Lore
Robots Weekly ?: A Glossary
The world of Artificial Intelligence/Machine Learning/Deep Learning is full of technical terms and can get confusing fast. It also uses some familiar terms in completely unfamiliar ways. Here’s a quick overview of a few of them. Let’s get confused!
Adam ?
You might have a friend, uncle, cousin, or frenemy named Adam. This isn’t them. Unless their name also stands for Adaptive Moment Estimation, in which case this may be them. Also, weird.
Adam is an optimization algorithm for deep learning models, a.k.a. the thing that helps optimize your model’s performance. Part of the secret sauce if you will.
If you’re interested in a more technical dive into Adam, you can check out this post at Machine Learning Mastery.
CNN ?
No, CNN hasn’t pivoted away from being a news network to become an artificial intelligence entity. I don’t think…
This CNN stands for Convolutional Neural Network. A CNN is a type of neural network that is typically used for image classification, so it’s hot like Hansel (so hot right now). CNNs are also used for image and video recognition, recommender systems, and natural language processing. So, what exactly are they used for? One use might be helping self-driving cars recognize pedestrians. You can dive deep over at Wikipedia.
If control of CNN were handed over to a CNN, would we know?
Dropout ?
This is not the term for when you get so overwhelmed by AI stories and terms that you “dropout” of the conversation. Dropout in deep learning is a form of “regularization” where nodes/neurons within a neural network are turned off at random during training steps. So they “drop out” of the network temporarily.
Why do this? It helps prevent a neural network from specifying too much to the training set (called “overfitting“) by reducing the co-dependency of neurons on each other. This is a very real problem in deep learning. One example of overfitting is a model that was trained to differentiate wolves from huskies, or so they thought. What it really learned is that snow = wolf since all the training images had wolves in snowy landscapes and huskies in non-snowy landscapes. Here’s another fun example with tanks.
Amar has a nice dropout overview on Medium.
RMS ⚕
This does not stand for the Royal Medical Society. Or the Ramanujan Mathematical Society. Or the Russian Mineralogical Society. It stands for Root Mean Square propagation. Fun!
It’s another optimization algorithm like our buddy Adam. Except this one was proposed by our boy Geoff Hinton (remember him?).
Here’s an overview of RMSprop and a bunch of other optimization algorithms, if you’re feeling nerdy.
Supervised Learning ??
This is when all the little neural networks go to school and the teacher watches over them as the learn. Ok, not really, but how adorable were the neural nets you imagined?
Supervised learning is one of the three main buckets of machine learning. Supervised learning requires a bunch of labeled data, using one of the earlier examples this would be an image labeled “wolf”. These labels are used to train the network that a given x equals the label y so that it can predict y in the future when only give x. This is the most common type of machine learning at this point.
And here’s your Wikipedia link.
YOLO ?
This is exactly what you remember people yelling before crossing the street, “You only look once!” Am I remembering that incorrectly?
I was going to come up with some witty write-up but then got distracted by this video from Joseph Redmon, so go read what he said while I get my Friday dance party on.
TL;DR: “YOLO is a state-of-the-art, real-time object detection system”