Best LSTM explanation I have watched! All your videos are superb! I want to watch them all from beginning to end! Thank you for such detailed and intuitive explanations! :D
At 19:31, he mentioned how many units of LSTM , the units parameters is not for how many units of LSTM in any layer, it is for hidden state dimension. And for how many LSTM depends on input shape[0].
So if I understand well, if we consider the input to be a sequence of x elements, each "LSTM" unit contains x states, and returns a list of x vectors passed to the LSTM units of the next hidden layer. Am I right ?
@@Droiduxx yes, but consider return_sequence, and return_stae arguments also, their default values false , to see the full picture, kindly turn on return sequence. Example - x = tf.range(60) x = tf.reshape(x,(5,3,2)) # shape - ( batch, time, num-features) lstm = tf.Keras.Layes.LSTM( 7, return_sequence= True) Output = lstm(x) Print(Output.shape) # answer (5,3,7)
Amazing tutorial! I got a question: At 14:59 you explain the forget gate. In the lower-left corner, the cell gets ht-1 (last timestep) as input. Is it possible to have a sequence of past days as input? For example ht-1 & ht-2 & ht-3 ... etc. to spot potential trends in the data. Maybe with multiple variables. Giving every single timestep an additional weight.
11:40 What is going on with the arrows? Signal from previous cell merges with current Xt, but there is no operator. Signal from left and signal from bottom Xt. And they both go to 3 gates? Edit: ok I see, its explained later
Thank you very much. No amount of money is little. Every penny counts :) Bulk of the money goes to charities that help with cancer research and eye surgeries for poor people. So the society benefits from any amount that is contributed. Thanks again.
Nice video, so well explained and not too long, along with a full tutorial. Probably one of the best ones about LSTM. Thanks and please keep up the good work! Greetings from France!
I was struggling to understand the basic concept of LSTM and watched dozen of videos and finally found the best one so far. Thank you so much for letting us understand. Greetings from GIST!
Hi DigitalSreeni...I am a PhD candidate investigating applications of MLPs, CNNs and LSTMs. I see that you have amazing graphics for these model types in your videos. Would you be willing to share these graphics for the model architectures with me so that I may use them in my dissertation and defense presentation? I certainly would give you credit for them. Thank you for your time!
Thank you, honestly it s very clear. Please I am looking for a tutorial on image classification but using local images dataset. Have y made a one before. Thank you again
Thanks for your videos! It's really helpful. I have a small question. Could you explain a little more about the meaning of units? Is it mean the number of hidden layers or the number of neurons in a layer?
LSTM is primarily used for processing sequential data. While it is possible to use LSTM for image classification tasks, it is generally not the best choice as it is designed to model sequential dependencies in data, whereas images are inherently spatial and do not have an obvious sequential structure. Images are typically processed using CNNs, which are specifically designed to handle spatial data and can effectively extract features from images using convolutions.
I'm using RNN for my PG thesis work. I've a query. Do we have to run stationarity test for our time series data before feeding it in the neural network model... or this step is only required in traditional time series models like ARIMA?
RNNs are capable of learning nonlinearities (compared to ARIMA) and therefore should be able to learn from the input data without doing any stationarity pre-processing. This is especially true if you use LSTMs. Also, please note that you need lot more training data for RNNs compared to ARIMA. You may find this blog useful to understand the effectiveness of RNNs: karpathy.github.io/2015/05/21/rnn-effectiveness/
I've watched dozen of videos on LSTM and this is the best one so far. Thank you so much sir. Greetings from UCLA!
Glad it was helpful!
The first youtube tutorial I saw which explains a LSTM in detail, e.g. why a Sigmoid or why a tanh is used within the cell. Great!
Can't believe that this is free. Thanks a lot. You are building a community of future researchers and innovators here!
My pleasure!
Best LSTM explanation I have watched! All your videos are superb! I want to watch them all from beginning to end! Thank you for such detailed and intuitive explanations! :D
I get valuable Understanding. I realy appriciate the way of your explanation.
amazing work, thank you so much!
Best explanation out there, i understood, what is happening both conceptually and mathematically
Thank you very much for this video sir!
Thank you very much! It is well explained!
Very intuitive video!
Great presentation sir! thank you so much!
One of the best explanation ever on LSTM! Greetings from Politecnico di Milano!
Thank you Sir, Nice explanations.
Thank you, it is really helpful
You’re welcome.
At 19:31, he mentioned how many units of LSTM , the units parameters is not for how many units of LSTM in any layer, it is for hidden state dimension.
And for how many LSTM depends on input shape[0].
So if I understand well, if we consider the input to be a sequence of x elements, each "LSTM" unit contains x states, and returns a list of x vectors passed to the LSTM units of the next hidden layer. Am I right ?
@@Droiduxx yes, but consider return_sequence, and return_stae arguments also, their default values false , to see the full picture, kindly turn on return sequence.
Example -
x = tf.range(60)
x = tf.reshape(x,(5,3,2))
# shape - ( batch, time, num-features)
lstm = tf.Keras.Layes.LSTM( 7, return_sequence= True)
Output = lstm(x)
Print(Output.shape)
# answer (5,3,7)
Great work sir. keep on doing great job
Amazing tutorial! I got a question:
At 14:59 you explain the forget gate.
In the lower-left corner, the cell gets ht-1 (last timestep) as input. Is it possible to have a sequence of past days as input?
For example ht-1 & ht-2 & ht-3 ... etc. to spot potential trends in the data. Maybe with multiple variables. Giving every single timestep an additional weight.
11:40 What is going on with the arrows? Signal from previous cell merges with current Xt, but there is no operator. Signal from left and signal from bottom Xt. And they both go to 3 gates?
Edit: ok I see, its explained later
I can´t help but find this channel incredibly undersubscribed!!!
I’m glad you like the content. I rely on you guys to spread the word :)
Thanks!
I know this little amount of money is not enough to say thank you. Keep the good works ser, 🥰
Thank you very much. No amount of money is little. Every penny counts :)
Bulk of the money goes to charities that help with cancer research and eye surgeries for poor people. So the society benefits from any amount that is contributed. Thanks again.
Nice video, so well explained and not too long, along with a full tutorial. Probably one of the best ones about LSTM. Thanks and please keep up the good work! Greetings from France!
Hi, well explained! Could I have your slides?
Why is there a dropout after the final LSTM layer?
I was struggling to understand the basic concept of LSTM and watched dozen of videos and finally found the best one so far. Thank you so much for letting us understand. Greetings from GIST!
Great to hear!
Awesome! Thanks sir.
Good, thanks a lot.
You are welcome
great. thx a lot
Hi DigitalSreeni...I am a PhD candidate investigating applications of MLPs, CNNs and LSTMs. I see that you have amazing graphics for these model types in your videos.
Would you be willing to share these graphics for the model architectures with me so that I may use them in my dissertation and defense presentation? I certainly would give you credit for them.
Thank you for your time!
Thank you for the video.
I have a question.
The number of units (50) is the number of the so called "hidden units", also known as "hidden size"?
Thank you so much :)
Subscribed after watching your first video.
Can you teach us how to use LSTM and ARIMA in ensemble learning in forecasting time series data?
Thank you, honestly it s very clear.
Please I am looking for a tutorial on image classification but using local images dataset.
Have y made a one before.
Thank you again
Great explanation! Thank you so much!! : )
thank you, nice video for LSTM new learners :)
谢谢老师
really, thank you for your more clarification!
i love your video...i am just starting to learn machine learning and its very useful'
Thanks for your videos! It's really helpful. I have a small question. Could you explain a little more about the meaning of units? Is it mean the number of hidden layers or the number of neurons in a layer?
May be this helps... stats.stackexchange.com/questions/241985/understanding-lstm-units-vs-cells
@@DigitalSreeni Thanks a lot! It's very helpful.
Dear Dr. S. Sreeni,
Thanku for your informational videos regarding cnn.
Kindly make LSTM for image classification tasks.
Thanku.
LSTM is primarily used for processing sequential data. While it is possible to use LSTM for image classification tasks, it is generally not the best choice as it is designed to model sequential dependencies in data, whereas images are inherently spatial and do not have an obvious sequential structure. Images are typically processed using CNNs, which are specifically designed to handle spatial data and can effectively extract features from images using convolutions.
Thank you so much for this video...
Nice tutorial! Thank you!
Hi sir. thank you for much for all your videos. Could you provide us with tutorial to implement LSTM & RNN with Python Please?
Yes... they should be out this week.
I've viewed several vids on LSTM but this breakdown is the best!!
I'm using RNN for my PG thesis work. I've a query. Do we have to run stationarity test for our time series data before feeding it in the neural network model... or this step is only required in traditional time series models like ARIMA?
RNNs are capable of learning nonlinearities (compared to ARIMA) and therefore should be able to learn from the input data without doing any stationarity pre-processing. This is especially true if you use LSTMs. Also, please note that you need lot more training data for RNNs compared to ARIMA. You may find this blog useful to understand the effectiveness of RNNs: karpathy.github.io/2015/05/21/rnn-effectiveness/
Best teacher ever.
Thanks
Sir you are a gem!
thanks!
Nice Explanation Sir!
I feel very gifted that I got the suggestion from RUclips, the right video....
I am glad you found my channel :)
I've watched many videos and read a lot about LSTM but this is the first time i really understand how LSTM works. Thumbs up thank you!
Great to hear!
awesome explanation thank you very much
Glad it was helpful!
Amazing Sir.
Lol ever heard of transformers???
Now sure what your meant by your comment, was that a question?
please make a video about attention in images
I got your attention :)
We are infinitely grateful
Thank you :)
First like a video then watch it !
Thanks for your blind confidence in the video, I hope your opinion doesn’t change after watching the video :)
nice explanation!
Thanks! 😃
I am so happy to discover this channel! :)
His continuing use of "ok?" "ok?" "ok?" "ok?" is incredibly annoying.
And you are not annoying at all.
Poor choice to comment on personal trait rather than content of the tutorial, ok?