AI and Predictive Analytics in Data-Center Environments Supervised - PowerPoint PPT Presentation

AI and Predictive Analytics in Data-Center Environments Supervised Learning Methods Josep Ll. Berral @BSC Intel Academic Education Mindshare Initiative for AI

Introduction “If we have data and it is labeled, we can learn their relation and predict future labels”

Supervised Learning • Supervised Learning • Training data is already labeled • Want to predict new unlabeled data

Supervised Learning Observe Set of Examples <features> + <label> Label Later... New Set of Automatic Examples Labeling Observe Model <features> <labels>

Supervised Learning • Labeling data: • By hand • By known methods • By posterior metrics • From known data Observe Label dataset “cat” “cat” “cat” in out “cat” “cat” “dog” “dog” “dog” “cat” “dog”

Typical Flow What is “dog” Model this? Right! [Reinforce]

Typical Flow What is “dog” Model this? Right! [Reinforce] What is “dog” Model this? No! It is a “cat” [Adapt]

Typical Flow What is “dog” Model this? Right! [Reinforce] What is “dog” Model this? No! It is a “cat” [Adapt] What is “cat” Model this? Right! [Reinforce]

Typical Flow What is “dog” Model this? Right! [Reinforce] Iterate What is until the “dog” Model this? model is accurate No! It is a “cat” [Adapt] What is “cat” Model this? Right! [Reinforce]

Good Procedures • Keep some data “unseen”, to avoid overfitting / memorizing Algorithm Model Training Set

Good Procedures • Keep some data “unseen”, to avoid overfitting / memorizing Tune the Algorithm Algorithm Model Training Set No Is it Predict Model Evaluate good? Validation Set

Good Procedures • Keep some data “unseen”, to avoid overfitting / memorizing Tune the Algorithm Algorithm Model Training Set No Is it Predict Model Evaluate good? Yes Validation Set Final Predict Model Evaluation Test Set

Supervised Learning Algorithms & Methods!

Algorithms & Methods • Classification • The outputs are “classes” “cat” E.g.: “dog” • Regression • The outputs are “quantities” Max Speed E.g.: Car Properties

Some Methods • Regression algorithms • Linear & Polynomial Regression, Gaussian Processes, ... “Attempt to find a function/set -of-functions that match with the example points” • The learning process minimizes the regression error

Some Methods • Trees and Forests • Decision Trees, Regression Trees, Random Forests “Attempt to find a set of recursive partition that minimize the classification or regression error” “A” “C” “B” “A” “A” “B” “C” “B” “B” “C” “A” “C”

Some Methods • k – Nearest Neighbors “ Compare new samples with some memorized ones, and classify/predict as the ‘k’ nearest ones” ?

Some Methods • Bayesian Methods • Naïve Bayes, Bayesian Networks, ... “Compute probabilities of classes, events and relations, then apply Bayes theorem” P(A|B) = P(A) · P(B|A) / P(B) P(Class|Example) = P(Class) · P(Example|Class) / P(Example) = P(Class & Example) / P(Example)

Some Methods • Support Vector Machines “Find the function that best divides classes, with a minimal tolerance for errors”

Summary • Supervised Learning: • Models learn from labeled data, and “human - driven” tuning • Methods for Regression & Classification • Lots of Algorithms to be applied • Each with its characteristics • Strong and weak points • Different consumption of resources