2. 2
When a solution to a problem can only be modelized by the data that defines it, you should use machine
learning to reach that solution.
There exist many different machine learning algorithms , but they often fall into one of those categories :
o Supervised learning
o Unsupervised learning
o Semi supervised learning
o Reinforcement learning
What is machine learning ?
Machine learning provides systems the ability to automatically learn and improve from experience
without being explicitly programmed.
4. 4
o In supervised learning all the data is labeled
o Mostly used to learn patterns in data in order to generate predictions
o A loss function can be computed based on true labels
Supervised learning
Learning with fully labeled data
Image Source : https://www.quora.com/What-is-machine-learning-and-AI#
5. 5
Mostly used on regressions and classifications :
o Price predictions
o Age prediction
o Gender classification
o Spam detection
o Medical diagnosis
Supervised learning applications
Image Source : https://www.kaggle.com/ashishtripathy/a-data-science-primer
7. 7
o The data is not labeled
o The network must find patterns and similarities in the data
o Unsupervised learning is widely used for clustering and to generate data visualizations
Unsupervised learning
Finding similarities between unlabelled data
Image Source : https://recast.ai/blog/the-future-with-reinforcement-learning-part-2-comparisons-and-
applications/main-qimg-e510d5175c56d0b7d78e8c59a7a8c8d5/
8. 8
o Clustering similar images together
o Market segmentation
o Data visualization generation
o Association problems
o Dimensionality reduction
Unsupervised learning applications
Clustering , visualization and association problems
Image Source : http://hpssociety.info/news/dbscan-scikit.html
10. 10
Getting a big labeled dataset is no east task, Semi supervised learning allows us to work with mostly non
labeled data.
o The data is mostly non labeled
o Two pass algorithm to combine the strenght of unsupervised and supervised learning
1. Run an unsupervised machine learning algorithms to cluster the data based on the labeled data
2. Easily assign labels based on the clusters
3. Run a supervised machine learning algorithm with the newly created labels
Semi supervised learning
Merging both supervised and unsupervised learning to improve performances
12. 12
o Experience driven learning
o Closer to human learning mechanism
o Sparse delayed feedback
o Value network : defines the win situation
o Policy network : defines the actions to take in order to win
o Behavioral learning
Reinforcement learning
A radically different approach
Image Source : https://playlearnanalytics.com/
13. 13
o AlphaGo
o Video games AI
o Autonomous car behavior
o Autonomous drones / swarm behavior
Reinforcement learning applications
Capable of complex real time decisions