Intro

Image Classification

Uber's Dispatch Algorithm

Why apply evolution to neural network design

Evolutionary Algorithms

Inter-life learning and intra-life learning

How do we perform Neuro-Evolution?

Population Selection

Fitness Function

Recap

Tutorial Overview

Supervised learning

Maximum-a-posteriori estimates

Case-shot Classification Problem

Neural Network Design

Evaluating fine-tuning

Second-order Optimization

Prototypical Networks

Bayesian Meta Learning Approaches

What if we don't have a supervised loss function?

intro

what is metalearning?

the "why" of metalearning

the "what" of metalearning

difference between traditional learning and metalearning

the "how" of metalearning - benchmark

the "how" - emphasize / exclude

What is meta-learning?

Meta-Learning and Multi-task Learning

How does meta-learning work?

One-Shot Image Recognition

Learning as a function

Representing Functions

Generic learning

Learned learning function

Meta-Learning Categories

Gradient Descent-Based Meta-Learning

Opening Remarks

What is meta-learning?

History of meta-learning

Categorization of Methods

How has meta-learning changed each of the components of machine learning?

Categories of meta-learning

Model-based meta-learning

Recap

Neural architectures

Q&A session

One-Shot Learning

Reason One-Shot Learning Is Interesting

Meta Learning

Human Level Concept Learning through Probabilistic Program Induction

One-Shot Learning Task

Feed-Forward Neural Network

Recurrent Neural Network

Model Agnostic Meta Learning for Fast Adaptation of Deep Networks

Optimization as a Model for Few Shot Learning

Easy To Use

Logistics

Plan for Today

Multi-Task Learning vs. Transfer Learning

From Transfer Learning to Meta-Learning

Two ways to view meta-learning algorithms

What is meta-learning?

High-parameter optimization

Summary of meta-learning research

Optimization-based meta-learning

Meta-Objective

What happens at test time?

Background on Reinforcement Learning

What is meta-optimized?

Optimizer performance

General announcement

Agenda

Graphical Model of Shared Information

What happens when entropy is zero?

Is Active Learning equivalent to exploration in RL?

Does a Bayesian Meta-Learning approach work only with optimization-based meta learning?

Variational Inference

How might we do meta-learning leveraging these techniques?

Bayesian Optimization-based Meta-Learning

Evaluating Bayesian Meta-Learning

Question: How do you interpret uncertainty in the model?

Intro

Supervised Learning Pipeline

Goal of Reinforcement Learning

Goal of Multi-task Reinforcement Learning

Policy Gradients

Approximating Expectation of Samples

Policy Gradient: Q&A

Bellman Update

Hindsight Relabeling or Hindsight Experience Replay

Goal-Condition Setting

Intro: Issues with Classical Deep Learning

Alternative to Classical Learning: Few-Shot learning

Classical Learning vs Meta-Learning

Concept: Prototypical Networks

Code: Prototypical Networks - Train & Test

End

Course Reminders

Memorization

Importance of labels

Anchoring: Labels

Applying regularization to activations

What if we only have unlabeled data?

Clustering to automatically construct tasks for unsupervised meta-learning

Test Construction

Results & Conclusion

Next Lecture: Bayesian Meta Learning

Dynamic Forecasting

Forecasting

Physics Guided Ai

Idea behind Physics Guided Ai

A Deep Learning Model

Multitask Learning

Adaptive Padding

Test Error

Adaptation Error

Are Models Able To Predict Dynamics with Phase Transitions