際際滷

際際滷Share a Scribd company logo

Machine learning a developer's perspective

Download as PPTX, PDF
Rupak Chakraborty
Rupak Chakraborty

This document provides an overview of machine learning from a developer's perspective. It begins by stating what ML is not, such as a silver bullet or rule-based expert system. Two definitions of ML are given as learning from data without explicit programming and developing algorithms that can teach themselves tasks from examples. The math concepts behind ML like statistics, probability, and optimization are listed. Different types of ML like supervised, unsupervised, and reinforcement learning are described. Tools for ML in Python are mentioned. Best practices like splitting data and adjusting hyperparameters are recommended. The document provides online resources for ML notebooks, tutorials, books, and blogs.

1 of 21
Download to read offline
Machine Learning A Developers Perspective
Walk forward in pure radiance of the past
What ML is not ? A silver bullet to solve all your automation needs. Rule based expert systems (Classic Business Analysts) Dark Magic which is able to infinitely generalize from a small dataset. Skynet type of algorithm with superhuman intelligence.
So whats ML anyway.? Machine Learning is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it - Dan Ariely Machine Learning is the art and science of developing algorithms which are able to teach computers how to do a task without being explicitly trained for it (i.e. without exposing it to all possible patterns)
The Math Behind ML Statistics Linear Algebra Probability Multivariate Calculus Optimization Theory Linear/Integer Programming Machine Learning
Taxonomy of AI Machine Learning Symbolic Logic Game Theory Social Network Analysis General Game Playing Decision/Control Systems
Hues of ML Supervised Learning Semi- Supervised Learning Unsupervised Learning Reinforcement Learning Deep Learning Active Learning Zero-Shot One Shot Learning Transfer Learning Classification Regression ClusteringLabel Propagation Q-Learning
Building Blocks of ML Feature Extraction Class Imbalance Handler Classification Evaluation Ensemble Feedback
Continued Bias-Variance Tradeoff Regularization Hyperparameter Optimization PAC Learning &VC-dimension
In the church of Reverend Bayes
Continued Despite the assumption of mutually independent feature set, it works wonderfully well most times. Theoretically the best classifier (lowest error rate) among any possible classifier (given enough data) Good for easily establishing fairly strong baselines. Not good at handling noise, missing feature values and feature correlations
Like Breeds Like
Continued Highly non-linear, lazy, online instance based learner Theoretically the error rate is twice the bayes error rate Optimal value of K has to be learned in order to balance the bias-variance trade-off Fails in high dimensional spaces due to the curse of dimensionality
Tools of the trade
Python Ecosystem
Enough Theory Now Show Me Some Code
Online Resources http://sebastianraschka.com/notebooks/ml-notebooks.html http://www.erogol.com/machine-learning-pathway/ http://www.datasciencecentral.com/profiles/blogs/top-10-ipython-tutorials-for-data-science-and-machine- learning http://www.kdnuggets.com/2017/01/blogs-analytics-big-data-mining-data-science-machine-learning.html http://www.datasciencecentral.com/profiles/blogs/15-deep-learning-tutorials https://github.com/rasbt/python-machine-learning-book/blob/master/docs/references.md https://charlesmartin14.wordpress.com/ http://people.cs.pitt.edu/~milos/courses/cs3750/
Best Books to Read Pattern Classification Duda, Hart, Stork Machine Learning Tom Mitchell Elements of Statistical Learning Friedman, Hastie, Tibshirani Machine Learning, A Probabilistic Approach Kevin Murphy Introduction to Statistical Pattern Recognition K Fukunaga
Best Practices Always split your data into train, test, validation sets Take care of bias-variance tradeoff Adjust your hyperparameters for optimal performance Use domain/problem specific evaluation metrics Combine different classifiers/regressors to form an ensemble.
Questions ???
Thank You

More Related Content

Machine learning a developer's perspective

  • 2. Walk forward in pure radiance of the past
  • 3. What ML is not ? A silver bullet to solve all your automation needs. Rule based expert systems (Classic Business Analysts) Dark Magic which is able to infinitely generalize from a small dataset. Skynet type of algorithm with superhuman intelligence.
  • 4. So whats ML anyway.? Machine Learning is like teenage sex: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it - Dan Ariely Machine Learning is the art and science of developing algorithms which are able to teach computers how to do a task without being explicitly trained for it (i.e. without exposing it to all possible patterns)
  • 5. The Math Behind ML Statistics Linear Algebra Probability Multivariate Calculus Optimization Theory Linear/Integer Programming Machine Learning
  • 6. Taxonomy of AI Machine Learning Symbolic Logic Game Theory Social Network Analysis General Game Playing Decision/Control Systems
  • 7. Hues of ML Supervised Learning Semi- Supervised Learning Unsupervised Learning Reinforcement Learning Deep Learning Active Learning Zero-Shot One Shot Learning Transfer Learning Classification Regression ClusteringLabel Propagation Q-Learning
  • 8. Building Blocks of ML Feature Extraction Class Imbalance Handler Classification Evaluation Ensemble Feedback
  • 9. Continued Bias-Variance Tradeoff Regularization Hyperparameter Optimization PAC Learning &VC-dimension
  • 10. In the church of Reverend Bayes
  • 11. Continued Despite the assumption of mutually independent feature set, it works wonderfully well most times. Theoretically the best classifier (lowest error rate) among any possible classifier (given enough data) Good for easily establishing fairly strong baselines. Not good at handling noise, missing feature values and feature correlations
  • 13. Continued Highly non-linear, lazy, online instance based learner Theoretically the error rate is twice the bayes error rate Optimal value of K has to be learned in order to balance the bias-variance trade-off Fails in high dimensional spaces due to the curse of dimensionality
  • 14. Tools of the trade
  • 16. Enough Theory Now Show Me Some Code
  • 17. Online Resources http://sebastianraschka.com/notebooks/ml-notebooks.html http://www.erogol.com/machine-learning-pathway/ http://www.datasciencecentral.com/profiles/blogs/top-10-ipython-tutorials-for-data-science-and-machine- learning http://www.kdnuggets.com/2017/01/blogs-analytics-big-data-mining-data-science-machine-learning.html http://www.datasciencecentral.com/profiles/blogs/15-deep-learning-tutorials https://github.com/rasbt/python-machine-learning-book/blob/master/docs/references.md https://charlesmartin14.wordpress.com/ http://people.cs.pitt.edu/~milos/courses/cs3750/
  • 18. Best Books to Read Pattern Classification Duda, Hart, Stork Machine Learning Tom Mitchell Elements of Statistical Learning Friedman, Hastie, Tibshirani Machine Learning, A Probabilistic Approach Kevin Murphy Introduction to Statistical Pattern Recognition K Fukunaga
  • 19. Best Practices Always split your data into train, test, validation sets Take care of bias-variance tradeoff Adjust your hyperparameters for optimal performance Use domain/problem specific evaluation metrics Combine different classifiers/regressors to form an ensemble.