Comments on: Understanding Convolution in Deep Learning

By: Tim Dettmers

Tim Dettmers — Mon, 29 Nov 2021 03:17:25 +0000

In reply to bruce. Thanks for letting me know! I did not realize the images were broken!

By: bruce

bruce — Sat, 20 Nov 2021 12:39:55 +0000

hey , can please update your article , plots and equations are missing

By: MJ

MJ — Wed, 27 Mar 2019 23:51:26 +0000

Just recently found this gold of an article. Concepts are explained excellently. Very nice read. Thanks!

By: Tim Dettmers

Tim Dettmers — Fri, 12 Oct 2018 16:48:10 +0000

In reply to Shuvaday Chatterjee. Filters can have both negative and positive weights. Quick changes from very negative to very positive values are useful for detecting edges because usually two different objects have different colors and this difference is amplified by patterns of positive and negative weights.

By: Shuvaday Chatterjee

Shuvaday Chatterjee — Thu, 11 Oct 2018 14:43:27 +0000

Hello Tim

Thank you for the wonderful post.
I have one question.Could you please help me to understand why the filter has negative weights?

Thank You
Shuvaday Chatterjee

By: Tomasa Finely

Tomasa Finely — Wed, 26 Sep 2018 22:51:26 +0000

Thanks for sharing your thoughts abouut foods. Regards

By: phim con heo

phim con heo — Sun, 02 Sep 2018 09:09:03 +0000

I am genuinely glad to glance at this website posts which contains plenty of helpful information, thanks for providing these kinds of information.

By: peyman

peyman — Sat, 02 Dec 2017 06:33:00 +0000

I am a graduate student in computer network engineering and study in Iran
And I’m investigating cryptography and decryption with the convolution of the code
I need an algorithm to capture and hide the data and submit an image file, and eventually send the data hidden behind the photo to be detected in the receiver.
please guide me

By: Yun Teng

Yun Teng — Sat, 30 Sep 2017 06:56:28 +0000

In reply to Tim Dettmers. Thanks! I will look into Winograd.

By: Tim Dettmers

Tim Dettmers — Fri, 29 Sep 2017 12:01:03 +0000

In reply to Yun Teng. For smaller kernel sizes FFT is not as fast (or about as fast) as normal convolution. However, if you use Winograd FFT then you can improve the performance further for 3x3 convolution. Currently, it is the fastest algorithm and is used in most software libaries as the default (sometimes others are used, as Winograd FFT consumes a bit more memory). Also consider, that the complexity is not a very relevant measure if the variable have small sizes. If you process small kernels and images then performance depends mainly on the hardware, for example, it will depend on cache size, and access speed between different levels of memory and so forth.