Transfer Learning

Le tranfer learning consiste à réutiliser un modèle de deep learning déjà appris sur une grande base de données pour un problème aux dimensions plus modestes. C’est la première chose à faire quand on commence le deep learning : c’est souvent très rapide pour des résultats déjà acceptables. Cela a aussi l’avantage d’être peu coûteux comparé à l’apprentissage d’un réseau de neurones profond complet sur une grande bases de données.

Notebooks

• Transfer Learning (Olivier Grisel)

• Search images with deep learning

Lectures - introduction

• Building powerful image classification models using very little data

• Deep Learning : choisir son framework et entrainer ses modèles dans Azure, présenté avec Olivier Grisel

Lectures - articles

• Unsupervised and Transfer Learning Challenges in Machine Learning, Volume 7

• ICML2011 Unsupervised and Transfer Learning Workshop

• Transfer Learning

• Deep Learning of Representations for Unsupervised and Transfer Learning

• Unsupervised and Transfer Learning Challenge: a Deep Learning Approach

• Transfer Learning by Kernel Meta-Learning

• A Survey on Transfer Learning

• Domain-Adversarial Training of Neural Networks

• Stability and Hypothesis Transfer Learning

• Transfer Learning Decision Forests for Gesture Recognition

• Learning Transferable Features with Deep Adaptation Networks

• Asymmetric Transfer Learning with Deep Gaussian Processes

• Transfer Learning in Sequential Decision Problems: A Hierarchical Bayesian Approach

• Transfer Learning for Reinforcement Learning Domains: A Survey

• Unsupervised dimensionality reduction via gradient-based matrix factorization with two adaptive learning rates

• Transferability in Machine Learning: from Phenomena to Black-Box Attacks using Adversarial Samples

• Semi-supervised Knowledge Transfer for Deep Learning from Private Training Data

• Knowledge Concentration: Learning 100K Object Classifiers in a Single CNN

Lectures - hors nn

• Transfer Learning Decision Forests for Gesture Recognition

Modèles pré-entraînés

• Places CNN, Pre-release of Places365-CNNs (deep learning)

• CNTK (sur github)

• Model Zoo

• Model Gallery CNTK

• tensorflow/models