Lesson 3: Deep Learning 2019 - Data blocks; Multi-label classification; Segmentation

One important feature of the Planet dataset is that it is a *multi-label* dataset. That is: each satellite image can contain *multiple* labels, whereas previous datasets we've looked at have had exactly one label per image. We'll look at what changes we need to make to work with multi-label datasets.

Next, we will look at *image segmentation*, which is the process of labeling every pixel in an image with a category that shows what kind of object is portrayed by that pixel. We will use similar techniques to the earlier image classification models, with a few tweaks. fastai makes image segmentation modeling and interpretation just as easy as image classification, so there won't be too many tweaks required.

We will be using the popular Camvid dataset for this part of the lesson. In future lessons, we will come back to it and show a few extra tricks. Our final Camvid model will have dramatically lower error than an model we've been able to find in the academic literature!

What if your dependent variable is a continuous value, instead of a category? We answer that question next, looking at a keypoint dataset, and building a model that predicts face keypoints with high accuracy.

[TheEbbemonster]

People who are true experts at their field usually know how to explain complex topics is simple terms. I have read tons and tons of books about deep learning, and Jeremy Howard is with a distance the person who has been the best at explaining deep learning in such term. Thanks for an incredible course!

[maskrey]

This course should be the standard for how to teach machine learning. I have taken many courses in machine learning, both through formal education and online, and all of them separate the theory and the hands-on application in different lectures. Jeremy's approach is SO much better there is no comparison!

[tushihahahi]

Jump to 56:00 for * Camvid segmentation Example
01:34:00 for * Regression with BIWI head pose *
01:41:00 for * NLP IMDB *

[trangnv]

Give Jeremy the Nobel price for democratizing ML and DL

[kawingchan]

At 1:55:00, the fundamental operational aspect of DL is simple. Andrew Ng also mentioned that his Stanford students were also a bit disappointed, after achieving the mathematical “peak” of seeing backprop in full glory details. The analogy can be made with Feynman saying, to the effect, law of nature is simple, but her actions are complicated. Given interactions of millions of these neurons, our intuition about their collective actions may not work, just as Jeremy said in this lesson.

[kevalan]

I am puzzled why transfer learning would work if you double the resolution. Basically all the low-level filters should be fairly incorrect, right? Unless they are scale-independent for some weird reason.

[mjafar]

Run the sound through a compressor filter PLEASE for next workshop.

[prizmaweb]

There was a question about unsupervised approaches for segmentation. Do look at use of expectation maximisation for segmentation

[abhinavsingh]

He talks about how to choose the learning rate at the 48 minute mark; saving for future reference.

[sanjeetpatil]

@13.50 can you kindly paste the whole code of the cell number 6
Since it is not visible in the whole video

[ilanaizelman]

Every lesson has around 30% decrease of people watch it. Amazing Course!

[usmabhatt]

Kindly update us with new function names.
Is not there any forum or web page type, where we check which function names are updated?

[apolloapostolos]

1:51:21 —>> This is why Mr. Howard is valued.

[sandeepms]

Thanks for the great tutorials ! Can we access these Notebooks from somewhere?

[jonatani]

1:27:00
I assume that an image's size is really its width or height, right? If so, by doubling the size, its actual size will be 4 times bigger, right? If so, batch size should be quartered. Am I right or is any of my assumptions wrong?

[albertotono]

42:32:00 nice explanation for fine-tuning. were someone able to implement it properly in their projects?

[alexturk]

I don't know if Jeremy misspoke, but at 8:40 he said that the error had decreased by 500%. Nothing can decrease by more than 100% of itself, it is basic percentages.

[kawingchan]

At 1:33:16, the use of 16 bit improve generalization, this is rather a surprise as I came across similar with quantization, but it tends to sacrifice accuracy. I speculate the 16 bit op is a form of regularization. after all, our brains don’t need high precision signaling to perform so well.

[toequantumspace]

Sadly, my nvidia gtx1070 does not support fn16

[sardarvallabhaipatel]

Hello !
how can I save the segmented mask so that I can output the segmented image as jpg format in my web applications ??