2. OUTLINE
• What is Machine Learning?
• Applications in Machine Learning
• (The Machine Learning) Model
• Machine Learning Models in Action
• Training Data
• Model / Data Considerations
• Models
• DecisionTree
• Random Forest
• Clustering
• Linear Models
• SupportVector Machines (SVM)
• Artificial Neural Networks
• Deep Learning (CNN)
• Reinforcement Learning
3. WHAT IS MACHINE LEARNING?
“Field of study that gives computers the ability to
learn without being explicitly programmed.”
- Arthur Samuel
A computer program is said to learn from experience E with respect to
some taskT and some performance measure P, if its performance onT,
as measured by P, improves with experience E.
-Tom Mitchell1959 1998
5. APPLICATIONS INTHE MODERN WORLDAPPLICATIONS INTHE MODERN WORLD
Optical Character
Recognition
Recommendation
Engines
Facial Recognition
Autonomous
Vehicles
Personal Assistants /
Chat Bots
9. MODEL
A REPRESENTATION OF A REALWORLD PROCESS
Water Cycle
Evolution
Neuron-McCulloch & Pitts Model, 1943
10. MACHINE LEARNING MODELS IN ACTION
Untrained Model(Old)
DATA
Trained Model New
Data
Info?
Prediction?
Decision?
Expert
Knowledge
11. TRAINING DATA
Feature =Variable = Predictor Objective Measurement
Height (in) Weight (lb) Color Claws retract Class
11.2 10.1 black yes cat
23.1 45.2 black/white no dog
13.0 20.1 black/white yes cat
9.7 7.2 white yes cat
… … … … …
12. TRAINING DATA
Feature =Variable = Predictor Objective Measurement
Height (in) Weight (lb) Color Claws retract Class
11.2 10.1 black yes cat
23.1 45.2 black/white no dog
13.0 20.1 black/white yes cat
9.7 7.2 white yes cat
… … … … …
13. TRAINING DATA
Feature =Variable = Predictor Objective Measurement
Height (in) Weight (lb) Color Claws retract Class
11.2 10.1 black yes
23.1 45.2 black/white no dog
13.0 20.1 black/white yes
9.7 7.2 white yes cat
… … … … …
14. TESTING DATA (NO PEEKING!)
Training and testing sets must
ALWAYS be disjoint
• Cross-validation
• Leave-one-out
• OOB (Out-of-bag for
ensembles)
15. MODEL/DATA CONSIDERATIONS
(RELEVANT TO MODEL SELECTION)
Each model can/cannot handle certain data characteristics / analysis needs
• Supervised vs. Unsupervised data?
• Class Imbalance (200 cats vs. 3 dogs)
• 2-class vs. Multiclass (say 200 cats, 146 dogs, 25 sugar gliders, 5 platypuses)
• Scale issues (see Distance-based Clustering; Normalization / Standardization)
• FeatureType (Categorical, Continuous, etc)
• Dimensionality (# of features / measurements)
• Cost Sensitivity (Miss / False Alarm – can the model adjust?)
• Propensity to Overtrain (fitting to noise – see Bias vs.Variance)?
• Need to estimate uncertainty?
• Ability to adapt to changing conditions (parameters)?
• Robustness to sparse data (parameter estimation)?
16. DECISIONTREE
1) At each node, a question is asked
about a specific feature
2) The answer directs data left/right
3) Decision trees must be pruned to
prevent overtraining
18. RANDOM FOREST
Random Forest is an ENSEMBLE of DecisionTrees
Node Splits (Training)
• Bagging (resampled data for each
tree)
• “Best” univariate split on random
subspace (subset of all features)
• Gini Impurity
• Leaf nodes are class homogeneousLeo Breiman
19. RANDOM FOREST
Random Forest is an ENSEMBLE of DecisionTrees
Leo Breiman
Classification
1) Samples propagate through
each tree
2) Tree “votes” for a class
based on leaf node
3) Final decision based on class
conditional probability
26. SUPPORTVECTOR MACHINES (SVM)
Maps linearly nonseparable data to a higher dimension
Kernel trick
makes this mapping more
efficient
Also: sub-gradient descent, coordinate descent
27. SUPPORTVECTOR MACHINES (SVM)
Support vectors in the feature space used for classification
Support vectors are
determined by the
most difficult points to
classify…
31. DEEP LEARNING (CONVOLUTIONAL NN)
From the Latin convolvere,“to convolve” means to roll together
We convolve an image with multiple kernels (filters) at each layer
34. REINFORCEMENT LEARNING
A reward-driven approach for a
machine to “self-learn”
• At each step, the agent takes an
action based on environment state
• The agent receives a reward based
upon the new state (post-action)
• The agent’s goal is to maximize his
reward
35. REINFORCEMENT LEARNING
Donald Michie creates MENACE, 1963
(Machine Educable Noughts And
Crosses Engine)
MENACE learned to play TicTacToe
using stacks of matchboxes