Thread by @svpino

Let's talk about learning problems in machine learning:

▫️ Supervised Learning
▫️ Unsupervised Learning
▫️ Reinforcement Learning

And some hybrid approaches:

▫️ Semi-Supervised Learning
▫️ Self-Supervised Learning
▫️ Multi-Instance Learning

Grab your ☕️, and let's do this👇

Supervised Learning is probably the most common class of problems that we have all heard about.

We start with a dataset of examples and their corresponding labels (or answers.)

Then we teach a model the mapping between those examples and the corresponding label.

[2 / 19]

The goal of these problems is for a model to generalize from the examples that it sees to later answer similar questions.

There are two main types of Supervised Learning:

▫️ Classification → We predict a class label
▫️ Regression → We predict a numerical label

[3 / 19]

A Supervised Learning Classification example:

Given a dataset with pictures of dogs and their corresponding breed, build a model that determines the breed of a new picture of a dog.

Notice how the goal is to predict a class label (the breed of the dog.)

[4 / 19]

A Supervised Learning Regression example:

Given the characteristics of a group of houses and their market value, build a model that determines the value of a new house.

Notice how the goal is to predict a numerical label (the value of the house.)

[5 / 19]

Unsupervised Learning is about finding relationships in data.

There are no labels involved in this process. We aren't directly teaching the algorithm through labeled examples. We are expecting it to learn from the data itself.

[6 / 19]

An example of Unsupervised Learning:

Given a list of prospective customers, group them into different segments so your marketing department can reach out to them.

Here the algorithm will determine different groups for your customers based on existing relationships.

[7 / 19]

Clustering is the most common example of Unsupervised Learning.

You have probably heard of k-Means as one of the most popular clustering algorithms. Here, "k" represents the number of clusters we want to find.

[8 / 19]

Reinforcement Learning is pretty cool:

An agent interacts with the environment collecting rewards. Based on those observations, the agent learns which actions will optimize the outcome (either maximizing rewards or minimizing penalties.)

[9 / 19]

An example of Reinforcement Learning:

A robot learning its way from point A to point B in a warehouse by walking and exploring the different paths between the two locations.

Every time the robot gets stuck is penalized. When it reaches the goal, it is rewarded.

[10 / 19]

But of course, AlphaZero (Chess) and AlphaGo (Go) are probably two of the most popular Reinforcement Learning implementations.

DeepMind is the company behind all of this research. Check out their website for some really cool articles.

[11 / 19]

In Semi-Supervised Learning, we get a lot of data but only a few labels. Sometimes, even the labels we have are not completely correct.

The goal is to build a solution that takes advantage of all the data we have, including the unlabeled one.

[12 / 19]

A few days ago I posted a thread about Active Learning, a semi-supervised approach.

Check it out if you are looking for more information about one possible way to approach this problem.

[13 / 19]

https://twitter.com/svpino/status/1365307434856820736?s=20

Self-Supervised Learning is a subset of Unsupervised Learning.

The idea is to use Supervised Learning to solve a task (pretext task) that can later be used to solve the original problem.

We frame the problem in a way where we can take advantage of existing labels.

[14 / 19]

Three popular examples of Self-Supervised Learning problems:

▫️ Autoencoders
▫️ Transformers
▫️ Generative Adversarial Networks

I'm planning to talk more about these 3 in the future, so stay tuned for that.

[15 / 19]

The last one on our list is Multi-Instance Learning, a subset of Supervised Learning.

The difference here is that instead of having labels for every specific example, we have labels for groups of examples.

These groups are called "bags."

[16 / 19]

Multi-Instance Learning models focus on predicting the class of a bag given the individual instances contained in it.

An example of this type of problem is content-based image retrieval, where we aim to find images based on the object it contains.

[17 / 19]

Whenever you are in front of a problem, you could identify potential solutions if you can recognize its type given its characteristics.

That's the value in understanding these types of machine learning problems and possible applications.

[18 / 19]

I have a ton more of this type of content for you. Like, pounds and pounds of it!

Just follow me, so you don't miss the fun part of Twitter.

And thank you for the support!

[19 / 19]

🦕