What are Algorithms?

AI systems can’t really function without them. They reach out across just about every part of our lives.

The importance of Algorithms can’t be overstated and yet, it seems like for many people who are new to the AI space, understand what they are isn’t so easy. With this primer, we hope to help you get there.

Consider recipes. Yes, recipes. Seemingly, the most popular way to describe what algorithms are to someone who has little to no prior knowledge of AI is to use the analogy that an algorithm is like a recipe.

This is because an algorithm is a set of instructions that lead to some sort of finished product that represents a sum of those instructions. Think about what you do to make a cake. You input eggs, flour, sugar, and all of the other ingredients, mix them, pour them into a pan, stick it in the oven at a certain temperature, and wait for a specific amount of time for the ingredients to come out as the finished product.

In a basic sense, AI algorithms are no different, though the more precise the instructions are that are fed to the system, the better. Consider the backpropagation algorithm, which is arguably the most important algorithm to most AIs in existence(at least those that are structured to benefit from Deep Learning). Without getting too lost in the weeds, inputting a backpropagation algorithm into an AI system involves feeding it a set of instructions related to how it should process any data that it takes in at any time.

More specifically, this means that the set of rules that make up a backpropagation algorithm tell it to consistently take “already-processed” data back through an AI system’s neural nets, in order to train it on how to perform better. On a more technical level, this means that the “weights” in the “hidden layer” of a neural net are adjusted to allow it do so.

If you’re not exactly familiar with what these weights are, they represent the “neurons” that make up the middle layer of a neural net. Their function is essentially to take the raw data that comes in from the first layer and assign numerical values to it so that the entire AI system can understand it on a quantitative level. For those of you who have had any experience in the blockchain industry, especially with crypto mining, you can think of adjusting the weights of a neural net like adjusting the numbers you put in to achieve the hash(range of numbers) that are ideal for a crypto transaction to be verified.

Adjusting the weights means adjusting the values that are input into the final group of equations that the AI computes before taking whatever action it is designed to take. Again, like miners in a crypto network, an AI needs to play around with the weights in its’ neural nets and therefore, test out what numerical values help it achieve its’ ideal outcome.

It’s not an easy process and by all accounts, training an AI system takes a large amount of time and effort. Because of this, it’s no accident that AI teams often start their systems out in “black-boxes,” which are closed environments in which the team can tweak the system as much as possible before letting it out into the wild.

Put this together with backpropagation algorithms and you have an effective equation for achieving the “deepest” possible learning, at least for the foreseeable future. Once AI is able to truly stand on its’ own and we can say that we have reached the time of “Artificial General Intelligence,” then things may change, even at the structural level.

Supervised Learning likely won’t hold a dominant sway over the industry forever. Something has to give eventually. It’s just a question of when and whether we’ll be ready with the right tools to deal with autonomous AIs.

In future posts, I’ll dig more into all of these topics including black boxes and their importance as well as backpropagation. For now, suffice it to say that algorithms are the fuel for AI systems and without them, there arguably wouldn’t be AI at all. Until our next post, ask yourself: what recipe would you like to see fulfilled in future AI systems?