that were made, leaving only the “proper” connections.
A mechanism used by biological neurons to determine which connections are the “proper ones” is based on Hebbian Rules, originally articulated by Donald Hebb, a psychobiologist, in the 1940s.
A somewhat non-intuitive implication of having a brain wire itself together by simply allowing the individual neurons to follow Hebbian rules locally, is that the brain will end up with the capability to efficiently detect patterns of sensory stimulation that are correlated with one another. The end result is a brain that is very efficient at picking up the correlated sense impressions caused by objects and events that are present in the environment in which brain development takes place.
Learning to respond to these correlations allows the brain to achieve partial completion. The brain does not have to detect every single feature to know that a certain object is present or that a certain event is taking place. Detection of the correlated features allows the brain to infer what is happening based on incomplete information.
Creating an ANN that functions similarly should be able to be accomplished by employing Hebbian-like procedures to adjust the weights of formal neurons (described in my Part 1 post) while the ANN is given extensive exposure to environments similar to those in which the ANN will need to operate.
Human brains require years of postnatal experience with the environment to wire up its brain connections. ANNs will most likely require similar quantities of training experiences, but the training can presumably be delivered much more rapidly. For example, Wikipedia reports that Tesla has accumulated over a million video clips taken from cameras operating in Tesla cars in the real world. This massive amount of information can probably be loaded into the Tesla Vision Version 12 ANN training model within a matter of hours or days, certainly much shorter than the years human brains need to form their connections.
What has been described so far involving wiring a brain together using Hebbian mechanisms results in what psychologists call perceptual associative learning. This kind of learning is necessary but not sufficient for carrying out a task such as operating an automobile. Associative learning can be used to create attractors (described in my Part 3 Post), but does not specify what response should be performed when a particular attractor state is engaged. A biological neural network, or an artificial one (ANN), needs to know not only what situation is present in the current environment, but also whether the response that will be elicited by the attractor state that has been engaged is the correct one.
Biological brains learn to adjust their responses based on learning from the consequences of those actions. Reward systems in the brain keep track of rewards and punishments and, based on those, increase the likelihoods of responses that resulted in favorable outcomes and eliminate responses that resulted in unfavorable ones.
Similarly, numerous procedures have been developed over the past few decades that allow ANNs to adjust their weights automatically based on supervised training in which the ANN is informed after each response whether the outcome was a favorable or unfavorable one.
The past year has seen unprecedented advancements in Large Language Model (LLM) ANNs such as ChatGPT. It looks promising that the near future will bring about similar impressive improvements in Machine Vision ANNs.
Ronald Boothe, psyrgb@emory.edu
NOTE:
- Ronald G. Boothe. Perception of the Visual Environment. Springer-Verlag New York, 2002.
