Цитаты из книги Грокаем глубокое обучение

“The only relevant test of the validity of a hypothesis is comparison of its predictions with experience.”Milton Friedman, Essays in Positive Economics (University of Chicago Press, 1953)

Learning is just reducing error

how do machines learn?

I’m very excited that you’re here! You should be, too!

I’m also grateful to Mat and Niko at Udacity, who included the book in Udacity’s Deep Learning Nanodegree,

There are three gates—f, i, o—and a cell-update vector u; think of these as forget, input, output, and update, respectively.

f is the “forget” gate.

embeddings, which are vectors mapped to words that you can forward propagate into a neural network. Thus, if you have a vocabulary of 200 words, you’ll also have 200 embeddings.

Much of the research into neural networks is about finding new architectures that can find correlation faster and generalize better to unseen data.

But the different configurations of weights and layers between the input and output layers have a strong impact on whether the network is successful in finding correlation (and/or how fast it finds correlation).The particular configuration of layers and weights in a neural network is called its architecture

Why does it turn out this way?

I’m going to do what I typically do in real life (cuz I’m lazy—I mean, efficient):

So, it turns out that learning all the methods for taking any two variables in any function and computing their relationship takes about three semesters of college. Truth be told, if you went through all three semesters so that you could learn how to do deep learning, you’d use only a very small subset of what you learned.

Notice that you always have the derivative between two variables. You’re always looking to know how one variable moves when you change another one.

Learning is all about automatically changing the prediction function so that it makes good predictions—aka, so that the subsequent error goes down to 0.

That wouldn’t do.

This is silly, but totally doable.

forward propagation, wherein a neural network takes input data and makes a prediction. It’s called this because you’re propagating activations forward through the network.