Outline • Human learning behavior • Introduction to meta learning • Realizing meta learning • Experiment result • conclusion 1 Human Learning Behavior Human can recognize an object precisely by showing one or a few that object pictures . Learning to recognize Human can quickly grasp motor driving skill if already learnt how to ride bicycle. Learning new skill 2 Introduction to Meta Learning error Meta Learner info data Learner info tasks Base Learner Architectures of traditional AI model and meta learning Ø Traditional AI model – – – – large amount of data thousands iterations learning from scratch mastering one task Ø Meta learning – – – – a few data for learning a few iterations utilizing past experience Artificial General Intelligence 3 Realizing Meta Learning Three ways to realize Meta Learning 4 Learning the Optimizer Architecture of Meta Networks An augmented layer 5 Learning the Initialization Viewing the process as : •Searching internal representation •Maximizing sensitivity to loss of parameters The process of MAML algorithm 6 Experiment Result on Meta Networks Meta Networks performance on Omniglot and Mini-ImageNet 7 Experiment result on MAML MAML performance on Omniglot and MiniImagenet 8 Conclusion Advantage: pDrawing on the processing of human learning pFew data to obtain new skill pFast adapting to new tasks with previous knowledge pThought as stepping stone for Artificial General Intelligence Disadvantage: pPerforming slightly poor pHard to train pDistance way to catch up with humans 9 Reference [1] Tsendsuren Munkhdalai and Hong Yu. "Meta networks." ICML, 2017 [2] Andrychowicz M , Denil M , Gomez S , et al. Learning to learn by gradient descent by gradient descent. NIPS, 2016 [3] Finn C , Abbeel P , Levine S . "Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks. " ICML, 2017. [4] Nichol A , Achiam J , Schulman J . On First-Order Meta-Learning Algorithms[J]. 2018. [5] Gregory Koch, Richard Zemel , Ruslan Salakhutdinov "Siamese Neural Networks for One-shot Image Recognition ." ICML, 2015 10