Can there be an example notebook for multi-label classification as well. It seems multiclass and multi-label NN are quite similar in their setup but it will be useful to see an example. Also can it be assumed that multi-label classification can also be called object detection? If thats true, can object detection be explained as well. How in the lecture the borders are drawn around a pedestrian, car or a bus.
Hey @ftayyab,
Let’s answer your query part by part.
MLS Course 2 W1 and W2 are supposed to introduce the newbies in the world of ML to the concept of Deep Learning and Neural Networks. Object detection is something that can be most easily learnt once you have a working knowledge of Convolutional Neural Networks (CNNs), something which hasn’t been touched upon in the course at all. This is because most of the algorithms that deal with the task of Object Detection involved CNNs. If you are interested in learning more about CNNs & Object Detection, allow me to suggest you one of the courses focused specifically on CNNs, which you can find here. Also, if you are interested in learning more about deep learning in general, you may wanna take a look at the entire Deep Learning Specialization. The course that I have mentioned is the 4th course of this specialization.
For this question, I would like to highlight the philosophy of Prof. Andrew on the release of this specialization, which you can found in this LinkedIn post. This clearly explains why some topics have been included in the course and some topics have been excluded.
In both binary classification and multi-class classification, each example belongs to a single class only, but in multi-label classification, an example can belong to more than one classes simultaneously. Now, let’s come to object detection. In object detection, given an image, we are supposed to find and locate multiple objects in an image.
Now, first of all, I would like to highlight here that Multi-label classification is a type of classification task, while object detection is a very particular task. So, it would be completely wrong to assume that all tasks that fall under the purview of multi-label classification are object detection tasks. In other words, this question is logically incorrect, since what you are essentially asking is whether or not “a set comes inside a subset or not”. A logically correct question would be “Can it be assumed that all object detection tasks are multi-label classification?”. Now, let’s answer this question.
To answer this question, I have borrowed the definition of multi-label classification from Wikipedia, and stated it here for your reference:
Formally, multi-label classification is the problem of finding a model that maps inputs x to binary vectors y ; that is, it assigns a value of 0 or 1 for each element (label) in y .
Now, there are 2 important distinctions according to me that may or may not refute the claim “all object detection tasks are multi-label classification tasks”. First, in an object detection task, the examples might have more than one instance of the same class, for instance, there might be 2 or more cars in a single image. So, the image is assigned a single label only, but there are multiple instances of the same class. According to the formal definition as stated above, multi-label classification doesn’t have any support for this, unless and until you modify how the multi-label classification algorithms work. Second, in an object detection task, you are also supposed to output information which can localize the instances, something which is not an inherent part of multi-label classification.
Now, if you hold these distinctions firmly, the claim will be refuted. On the other hand, if you hold these distinctions tenderly, and are open to slight modifications in the algorithms, the claim will stand true.
Just for your reference, I found this from quick search on the web:
Object detection is inherently multi-label ; the difference is that it additionally adds and predicts localization information for each labeled box (and can can predict/count multiple instances of each class in an image).
You can find the source for this here. I hope this helps.
Regards,
Elemento
Hi Elemento, Great reply. really useful and thanks for the links