AI as Infant: The Layman’s Guide to Neural Networks

As AI becomes more ubiquitous, many of us are scrambling to learn to code and master mathematics. Your colleagues may have started talking about potential AI initiatives. Maybe, late at night, your vague technological FOMO leaves you scouring wikipedia.

But AI are made in our image. So rather than looking outward, we can understand them best by looking at ourselves.

With that in mind, let me introduce you to the most exciting form of AI: the neural network.

The apt NN/baby comparison isn’t coincidental. As the name suggests, NNs are structured like human brains. And like us, they start out as blank slates, gaining understanding as they’re exposed to more data.

Given the fundamental similarities, they also share many quirks, strengths, and downfalls with babies. For example, they’re….

If your baby has decided to shriek, you probably don’t know why. You may ask, but it’s not going to get you anywhere.

Similarly, even compared to other algorithms, NNs can’t explain their reasoning. You just have to look at the output, look at the input, and guess why it’s doing what it’s doing. This is known as the black box problem.

If a baby has only ever seen dogs, when she sees her first cat, she’s likely to call it a “doggie.”

NNs often make similar mistakes when their datasets aren’t carefully designed. Be sure to include edge cases — hairless cats, cats with 3 legs, etc.

Kids are always tugging down lamps, scattering Tupperware, and picking up curse words.

NNs can be mischievous too. Once in a while, when they’re supposed to practice English, they invent a new language. When you want them to transform and revert images, they cheat by encoding the original into the elaborations.

Humans aren’t built for reading. Nothing in our neurology or our evolutionary past suggests we should be able to do it. But kids learn it anyway.

Similarly, NNs are also able to process all forms of data — images, numbers, video, text, and audio — instead of specializing in just one or two of them.

NNs require larger datasets and more computing power than other models. So if another algorithm will get the job done, don’t build a NN. Be like your aunt who skipped kids to travel the world.