Week 11 Discussion Thread

This post contains all comments and discussions during our FastBook Week 11 session on convolutions.


Blog posts from last week

Links from this week


Hi folks! Welcome (back) to Week 11 :smiling_face_with_three_hearts:


I am going to make sure that I write 20 characters. TEST.


This blog article wasn’t completely tied to fastbook, but it was spawned from the Normalization discussion from last week: Normalizing fastai Tabular Inputs | The Problem Solvers Guild


Hi everyone,

Here we are having explicit ratings on a scale of 1-5 or 1-10;

How can we build collab_learner or a recommendation system where we don’t have explicit_ratings but some implicit feedback like number of likes or number of cart adds or number of purchases etc. on a retail website?


Post by Jay Alammar on Embeddings: The Illustrated Word2vec – Jay Alammar – Visualizing machine learning one concept at a time.

Kevin’s Blog: https://blog.problemsolversguild.com/

Convolution Blog: Image Kernels explained visually


Is this collaborative filtering we are discussing?

This was a great read, Thanks for sharing! :slight_smile:

Edit: Please feel free to also share in #projects-resources :tea:

I would probably handle that differently depending on what the distribution of your number of likes looked like. If a lot of posts have 0 likes, but some of them have a large number, you may consider using a scaled version. Another option would be to put them into buckets so bucket 0 would equal 0 likes, bucket 1 would equal 1-10 etc. You could create that rank yourself since it’s just an arbitrary number.


I think using a sigmoid to convert this implicit feedback into some range could be a good rating measure


Yes, I agree. If we have multiple fields though like we want to take into account cart adds, purchases, clicks, likes etc. how can we combine them into one representative rating?

Should it then be a weighted sum of the sigmoids applied to individual features? The main challenge here is sparsity would be more than movies I think because many users buy once or twice and very few users buy frequently for a small online store…


I was checking on ways to address this cold start problem in the FastAI forum. Look at who initiated the discussion! :smiley:


Are Kernels randomly chosen or are they chosen from a limited set that are known to perform well?


kernels are typically initialized with random values. The size of the kernels on the other hand is chosen from a small set (typically) such as 3x3, 5x5, 7x7 - pros/cons for choice of these sizes will become clear as we proceed through the remaining chapters.

1 Like

I liked this blogpost which allows you to play around and understand kernels.


OK. I thought some of the standard image processing ones would be chosen. e.g Kernel (image processing) - Wikipedia

Question from Vasudev S.

What if the values are negative after doing conv operation on the image? How to handle this?

A follow up question: what does the the negative values implies after doing the conv operation on an image?


Just the fact that convolution was not activated at that particular point. For example for top-edge/bottom-edge convolution kernels! But apart from that I am not sure.

1 Like