Neural Networks, Pavlov's Dogs, and Elections

Assorted Flowers

Today's blog post is a bit technical. My intention is to explain some of my thoughts on how people behave unconsciously (and sometimes even consciously). At its heart, this is about group psychology, and applying the ideas of classical conditioning. I really came to this understanding from a different direction, however; one informed more by my experience with computer science and physics than by psychology.

Image by BruceBlaus. Obtained from Wikimedia Commons.

To begin with, I need to describe the humble neuron. Neurons are the basic unit that makes up the brain. They are cells. By themselves, neurons are relatively simple things. The basic feature that makes neurons special is the way they link up and pass signals to one another. A typical neuron is made up of a cell body with an axon and dendrites. The axons and dendrites are basically like tentacles. Typically, the axon of one neuron will connect to one or more dendrites of other neurons. The dendrites are typically thought of as inputs and the axons are typically thought of as outputs.

A neuron's signalling activity is often referred to as firing. It consists of the generation and release of an electric potential between dendrites and axon. When a neuron detects, via its dendrites, that there has been firing on the input side, it can excite the neuron so that it also fires, which will then be detected by any neurons connected to its axon, on the output side, and possibly excite those neurons.

When a neuron detects that an input has fired, and then, shortly thereafter, it detects that an output has fired, the neuron will, thereafter, be more likely to fire off a signal in response to that input. However, if the neuron detects that an input has fired, and the output does not fire, then the neuron will, thereafter, be less likely to fire off a signal in response to that input. It's as if the neurons are trying to be accurate predictors of what the output will do, and signal a prediction before the output even acts.

This kind of frequent firing mechanism is hard and inefficient to model in a typical computer, so usually, each model neuron has some kind of number attached to it, representing its level of excitement. It weights the value of each of its input connections and combines the results in order to determine its own level of excitement. A method called backpropagation is often used to readjust the weights as a result of the calculated success or failure of the overall network of neurons. It is a model that has been extremely successful in generating algorithms that are capable of learning. The basic strength of neural networks is the ability to discover and recognize complex patterns autonomously, or semi-autonomously. It is the reason why human brains are so good at processing images and sounds, and the application of these principles is the reason why computers are getting so much better at processing images and sounds.

One of the things that struck me, early on, about this neuronal behavior, is that it is not terribly different from the way people-sized organisms behave. In the 1800's Pavlov had developed the fundamentals of classical conditioning in his work with dogs. Since people and all other known animals are also subject to the principles of classical conditioning, it seems difficult to avoid the suggestion that neurons are a system that shows scaling symmetry. (By which I mean, that the system looks very much the same when observed from a distance. In this instance, an organism driven by neurons exhibits neuron-like behavior, and then, quite probably, a system composed of such organisms may continue to exhibit neuron-like behavior.) Apart from using this observation to motivate training, which is how classical conditioning is often applied, this suggests that some analogy with neural networks should also be useful for informing our development of social organization.

When Neurons Vote

This brings me to elections. Way back in high school, I had an encounter that I will not soon forget. We had one particular student who was very vocal about political issues. That is, politics, candidates, policy, and elections were, as a whole, something he was particularly passionate about. His political leanings were rather opposite mine. Somehow I ended up briefly speaking with this student on the subject of the presidential election. I think I was probably a Bob Dole supporter, while this student would have supported Bill Clinton. He attacked with the assertion that Bob Dole was going to lose, and I replied by agreeing with him, but that I thought that had little to do with whether he was the better choice. At that the other student seemed dumbfounded. The impression I was left with, (whether it was accurate or not) was that, to him, he voted for the sake of being able to say, or congratulate himself, that he had supported the winner, rather than for the sake of promoting his ideals.

The idea then is, that there are probably many people voting, who vote for the sake of being able to say, or congratulate themselves, that they had supported the winner, rather than for the sake of promoting their ideals. My years of observation since then have not been able to disprove that basic hypothesis, or come up with a significantly better explanation for the accumulated evidence.

Considering, recently, the analogy with neurons, it finally struck me that this irrational-seeming behavior made perfect sense by making the analogy that people were like neurons. We don't have a concrete message, but we feel as if our lives depend on being heard and accepted. So, all too often, we are simply watching, trying to guess what will be accepted, because, we feel deep down somewhere, perhaps unconsciously, that if we can attach to the right figure, and support him before he wins, then we will be relevant and accepted.

In many cases, this kind of thinking makes a lot of sense. Our survival, in a very real way, does depend on our ability to find acceptance and relevance with our peers. To see that people often behave this way, one only needs to look at the social interactions of highschoolers. It is preferable to avoid suggesting to ourselves that we are better than them and aren't fundamentally acting according to the same principles. In the case of elections, however, because our individual relevance, and worth, objectively has nothing to do with the outcome of an election, (unless one happens to be a politician, or otherwise professionally associated with electioneering) the behavior is sadly self-defeating. Specifically, the results of this group behavior are often distinctly contrary to the public and individual good.

An SB54 Analogy

In the wake of SB54, I have been thinking a lot more about how this kind of model might inform elections. I have long been a fan of republicanism, and clearly, this biases me against SB54. My gut tells me that the analogy with neural networks accurately suggests that removing the layer of representation can weaken society's ability to have good constructive elections.

Obtained from Wikipedia.
GNU Free Documentation License

Specifically, in computer science, neural networks are often described in terms of layers, and there are typically three kinds. We have an input layer, an output layer, and one or more hidden layers. In reality it can be much more complicated, but this basic model illustrates the point.

When simple neurons are organized in this way, it is objectively impossible for the network to accurately recognize anything but the simplest of patterns without the presence of hidden layers. Having a mixture of hidden layers and direct input-to output connections can also work.

I'm interested in working out how a neuronal model of elections would work, but I am certain that our history of gutting the power of our electors and delegates, all across the country, has left the people effectively powerless, even though one might naively think that the people are getting more of a voice.

Other Suggestions

Image by Tomwsulcer.
Obtained from Wikimedia Commons.

Analogy with neuronal models also suggests why our government has an entrenched two party system that most people are unhappy with. It shows that in order to get people out to vote, you don't need to be a good candidate, because people, by their nature do not look for the best candidate. In order to get people out to vote, you have to project the image of a winner. It is only possible for two candidates at a time to appear to be viable. The candidates have to get over 50% of the electors, so an effective three-way split would make all candidates look nonviable. Hence, only two parties in an election will ever draw popular support.

This also suggests the importance of a biased media. In order to keep even a two-party system intact under such circumstances, people have to be able to believe that their candidate is likely to win. This is untenable if every news source reports accurately, as, except for in the closest of elections, many people will quickly realize who the winner will be and vote for them (or simply not vote).

Of course, this begs the question: If the media figures are to be modeled as neurons, then why do they willingly fire inaccurately, undermining their own future influence and relevance? By analogy with neural networks, there are many reasons why this might be. It is worth pointing out here that my description of neurons was only a description of typical neurons, and there are deviant types that are poorly understood, (by me, at least) and which can create biases. Biasing neurons are often intentionally added to some designs to achieve certain behaviors. It is also true that some neurons don't receive input from other neurons, but receive input from external influences. (Neurons that detect light, in your eyes, or sound, in your ears. There are temperature sensitive neurons, and heat sensitive neurons, etc.)

To that extent we might suggest that inaccuracy in the media could stem from foreign powers providing perverse incentives, actors with unusual psychology being attracted to the field, feedback loops (networks don't always move in one direction, and hence internal actors may cause bias), or some other influence. Such bias is, however, an interesting phenomenon worthy of attention.

Another point to be made is that neural networks are known to suffer from a problem referred to as local minima. That is, they can sometimes learn the wrong thing and get stuck in the wrong solution to a problem, because any small change will represent a worse solution. (eg. the Nash equilibrium illustrated in A Beautiful Mind) Much work has been done to find ways to avoid local minima, or ensure that local minima are close to the optimal solution. If analogy is sufficiently apt, then such principles might be applied to help make elections a more effective tool for society.