Understanding terms in literature

Hello
Reading some literature online I’ve come across several terms that I hope someone can clarify. Specifically from this article below I am struggling to understand what they are talking about when they say:

  1. Feature maps (what we call filters in this course?)
  2. Positive and Negative filters
  3. Positive and Negative activations
  4. Optimization - what do they mean by “optimization” here?

thank you

1 Like

Can anyone help me with this please?
Having finished Course 4 of DLS/Convolutional NN’s I’d really hoped I’d be able to read this kind of articles - but I’m really at a loss here.
Moderators, do you have any thoughts?
thank you

You’re asking a question that is beyond the scope of the courses here. Please realize that we’re all volunteers here and are not being paid to answer questions, so it is up to each individual listening whether they want to invest the effort that it will take to answer a question like this. My first step was to make sure that this paper is actually by people who know what they are doing. There is plenty of material out there on Medium that’s written by people who’ve just taken one or two courses and are trying to pretend they have something useful to say. But the authors here do appear to be legitimate experts and the paper does look pretty interesting. So I scanned through most of it and can offer a few reactions:

Do you remember the lecture Prof Ng gave in Week 4 of ConvNets entitled “What Are Deep ConvNets Learning?” This paper is another way to approach the ideas in that lecture. In that lecture, Prof Ng describes work in which researchers instrumented neurons in the internal layers of a ConvNet that was trained to recognize images. Then they fed a collection of input images to the network and used their instrumentation to find which image patches in the inputs most strongly activate a given instrumented neuron. So that gives you a way to visualize what that neuron has learned to recognize.

The authors here make the point that using the above method has the limitation that you can only see what is triggered by those particular images. So they came up with the idea of using the trained network, picking some set of internal neurons and then defining a cost function based on the activations of those neurons. Then they use the gradients from that cost function to modify the input image. That’s the new idea here. They start from a randomly initialized input image and then run Gradient Descent based on their specific cost function to see what that does to the input image when the algorithm pushes the input to maximize the metric. So that’s what they mean by “Optimization” here: using the cost function to train the input to give the maximum activation, not training the filters and other parameters in the network itself (those are already trained and stay constant). Of course that means that the input is completely synthetic.

They then describe more details about how to do that in different ways.

I don’t see that they used the term “feature maps” anywhere.

They point out that you can train the input to give the largest positive value on the metric or you can flip it around and see what input generates the largest negative value or the smallest value on the metric.

I only spent maybe 10 minutes reading what they said, so I don’t claim to understand everything they said or explained, but that’s my summary of what I read. Feel free to discuss more based on my comments above.

Thank you, Paul
I know you are all volunteers, and I debated whether to post this question on an article from the Internet. I did because these researchers published many influential papers on the topics of ConvNets, and visualizing what ConvNets actually do helped me gain intuition tremendously: Dr. Ng had a few slides showing what conv filters look for like edges, lines, etc, and it was super helpful.
These terms I asked about appear often in other articles that I’ve read, and they were important to understand what was talked about. This course condenses a LOT of info and knowledge, so I use outside resources to get a better intuition of the things Dr. Ng presents.

Your answer was incredibly helpful, and thank you so much for the summary and the explanations.
Thanks for all you do!

P.S. Turns out Feature Map is the output of a convolutional layer representing specific features in the input image. It appeared in a few articles I read so far.

Hi, Svetlana.

It’s great that you are taking the knowledge from the course and expanding to other areas and research papers. It’s totally fine to post this kind of a question that goes beyond the courses here. My point was just that you can’t get your feelings hurt if no-one answers. The trick is to make it an interesting enough question that people will want to participate. I also learned a number of interesting things by reading the article that you found, so thanks for triggering the discussion!

Best regards,
Paul