Need of Batch Normalization || Lesson 18 || Deep Learning || Learning Monkey ||

#deeplearning#learningmonkey#neuralnetwork
In this class, we discuss the need for batch normalization.
In our neural network, we use batch gradient descent.
We send the data into our network in batches.
As the distribution of batch changes a little.
The data moved in layers and multiplied with different values.
The deep layers see a lot of change in data in each batch.
This we call it internal covariance shift.
Because of this covariance shift training, the data takes time.
To normalize the data at each layer so that it will see each time from the same distribution.
Weights are updated fast. That's the use of batch normalization.
Link for playlists:
/ @learningmonkey
Link for our website: learningmonkey.in
Follow us on Facebook @ / learningmonkey
Follow us on Instagram @ / learningmonkey1
Follow us on Twitter @ / _learningmonkey
Mail us @ learningmonkey01@gmail.com

Пікірлер: 15

  • @mouleshm210
    @mouleshm2103 жыл бұрын

    Wow...its very clear for me now👍👍👍👍

  • @makeshs9099
    @makeshs90993 жыл бұрын

    Awesome work brother!! Very clear and precise.

  • @LearningMonkey

    @LearningMonkey

    3 жыл бұрын

    Thanks Makesh

  • @farha5595
    @farha55952 жыл бұрын

    clear explanation 👍🏻

  • @shivamkumar-rn2ve
    @shivamkumar-rn2ve2 жыл бұрын

    Awesome cleared nicely

  • @LearningMonkey

    @LearningMonkey

    2 жыл бұрын

    Thank you

  • @datascienceandaiconcepts5435
    @datascienceandaiconcepts54353 жыл бұрын

    nice work

  • @LearningMonkey

    @LearningMonkey

    3 жыл бұрын

    Thanks

  • @debjyotibanerjee7750
    @debjyotibanerjee77502 жыл бұрын

    This is good explanation, bit I have 1 question, suppose if we keep our activation function in the hidden layers also as sigmoid function, then also distribution remains same, as the value will be in the range 0-1. So why not use sigmoid or tanh in the hidden layers to keep the distribution same?

  • @LearningMonkey

    @LearningMonkey

    2 жыл бұрын

    Keeping value in between 0-1 is not same as moving distribution between 0-1

  • @pratikneupane951

    @pratikneupane951

    8 күн бұрын

    Yes you are right saying the range of values given by Sigmoid to another layer is between 0 and 1.But the problem with Sigmoid is that as the updated weights moves away from the origin on both sides,the gradient vanishing problem arises. But in batch normalization ,we are trying to get the data points with mean and standard deviation close to 0 and 1 respectively.This doesn't mean that the range of values has to be in between 0 and 1 .Also the learnable parameters beta and gamma performs scale and shift to preserve the underlying magnitude of the inputs. (I am unsure if this is completely correct or not !).

  • @gsaidheeraj2229
    @gsaidheeraj22292 жыл бұрын

    Bro, I'm not sure what you mean but I heard "Normalization means Standardization" couple of times(time stamp: 4:05, 4:10). Normalisation is totally different from standardisation. Standardization means data is converted to standard distribution where values ranges from -inf to inf, where as in normalization data is converted to range of 0-1. Please try to correct it, because if we say same thing in interview, they'll directly disqualify in technical round.

  • @LearningMonkey

    @LearningMonkey

    2 жыл бұрын

    Here we use the term normalization. Batch normalization. But actually we are doing standardization. I explained not to get confused with the title.

  • @gsaidheeraj2229

    @gsaidheeraj2229

    2 жыл бұрын

    @@LearningMonkey okay.. got it. So you mean "in batch normalisation we srandardise the data" I agree with it...👍 thank you for clarifying.🙏

  • @surajrao9729

    @surajrao9729

    11 ай бұрын

    dayanand college? Is it the same dheeraj i am assuming?

Келесі