### Table of Contents

# Classification

## MLP

Multi Layers Perceptron, *PMC (Perceptron Multi-Couches)*

### Gradient Backpropagation

*Rétropropagation du Gradient*

#### Stochastic

#### with Inertia

#### Simulated Annealing

*Recuit Simulé*

### Newton

##### Objective

Converges faster than gradient descent

##### Quick Def

Second order

## RBFNN

Radial Basis Functions Neural Networks

#### First method

##### Objective

You have to chose `k`

##### Quick Def

k-means then gradient descent

#### Second method

##### Quick Def

incremental addition of neurons then exact method

## SVM

Support Vectors Machine

## Decision tree

*arbre de décision*

#### ID3

##### Quick Def

based on entropy

## k-nearest neighbors

*k plus proches voisins*

## Boosting

[Freund,Schapire, 1995]

##### Quick Def

Consists in combining a lot of weak classifiers to get a strong one.

### Boosting by majority

### AdaBoost

ADAptive BOOSTing, [Freund,Schapire, 1996]

The first and standard version is refered as Discrete AdaBoost.

##### Quick Def

Greedy approach

#### Discrete AdaBoost

[Freund,Schapire, 1996]

##### References

##### Full Definition

#### Real AdaBoost

[Friedman,Hastie,Tibshirani, 1998]

##### References

##### Full Definition

#### LogitBoost

[Friedman,Hastie,Tibshirani, 1998]

##### References

##### Full Definition

#### Gentle AdaBoost

[Friedman,Hastie,Tibshirani, 1998]

##### References

##### Full Definition

#### Probabilistic AdaBoost

[Friedman,Hastie,Tibshirani, 1998]

##### References

#### FloatBoost

##### Objective

AdaBoost is a sequential forward search procedure using the greedy selection strategy to minimize a certain margin on the training set. A crucial heuristic assumption used in such a sequential forward search procedure is the monotonicity (i.e. that addition of a new weak classifier to the current set does not decrease the value of the performance criterion). The premise offered by the sequential procedure in AdaBoost breaks down when this assumption is violated. Floating Search is a sequential feature selection procedure with backtracking, aimed to deal with nonmonotonic criterion functions for feature selection

##### Full Definition

#### AdaBoost.Reg

[Freund,Schapire, 1997]

##### Objective

An extension of AdaBoost to regression problems

##### References

#### Multiclass AdaBoost.M1

[Freund,Schapire, 1997]

##### Objective

Basic extension of AdaBoost to multiclass problems. A weak classifier needs to have an error rate less than 1/2, which is stronger than random guessing (1/k) and is often too difficult to obtain.

##### Quick Def

A weak classifier associates to an example a label in `{0,…,k}`

.

##### References

##### Full Definition

#### Multiclass AdaBoost.M2

[Freund,Schapire, 1997]

##### Objective

Tries to overcome the difficulty of AdaBoost.M1 by extending the communication between the boosting algorithm and the weak learner. The algorithm not only focuses on hard instances, but also on classes which are hard to distinguish.

##### Quick Def

A weak classifier associates to an example a vector in `[0,1]^k`

, and the pseudo-loss takes also into account weights according to the performance of the weak classifier over the different classes for the same example.

##### References

##### Full Definition

#### Multilabel AdaBoost.MR

[Schapire,Singer, 1998]

##### References

##### Full Definition

#### Multilabel AdaBoost.MH

[Schapire,Singer, 1998]

##### References

##### Full Definition

#### Multiclass AdaBoost.MO

[Schapire,Singer, 1998]

##### References

#### Multiclass AdaBoost.OC

[Schapire, 1997]

##### References

#### Multiclass AdaBoost.ECC

[Guruswami,Sahai, 1999]

##### References

#### AdaBoost.M1W

[Eibl,Pfeiffer, 2002]

#### GrPloss

[Eibl,Pfeiffer, 2003]

##### References

#### BoostMA

[Eibl,Pfeiffer, 2003]

##### References

#### SAMME

Stagewise Additive Modeling using a Multi-class Exponential loss function, [Zhu,Rosset,Zou, 2006]

##### References

#### GAMBLE

Gentle Adaptive Multiclass Boosting Learning, [Huang,Ertekin,Song, 2005]

##### References

### UBoost

##### Quick Def

Uneven loss function + greedy

### LPBoost

##### Objective

Not greedy, exact.

##### References

### TotalBoost

TOTALly corrective BOOSTing, [Warmuth,Liao,Ratsch, 2006]

##### References

### RotBoost

[Li,Abu-Mostafa,Pratap, 2003]

##### References

### alphaBoost

[Li,Abu-Mostafa,Pratap, 2003]

##### References

### MILBoost

(Multiple Instance Learning BOOSting), [Viola,Platt, 2005]

##### References

### CGBoost

Conjugate Gradient BOOSTing, [Li,Abu-Mostafa,Pratap, 2003]

##### References

### Bootstrap Aggregating

## Cascades of detectors

##### Quick Def

A cascade of classifiers is a degenerated decision tree, where at each stage a classifier is trained to detect almost all objects of interest, while rejecting a certain fraction of the non-object patterns (eg eliminates 50% of non-object patterns and falsely eliminates 0.1%, then after 20 stages it can be expected a false alarm rate of 0.5^20 and a hit rate of 0.999^20). It enables to focus attention on certain regions and dramatically increases speed.

## Trees of detectors

# Regression

**MLP (Multi Layers Perceptron)**

**RBFNN (Radial Basis Functions Neural Network)**

**SVR (Support Vectors Regressor)**