AI: How it works(Part 2)

In a neural network, not only does it contain a bias and a value, but is also has another numerical number of significance: the multiplier(which is also known as weight). The arrows that connect every neuron to every other neuron in the layer in front of it isn't only their for show. These arrows all hold a multiplier, which multiplies the value of the neuron(including the bias of the neuron). These multipliers are there in order to weigh the significance of certain neurons on others. Think of it like this: If you are hungry and craving food, what are some factors that determine what type of food you will get? Some examples would be what you are craving, how hungry you are, how much you can eat, etc. However, each factor isn't equally important when it boils down to your final decision. The type of flavor I am craving(salty, sweet, spicy) is far more important than something like "what is my favorite color." Although these two do contribute in some way to my choice of food, they have a different impact according to how important they are(to me). This is exactly the purpose of the multiplier, carried by the arrows. Some neurons are more likely to fire off(or to have a higher value) if one certain neuron fired off rather than a different one. This is key when it comes to neural networks, as it allows for flexibility and a large range of results when it comes to tweaking the values of the neural network(biases, # of neurons).

Surely, however, these AIs don't just start off the bat with all the precise values for biases and multipliers.

So how do these AIs learn and become so accurate?

When a reinforcement learning-based AI is first initiated, the values of its B's and M's(biases and multipliers/weights) are all random. The AI becomes more accurate over time because it is being "trained" to have higher accuracy. For every reinforcement AI, they will have something called a "training set." The training set will give the AI something to solve, and will compare the answer the AI presents with the correct answer. This helps the AI sharpen, as its answers start to show nearly 0 difference than the correct answers from the training set. The AI is capable of becoming smarter because it has a built-in tendency to optimize accuracy.

However, an AI cannot depend solely on its training set to become smart, as it needs to actually be able to adapt and "grow." The AI starts off with random values for its neurons and its arrows. From its starting accuracy, it begins to randomize every neuron's bias, and every arrow's multiplier(value of a neuron isn't something that is randomized; it is simply the output of previous neurons combined). As the accuracy of the AI increases, the AI begins to randomize less neurons than before, switching what neurons it changes. This way, the AI can realize which neurons are in the biggest need of change, in turn, maximizing accuracy. This process repeats until the accuracy begins to plateau, usually around somewhere is the high 90s(in percentage). This method may be useful, but there are also other strategies to further increase accuracy, such as Backpropagation.

Backpropagation is when the output of the AI and the correct answer of the training set are subtracted together, giving us the "Error" of the AI. The error can also be interpreted as how different the correct and given answer is. By getting this error after passing it through the neural network once, we can essentially reverse this process by sending the error of our result back from the output of the neural network to the input. We do this to minimize the value of the error while we send it back through the neural network, which will help us increase accuracy even further by ensuring AI doesn't repeat the same mistake over and over again.

AI can be upgraded through a wide range of methods, from methods covered in this blog, such as backpropagation, to other methods that I haven't mentioned. AI has so many unique and special interactions available according to the methods that any given person uses, whether or not it may be effective. AI should be explored upon further, as in the future, I believe that there could be a extremely diverse pool of different AI types.