ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

A General Overview of Machine Learning

?
?
Ashish Sharma
Ashish Sharma

This document provides an overview of machine learning, including: - The types of machine learning such as supervised learning, unsupervised learning, and reinforcement learning. Supervised learning involves predicting outputs from labeled inputs using techniques like regression and classification, while unsupervised learning finds patterns in unlabeled data using clustering and dimensionality reduction. - Common machine learning applications including speech recognition, machine translation, strategic gaming, computer vision, autonomous vehicles, manufacturing, and healthcare. - Effective machine learning involves reducing programming time through existing tools, customizing and scaling products as needed, and using statistics rather than logic to make decisions from real-world data.

1 of 19
Download to read offline
A General Overview of Machine Learning Boise Data Science Meetup -- September 18, 2018 Ashish Sharma ? Software Systems Engineer -- HomeCU, LLC. (2017 - present) ? Founder -- AI Developers, Boise ? City Ambassador -- AI Saturdays (global initiative of nurture.ai) ? Alumnus -- Boise State University (MS in Computer Science, 2015-2017) 1
Overview ¡ñ AI and Applications ¡ñ Intro to Machine Learning ¡ñ Types of Machine Learning ¡ñ Which algorithm should I use? ¡ñ Effective Machine Learning Image Source: Cousins of Artificial Intelligence -- Towards Datascience 2
AI Resurgence ? Computational Power (GPUs, cloud computing, distributed systems) ? Availability of large amount of Data (eg. Imagenet) ? Better theoretical understanding of the underlying techniques/algorithms ? Open and easily accessible research culture in academia and industry (NIPS, ICML, archiv.org) 3
AI Resurgence (contd..) ? Netflix Challenge (2009) $1 Million Prize (User ratings for films) ? Kaggle (2010) (over more than a million users today) ? Fei-Fei Li and team at Stanford open sourced ImageNet (2008-2010) ¡ô Imagenet Large Scale Visual Recognition Challenge (ILSVRC) ? Geoffrey Hinton¡¯s Deep Learning Team wins ImageNet 2012 (Alexnet) 4
Common Applications ? Speech recognition (virtual assistants) ? Advanced machine translation and natural language intelligence ? Strategic gaming algorithms (AlphaGo, chess) ? Computer Vision (image classification and object detection) ? Autonomous Vehicles ? Manufacturing Companies (landing.ai) ? Healthcare (Google¡¯s research on diabetic retinopathy -- with F-score of 0.95, surpassing the accuracy of 8 expert ophthalmologists) 5
Machine Learning ? Form of applied statistics with emphasis on the use of computers to learn complex mathematical functions. ? More formally, ¡°A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.¡± Image Source: xkcd 6
Types of Machine Learning ? Supervised Learning ? Unsupervised Learning ? Reinforcement Learning 7
Supervised Learning Terminologies: ? Input variable(s) ¡ô independent variable(s) ¡ô feature(s)/characteristic(s) of a single input object ¡ô Numerical -- continuous ( height, area of house) , discrete (grades, age) ¡ô Categorical (race, sex) -- nominal, ordinal ? Target variable(s) ¡ô Dependent variable(s), number/vector (eg. price of house, patient is diabetic, etc.) 8
Supervised Learning ? Function approximation ¡ô Mathematically: solve for coefficient(s) of a function ¡ô Search for a best performing model from a hypothesis space. ¡ô Make predictions based on historical (labeled) data ? Regression (predict continuous target variable) ¡ô Univariate Regression (1 input variable, 1 output variable) ¡ô Multiple Regression (>=2 input variables, 1 output variable) ¡ô Multivariate Regression (>=2 output variable) ? Classification (predict discrete/categorical target variable) ¡ô Email: Spam or not? ¡ô Is this image a dog or cat? 9
Unsupervised Learning ? Unsupervised Learning ¡ô Find hidden patterns and draw inference from (unlabeled) data ¡ô Essential for preliminary data analysis and visualization ? Clustering (grouping of similar data points) ¡ô K-Means, DBSCAN ? Dimensionality Reduction ¡ô Principal Components Analysis ¡ô Autoencoders 10
Reinforcement Learning ? AI, Animal Psychology, Control Theory ? Agents, Actions, Environment, Change in State, Reward/Punishment ? Eg. Deep Attari: ¡ô Input: Snapshots of Attari board images (State and Actions) ¡ô Algorithm: Convolutional NNs with no pooling ¡ô Output layer: tailored for regression score (Maximize Reward) 11
Beginner¡¯s Question! ? (Q)* Which Algorithm Should I Use? ? (A) The answer varies depending on many factors, including: ¡ô The size, quality, and nature of data ; ¡ô The available computational time; ¡ô The urgency of the task; and ¡ô What you want to do with the data(the problem). * towardsdatascience.com 12
Which algorithm should I use? ¡ô No one algorithm works best for every problem (Yes, not even neural networks!) 13
Important Concepts ? Model Selection: ¡ô K-crossfold validation ¡ô Train/Test/Evaluation Dataset ? Loss functions ? Convex Optimization ? Gradient Descent ? Model Complexity, Overfitting and Underfitting ? Regularization ? Training and Generalization Errors 14
Questions to ask when working on ML project! ? How much data do I have? What type/nature of data? ? How skilled and knowledgeable am I in this domain? ¡ô Will I be able to create more useful features from what I already have? ? How good am I in error analysis? 15
Questions to ask when working on ML project! ? Assumptions, Limitations and Adoption (ALA rule) of the algorithm. ¡ô Linear Regression (linear relationship, no or little multicollinearity, etc.) ¡ô Why does this particular loss function make sense? ? How good am I in debugging the chosen learning algorithm? 16
Effective Machine Learning ? Reduce time spent in programming (more experiments in short time) ¡ô Use off the shelf tools ? Customize and Scale Products ¡ô Start simple, scale as needed (again, choice of relevant toolsets) ? Think like a Scientist ¡ô Use statistics, not logic, to make decisions from the real world observations * ºÝºÝߣ content referred from Google¡¯s Machine Learning Crash Course 17
Thank You Ashish Sharma Email: accssharma@gmail.com /in/accssharma @accssharma AI Developers, Boise: https://github.com/aidevelopersboise/ai6-boise-materials HomeCU is hiring Software Engineers and Mobile Developers. https://www.homecu.net/company-jobs.html 18
Visual Demonstrations ? K nearest neighbor: http://vision.stanford.edu/teaching/cs231n-demos/knn/ ? CIFAR 10 Image Classification: https://cs.stanford.edu/people/karpathy/convnetjs/demo/cifar10.html 19

More Related Content

A General Overview of Machine Learning

  • 1. A General Overview of Machine Learning Boise Data Science Meetup -- September 18, 2018 Ashish Sharma ? Software Systems Engineer -- HomeCU, LLC. (2017 - present) ? Founder -- AI Developers, Boise ? City Ambassador -- AI Saturdays (global initiative of nurture.ai) ? Alumnus -- Boise State University (MS in Computer Science, 2015-2017) 1
  • 2. Overview ¡ñ AI and Applications ¡ñ Intro to Machine Learning ¡ñ Types of Machine Learning ¡ñ Which algorithm should I use? ¡ñ Effective Machine Learning Image Source: Cousins of Artificial Intelligence -- Towards Datascience 2
  • 3. AI Resurgence ? Computational Power (GPUs, cloud computing, distributed systems) ? Availability of large amount of Data (eg. Imagenet) ? Better theoretical understanding of the underlying techniques/algorithms ? Open and easily accessible research culture in academia and industry (NIPS, ICML, archiv.org) 3
  • 4. AI Resurgence (contd..) ? Netflix Challenge (2009) $1 Million Prize (User ratings for films) ? Kaggle (2010) (over more than a million users today) ? Fei-Fei Li and team at Stanford open sourced ImageNet (2008-2010) ¡ô Imagenet Large Scale Visual Recognition Challenge (ILSVRC) ? Geoffrey Hinton¡¯s Deep Learning Team wins ImageNet 2012 (Alexnet) 4
  • 5. Common Applications ? Speech recognition (virtual assistants) ? Advanced machine translation and natural language intelligence ? Strategic gaming algorithms (AlphaGo, chess) ? Computer Vision (image classification and object detection) ? Autonomous Vehicles ? Manufacturing Companies (landing.ai) ? Healthcare (Google¡¯s research on diabetic retinopathy -- with F-score of 0.95, surpassing the accuracy of 8 expert ophthalmologists) 5
  • 6. Machine Learning ? Form of applied statistics with emphasis on the use of computers to learn complex mathematical functions. ? More formally, ¡°A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E.¡± Image Source: xkcd 6
  • 7. Types of Machine Learning ? Supervised Learning ? Unsupervised Learning ? Reinforcement Learning 7
  • 8. Supervised Learning Terminologies: ? Input variable(s) ¡ô independent variable(s) ¡ô feature(s)/characteristic(s) of a single input object ¡ô Numerical -- continuous ( height, area of house) , discrete (grades, age) ¡ô Categorical (race, sex) -- nominal, ordinal ? Target variable(s) ¡ô Dependent variable(s), number/vector (eg. price of house, patient is diabetic, etc.) 8
  • 9. Supervised Learning ? Function approximation ¡ô Mathematically: solve for coefficient(s) of a function ¡ô Search for a best performing model from a hypothesis space. ¡ô Make predictions based on historical (labeled) data ? Regression (predict continuous target variable) ¡ô Univariate Regression (1 input variable, 1 output variable) ¡ô Multiple Regression (>=2 input variables, 1 output variable) ¡ô Multivariate Regression (>=2 output variable) ? Classification (predict discrete/categorical target variable) ¡ô Email: Spam or not? ¡ô Is this image a dog or cat? 9
  • 10. Unsupervised Learning ? Unsupervised Learning ¡ô Find hidden patterns and draw inference from (unlabeled) data ¡ô Essential for preliminary data analysis and visualization ? Clustering (grouping of similar data points) ¡ô K-Means, DBSCAN ? Dimensionality Reduction ¡ô Principal Components Analysis ¡ô Autoencoders 10
  • 11. Reinforcement Learning ? AI, Animal Psychology, Control Theory ? Agents, Actions, Environment, Change in State, Reward/Punishment ? Eg. Deep Attari: ¡ô Input: Snapshots of Attari board images (State and Actions) ¡ô Algorithm: Convolutional NNs with no pooling ¡ô Output layer: tailored for regression score (Maximize Reward) 11
  • 12. Beginner¡¯s Question! ? (Q)* Which Algorithm Should I Use? ? (A) The answer varies depending on many factors, including: ¡ô The size, quality, and nature of data ; ¡ô The available computational time; ¡ô The urgency of the task; and ¡ô What you want to do with the data(the problem). * towardsdatascience.com 12
  • 13. Which algorithm should I use? ¡ô No one algorithm works best for every problem (Yes, not even neural networks!) 13
  • 14. Important Concepts ? Model Selection: ¡ô K-crossfold validation ¡ô Train/Test/Evaluation Dataset ? Loss functions ? Convex Optimization ? Gradient Descent ? Model Complexity, Overfitting and Underfitting ? Regularization ? Training and Generalization Errors 14
  • 15. Questions to ask when working on ML project! ? How much data do I have? What type/nature of data? ? How skilled and knowledgeable am I in this domain? ¡ô Will I be able to create more useful features from what I already have? ? How good am I in error analysis? 15
  • 16. Questions to ask when working on ML project! ? Assumptions, Limitations and Adoption (ALA rule) of the algorithm. ¡ô Linear Regression (linear relationship, no or little multicollinearity, etc.) ¡ô Why does this particular loss function make sense? ? How good am I in debugging the chosen learning algorithm? 16
  • 17. Effective Machine Learning ? Reduce time spent in programming (more experiments in short time) ¡ô Use off the shelf tools ? Customize and Scale Products ¡ô Start simple, scale as needed (again, choice of relevant toolsets) ? Think like a Scientist ¡ô Use statistics, not logic, to make decisions from the real world observations * ºÝºÝߣ content referred from Google¡¯s Machine Learning Crash Course 17
  • 18. Thank You Ashish Sharma Email: accssharma@gmail.com /in/accssharma @accssharma AI Developers, Boise: https://github.com/aidevelopersboise/ai6-boise-materials HomeCU is hiring Software Engineers and Mobile Developers. https://www.homecu.net/company-jobs.html 18
  • 19. Visual Demonstrations ? K nearest neighbor: http://vision.stanford.edu/teaching/cs231n-demos/knn/ ? CIFAR 10 Image Classification: https://cs.stanford.edu/people/karpathy/convnetjs/demo/cifar10.html 19