The 5 Deep Learning Tasks creating thousands of DL use cases

There are infinite use cases for Deep Learning(DL). But all of them come from 5 basic DL tasks, and those are the really important ones to understand its thousands of variations. This article is focused on that: What are the main tasks we can perform with DL models and how they create thousands of use cases across all industries. The cherry of the cake: the role of DL frameworks.

The 5 DL Tasks are:

• Detection
• Classification
• Segmentation
• Prediction
• Recommendation

Let me start with an example and take three real-life use cases in three different sectors: Think of Cancer detection in Healthcare, Video Surveillance in Banking, and Face Detection in the Military/Defense sector. Well, all of them use the same DL task: detection. In the three cases, what the DL model will do is to detect something. In these cases, an image of a cancer tumor, a person, or a face. But the core of the model will be detecting.

Let's talk about inputs and outputs. In cancer detection, for instance, the desired output of my DL model would be to detect cancer. And the input? well, the input is the images taken by an MRI machine. Those images are fed to a DL model, which has been trained on the task of detecting cancer. Once the DL model analyses the image it will give us the answer to the question: is it present or not?. "it" being a cancer tumor. But "it" could be as well a person breaking into a bank vault (video surveillance), or the face of an authorized person trying to access a restricted area (face recognition).

Similar to detection, there are some other DL tasks, and can be applied to any input data in order to get a desired output. So, what about the other DL tasks?

In addition to detection, we could run a classification model, or segmentation, or prediction or recommendation model. Here is how each of them answering different questions: is it present? (detection) , what type of thing is it? (classification) , to what extend is it present? (segmentation) , what is the likely outcome? (prediction) or what will likely satisfy the objective? (recommendation). And you can use all of them. If you do so, you could not only detect the tumor but you could also identify what type of tumor is it, what's its size and shape, what is the survivability prediction and what the best therapy recommendation is.

And it gets even more interesting if you take into account that the input can change as well. It doesn't have to be pictures only. It could be video, or text data, or even audio. Keep in mind that at the end of the day, for computers, any input will be reduced to 1s and 0s. Look at this:

So, once you have clear which input you will work with, and what task you want to perform, you'll have to select a DL framework and the kind of neural network you want to use to approach the problem. If you are not clear on what a neural network is, check the article I did on the AI basics. Below the link.

The good news is that there are already very good untrained models publicly available today. They are DL models using Deep Neural Networks (DNN) for object detection, image classification, face recognition, and others. How does that help? For instance, if you want a model capable to receive images and distinguish between dogs and cats, there is a model already publicly available to do so. Do you want the model to recognize also ducks? then you will have to modify it a bit. But you won't have to start from scratch and that saves you a lot of time.

Typical applications are image classification, speech recognition and natural language processing. Other examples of DL use cases: in Medicine for diabetic grading, and drug discovery. In the Media and Entertainment business DL can be used for video captioning or content-based search as well as real-time translation. In security and defense, for video surveillance and cybersecurity, or pedestrian detection, lane tracking and recognizing traffic signs when it comes to Autonomous machines.

That's it for today. I hope you had a great weekend and are having a great start to the week, wherever (and whenever) you are reading this.

David