Machine learning
0 likes178 views
This document provides an overview of machine learning concepts including: 1) The main types of machine learning problems are classification, regression, clustering, and optimization. 2) Supervised, unsupervised, and reinforcement learning are the main approaches to teach a machine. 3) Building machine learning models involves data collection/preparation, model training/testing, and deployment of the trained model.
Machine learning
- 1. Machine Learning Nicolas P. Rougier July 2, 2018 Inria – Institute of Neurodegenerative Diseases – Bordeaux
- 2. What to learn? Considering an input x (that can be an image, a text, a vector, a sequence, a set of features, a scalar, etc.) Classification To find the category Cx of some input x Regression To find the value f(x) of some input x Clustering To divide all inputs {x} into distinct groups Optimization To find a sequence {ai} such as to optimize ∑ f(ai) 1
- 3. Classification We know that (x1, y1) is blue, (x2, y2) is blue, (x3, y3) is red., etc. What is the color of (x, y)? 2
- 4. Regression We know that f(x1) = y1, f(x2) = y2, f(x3) = y3, etc. x What is the value of f(x)? 3
- 5. Clustering We have (x1, y1), (x2, y2), (x3, y3), etc. What group (x, y) belongs to? 4
- 6. Optimization You chose C and got no reward, you chose F and got no reward, you chose N and got no reward, etc. A B C D E F G H I J K L M N O P What do you choose next? 5
- 7. How to learn? Supervised learning To learn a function that maps an input to an output based on example input-output pairs Reinforcement learning To find a policy that maps states to action such as to maximize long-term reward Unsupervised learning To infer a function that describes the structure of some (unlabeled) data 6
- 8. Machine learning ̸= Human learning To learn anything, at any time, using any method and a few samples. Explicit & implicit learning, lifelong learning, imitation learning, one-shot learning, declarative and procedural memory, short and long term memory, working memory, forgetting, motivation, emotion, etc. 7
- 9. The machine-learning pipeline Data (60% of your time) Analysis, acquisition, cleaning, preparation, augmentation, balance Model (30% of your time) Hyperparameters, pre-training, training, tweaking, testing Deployment (10% of your time) Optimization, dockerization, evaluation 8
- 10. Common pitfalls Data Small, unbalanced, biased, leakage, variance, etc Model Initialization, local minima, curse of dimensionality, etc Deployment Scalability, fragility, ethics, etc 9
- 11. How big is “big”? Deep convolutional network (Hinton, 2010) • 5 convolutional layers • 650,000 units • 60,000,000 parameters • 1,200,000 train images • 100,000 test images • 1,000 classes Latest OpenAI DOTA player (June 2018): 180 years worth of games, 256 GPUs and 128,000 CPU cores. 10
- 12. Do I need deep learning? If all you have is a hammer, everything looks like a nail. 11
- 13. Conclusion Lot of challenges ahead • Learn with a few samples • Lifelong learning • Transfer learning • Unbiased learning • Robust learning • Unsupervised learning • Reinforcement learning Lot of opportunities as well... 12