首先推荐大家看文章：From N to N+1: Multiclass Transfer Incremental Learning
核心细想就是Transfer Incremental Learning
原理图如下：Transfer learning的任务就是检测出小狗，并且系统已经学习到了几种动物（小猫、马）。从有n个类别学习第n+1个类别。
代码组成部分：
Contents -------- data/ -- demo data tmp/ -- temporary files (e.g. source classifiers) lib/ -- algorithm implementations lib/util -- utilities lib/dogma -- parts from DOGMA library lib/mktl -- Multi-Kernel Transfer Learning implementation files lib/multikt -- MultiKT implementation files lib/tl_baselines/ -- baseline TL algorithm implementation files lib/GenericClassifier.m -- abstract base class for classifiers (kernel computation and generic evaluation routines) lib/HyperSearch.m -- hyperparameter grid search utility class lib/MulticlassOneVsRest.m -- multiclass OVA classifier, where binary classfiers can be plugged in lib/MulticlassRLS.m -- multiclass LSSVM classifier lib/SimpleNplusOne.m -- Source+1 baseline implementation lib/SourcePlusOneHingeL.m -- Source+1 (hinge) baseline implementation lib/MTKL.m -- interface to MKTL (compatibe with generic evaluation framework) lib/MultiKT.m -- interface to MultiKT (compatibe with generic evaluation framework) lib/PmtSvm.m -- interface to PmtSvm (Tabula Rasa) (compatibe with generic evaluation framework) lib/MultisourceTrAdaBoost.m -- interface to MultisourceTrAdaBoost (compatibe with generic evaluation framework) lib/MULTIpLE.m -- The MULTIpLE algorithm implementation NplusoneBenchmark.m -- main experiment file; preamble contains its description
From N to N+1: Multiclass Transfer Incremental Learning 代码分析（1）
原文地址：http://blog.csdn.net/u013476464/article/details/41145083
