1. Learning from data
How does AI learn from data?
2. AI functions and areas involved
Once data has been acquired, almost every AI system nowadays relies on
3. AI functions and areas involved
Machine and Deep Learning techniques, our two next areas to explore!
4. Enter Machine Learning (ML)
Simply put, Machine Learning consists of learning from data to identify patterns.
5. Enter Machine Learning (ML)
One category of Machine Learning is supervised learning, for predictive tasks like classification and regression.
6. Enter Machine Learning (ML)
Unsupervised learning specializes in other pattern recognition tasks including clustering and anomaly detection,
7. Enter Machine Learning (ML)
and reinforcement learning investigates agents that learn by themselves from past experiences.
8. Enter Machine Learning (ML)
Finally, deep learning techniques use so-called neural networks to solve much more complex variants of all of these ML problems. Let's look at each category in more detail.
9. Supervised Learning: classification
Classification problems consist of assigning the right category or class to new data observations whose class is unknown.
In binary classification, there are two possible classes, and an observation may belong to one class or the other. For instance, we can build an ML model that, given data observations of penguin species described by different attributes like flipper length and body mass, predicts whether or not the penguin species is Gentoo.
10. Supervised Learning: classification
By contrast, multi-class classification assigns one out of several mutually exclusive classes to each new observation, for instance, to classify penguin data into three possible species.
Supervised learning approaches like classification require pre-labeled data observations to learn from, that is, data instances whose classes are already known.
11. Supervised Learning: regression and forecasting
Regression is the numerical counterpart of classification, where the target value to predict is quantitative: a number. Predicting a house value based on some attributes is an example of a problem solvable via regression.
Another supervised problem is time series forecasting, where the goal is to predict the future values of our target, based on observations from the past. For instance, forecasting the number of daily bus passengers for next month, based on historical data.
12. Unsupervised and reinforcement learning
So far, data needed to be labeled to perform supervised learning.
Can ML work when data is not labeled? Absolutely! That's the domain of unsupervised learning, which tries to learn properties or patterns behind the data, without assigning them to a predefined label.
This way, we can for example use clustering techniques to find subgroups of similar data.
Anomaly detection helps identify unusual data observations like suspicious credit card transactions.
Association rule discovery is a popular problem in retail to find items that are frequently bought together.
And in reinforcement learning, an AI agent is trained to solve complex problems like navigating a maze or playing videogames by trial and error.
13. How about Deep Learning?
So, what about Deep Learning? Deep Learning can be seen as the modern evolution of classical ML to solve particularly advanced and challenging problems. The idea behind it is that it teaches computers to learn from data as a human brain would do. Deep learning models look for patterns in pictures, text, sounds, and other data by using so-called deep neural networks.
These powerful models require huge amounts of data to be trained.
14. How about Deep Learning?
Thanks to technological advancements in computing and storing data, deep learning has gone a lot further in recent years. It can now perform impressive tasks like recognizing objects in videos, translating and summarizing text, and even generating realistic conversations, images, and songs.
15. Let's practice!
Over to you now!