Hello everyone! It’s April, which means we are ready again for a new unit. Let’s dive in.

You know what’s better than one neural network? TWO!!! Multi-modal approaches—making two neural networks interact for a certain result—dominate many of the current edge of neural network research. In this unit, we are going to introduce one such approach, Generative Adversarial Networks (GAN), but leave you with some food for thought for other possibilities for what training multiple networks together can do.

Woof. As I begin to write this I should add that this unit is going to be conceptually dense. Though we are teaching one particular algorithm (incidentally, named, REINFORCE), the world of reinforcement learning is build by one, if not many, very advanced treatments in maths.

So if anything, I would focus on getting the conceptual flavor of how these problems are formulated and discuses. If you can be along for the mathematical and algorithmic journey, then even better — but by no means required or expected… There’s still lots for all of us to learn together.

Welcome back! I think, over the last few days, we have been hyping up convolutional neural networks enough such that you are probably ready to dive right in. So… Let’s, uh, motivate it first!

Why do we use a CNN?

Let’s think of a toy problem to play with. Given a pattern made using two colours (let’s name them a and b, or perhaps black and white), let’s classify whether it is the “zebra” pattern" or the “checkerboard” pattern.

Require: analyze movie + quote [story + bellows]

ineffability of language vs. Sapire-Wolf

Foreignizing Time in Heptopod B

Louise’s ability to re-express her temporally-independent thoughts in English after learning Heptopod B represents a successful act foreignization of Heptopod thought for an English L1 audience despite this audience’s supposed limitations in understanding temporally-independent concepts according to the Sapir-Wholf Hypothesis.

Heptopod B does not have temporality

• RUSSIAN SCIENTIST: “Their final words translate to, “There is no time, many become one.” I fear we have all been given weapons because we answered the timeline wrong, please, if you - -”
• “Explain it by saying that light minimized the time needed to travel to its destination, and one saw the world as the heptapods saw it.”

So it seems like quintessential Sapir-Wholf: time hard to express with Heptopod, and so their way of thinking work around it.

CUTile is when NVIDIA wants to build a cute compiler too!

Takes a single thread, tile based abstraction, and maps it to different hardware; memory ands threads automatically managed.