NN Part 2 is now public!

Great work

Megan Britt

Any other optimization methods used? Simplex? Data dependent?

Hoping for this Friday...we'll see.

3blue1brown

HI! Fantastic videos - thank you so much! I am currently diving into data science and your videos really help me to understand the math behind. I am not a natural within the math. field. :) When will the next video about neural networks come out? Best regards, N.

I love your videos, guys! And it inspire me.... I’d like to work more my math... But now I am reviewing a Scrip theater play to a Theater in São Paulo... 😊 It’s difficult to conciliate both... But since I watch your video I could see that is possible to do both 🙂👍

I wanted to say, that I did something very similar at work intuitively with a much simpler problem. Maybe you could use this as an example in another video? We have a virtual environment that allows the user to chose a background color giving us 3 values (only 3 dimensions) for RGB. Based on the user's selection we wanted to automatically decide if the text on the page should be either black or white (1 output). So we found various values where the preference changed from black to white in order to develop a formula. We developed a cost function and adjusted our variables manually. In effect I got a very hands on approach at manually developing a very simple neural network.

Benjamin BairMoshe

Please consider doing a video on transforms like Laplace and Fourier.

That might be one valid way to try to express a certain output belief, but I'd imagine that just turning off all the output neurons ought to express the same belief just as effectively, right?

Jesse Thompson

Gavin - I'm currently working on a related problem, namely, application of ANNs to intrusion detection. In this case, the NN recognizes "normal" patterns, but you want it to alert on "non-normal" patterns. Thus, it has to generate a confidence in it's categorization and generate metadata when the confidence is sufficently low.

Phil Moyer

Wish this and the previous NN video were 60fps.

Alexey Badalov

My first thought during the discussion of what happens when a random input is put into the system was "Why not add an 11th 'None of the Above' output state". The part about what happens when you train a network with mislabeled data made a good counterpoint, though: the network would be identifying the specific "not a digit" test data, not the more abstract concept of "not a digit" itself.

Gavin Olson

Just amazing videos... Thank you!