Thursday, November 14, 2024
Latest:
Videos

Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM)

Artis Modus



Brandon Rohrer

Part of the End-to-End Machine Learning School course library at http://e2eml.school

Find the rest of the How Neural Networks Work video series in this free online course:
https://end-to-end-machine-learning.teachable.com/p/how-deep-neural-networks-work

A gentle walk through how they work and how they are useful.

Some other helpful resources:
RNN and LSTM slides: http://bit.ly/2sO00ZC
Luis Serrano’s Friendly Intro to RNNs: https://youtu.be/UNmqTiOnRfg
How neural networks work video: https://youtu.be/ILsA4nyG7I0
Chris Olah’s tutorial: http://bit.ly/2seO9VI
Andrej Karpathy’s blog post: http://bit.ly/1K610Ie
Andrej Karpathy’s RNN code: http://bit.ly/1TNCiT9
Andrej Karpathy’s CS231n lecture: http://bit.ly/2tijgQ9
DeepLearning4J tutorial: https://deeplearning4j.org/lstm
RNN/LSTM blog post: https://brohrer.github.io/how_rnn_lstm_work.html
Data Science and Robots blog: https://brohrer.github.io/blog.html

Follow me for announcements: https://twitter.com/_brohrer_

Source

Similar Posts

31 thoughts on “Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM)”

  12. 4:30 "it's Tuesday" -> Takes some Nasal Spray? 😀 Anyone else noticed?
    Great video, btw. Still watching but you'll probably end up saving my ass <3

  13. This was fantastic, thank you so much for taking the effort to do all of this, it has really helped me get started.

  15. What type of LSTM is he explaining?! It seems somehow different from the standard version of LSTM where there are two bus of information. It's something like GRU. The explanation of LSTM was more simpler, but it was also good.

  18. As a complete newbie to NN, your explanation of both RNN and LSTM was amazing! However, a couple of question are still unsolved:

    1. What does the ignoring gate do concretely? I suppose that, if the initial NN predicted 'Jane'/'Spot'/'.' the ignoring NN would stop the '.' from going through. But if the initial NN is trained like in the example of the video, the ignoring gate would be useless all the time…

    2. I suppose that the selection gate did nothing in the video because the prediction had only 2 elements. Its effect would be noticeable if there were 100 elements and the selection gate would only let through ~ 10 of them?

    3. This may be a dumb question, but… If all NN receive the same input (previous prediction + new input), the only way to get different results would be the initial random state, right? How can they play their role, if the inputs and the NN type are the same?

    4. Finally, I checked Chris Olah's tutorial and I saw that the diagrams differ. In particular, the names differ: going by chronological order, your "ignoring" gate would be Chris' "forgetting" one, your "forgetting" would be Chris' "input" and finally your "selection" would be Chris' "output". Is it right? Also, in your diagram the prediction that will become the output is squished by the tanh function and gated, while Chris' is untouched (tanh and gate are only present in what is feeded in the next iteration of the LSTM).

    Hope you can clarify some of my doubts, and thank you again for this video!

  24. I'm preparing my university exam about "Data Science and machine learning", these videos are pure gold for someone who is approaching this topics for the first time like me. Thank you so much, it was really worth the time spent to watch it.

  26. Dwight Schrute imposter dictionary: {Bears, Beets, Battlestar Galactica, .}

    Mistakes a Dwight Schrute imposter RNN can make:

    "Bears. Beets. Battlestar Galactica."

  29. Brandon has applied extraordinary skills in communicating the difficult (convoluted?) topic and concepts on LSTM, in simple and comprehensible language. The example on writing the "Children Text Book" brings out all of the major processes into sharp focus. I have learnt immensely. I will now be studying other videos by Brandon. Thanks a lot.

Comments are closed.

WP2Social Auto Publish Powered By : XYZScripts.com