1 x 1 Convolution (in Hindi)

[@Ajeet-Yadav-IIITD]

3:42 Shouldn't be n-f+1 ?

[@KshitizVermaDL]

Thank you for pointing it out! It should be n-f+1.

[@Omunamantech]

@@KshitizVermaDL if it is n-f -1 then it will be n - 1 - 1 = n - 2 = 26, am I right?

[@rajshaikh3505]

This channel is extremely underrated. All explanations are really really amazing, cant emphasize enough. Please keep making more such content. Thank you so much

[@KshitizVermaDL]

Thank you so much for the kind words!

[@samirz7]

I can't believe I didn't discover this channel earlier. Simple explanation of difficult concepts. Thank you. Learning in hindi really clears the concept on another level :)

[@KshitizVermaDL]

Thank you for this awesome comment! Such comments make me keep going!

[@somdubey5436]

very helpful video. I feel like you are gifted in making people easily understand complex things.

[@harshjain4256]

If we use 1x1 filter how it will detect features? For example edges?

[@raj-nq8ke]

Simply speaking, 1X1 filter reduces the size of channels

[@KshitizVermaDL]

You can use it to do so.

[@kazimali4273]

please tell me if we do not use relu function then whtas happen. I mean why we break linearity.

[@KshitizVermaDL]

There are two videos explaining this! Have a look at them

[@kazimali4273]

@@KshitizVermaDL thanks I will reply after seeing your videos

[@fierce10]

If you keep layering linear functions, result will also be linear only, then all layers can be represented as single layer. But we are using layers to get a non-linear prediction. We are only using linearity within a layer because it is easy to compute transformations. But we break linearity between each layer.

[@atiffaridi507]

jab koi image of size height x width x channel , single filter (eg. 3x3x3) ke sath convolve hoga to humein height-1 x width-1 x 1 image milega. Agar 16 filter ke sath convolve karenge to height-1 x width-1 x 16 size ka output milega. Plz correct at time: 2.41

[@KshitizVermaDL]

I tried to check, couldn't find anything wrong! Can you check your calculations with a 1x1 filter?

[@gender121]

A very good lecture but a slight doubt 28x28x192 * 5x5x192 (#32)=28x28x32 why not 28-5/1+1=23*23*32.Please explain Sir.Whether different padding has been used to make it uniform?

[@KshitizVermaDL]

Thanks for the comment! You are right. Padding has been used. I have defined the meaning of "same" padding in one of the earlier videos. It means you do padding so as to keep the output dimensions same as input.

[@eiesabyasachi]

@@KshitizVermaDL Sir, As per my understanding, can be it calculated like this In case of normal scenario, 28x28x192 * 5x5x192 ( #filters = 32 ) = (28-5)/1 + 1 = 24x24x32 No of operations will be ( 28x28x192 ) * ( 5x5x32 ) = 120 M In case of 1:1 convolution scenario, 28x28x192 * 1x1x16 = 28x28x16 * 5x5x16 ( #filters = 32 ) = (28-5)/1 + 1 = 24x24x32 No of operations will be ( 28x28x192 ) * ( 1x1x16 ) = 2.4M No of operations will be ( 28x28x16 ) * ( 5x5x32 ) = 10M All total it will be 12.4M Please correct me, if the calculation is wrong.

[@manishswami877]

thank you sir