Neural networks and genetic algorithms are a type of machine learning that mimics biological systems. Like the name suggests, the way neural networks work is similar to how the nervous system works. Same thing with genetic algorithms, they mimic how genomes evolve over time.
Nodes make up the different layers of a neural network. These nodes act like neurons and accept or process information. The diagram below is an example that you commonly see when talking about neural networks. The hidden layers between the input and output layers are the algorithms that process information.
Let’s say for example that you want to train a machine to recognize pictures of cats.
- You give the machine animal pictures and it tries to figure out which ones are cats and which ones are not.
- You then tell the machine the correct and incorrect ones.
- This is the part where things get cool… or weird.
Depending on whether an image is correct or incorrect, the nodes or neurons in the hidden layers are either strengthened or weakened. This is exactly how your brain strengthens synaptic gaps as you learn. Eventually, the machine figures out how to correctly identify cat pictures.
Genetic algorithms borrow concepts from evolutionary biology. In an ecosystem, you have a population of organisms that reproduce and are subjected to pressures from the environment. With genetic algorithms, this population isn’t a group of meerkats or whatever critter, but solutions or hypotheses.
The machine will test each of these solutions and combine (aka: breed) the most accurate ones or the ones that are the most “fit”. The offspring solutions are subject to crossover, mutations, and drift, just like a population of meerkats. The machine then tests the new generation of solutions and breeds the best ones. And so on and so forth.
What does all this look like? Well, some scientists decided to see what would happen if they asked a machine to build something.
Making Analog Circuits
In a research study from the University of Toronto, scientists used a genetic algorithm to design a circuit board. You can read the whole study, but the conclusion was that,
- the circuit board works
- they couldn’t figure out why
- there were parts that had, seemingly, no function
- if you took out these parts, it broke the circuit board
If this all seems a bit crazy, well, it really isn’t. Artificial Intelligence and machine learning are already fully embedded in our lives. Siri, Alexa, Roomba… not Furby though. Thank goodness. Those things were creepy.