Transcript Slides
m Machine Learning F# and Accord.net Alena Hall • Software architect, MS in Computer Science • Member of F# Software Foundation Board of Trustees • Researcher in the field of mathematical theoretical abstractions possible in modern programming concepts @lenadroid • Speaker and active software engineering community member Machine Learning Questions • Why machine learning? • What is the data? • How? Data Questions. Data reality :\ Path to grasping machine learning and data science… Contents • • • • Multiple Linear Regression Logistic Regression Classification K Means Clustering What’s next? F# for machine learning and data science! Why F#? 1. Exploratory programming, interactive environment 2. Functional programming, referential transparency 3. Data pipelines 4. Algebraic data types and pattern matching 5. Strong typing, type inference, Type Providers 6. Units of measure 7. Concurrent, distributed and cloud programming Data pipelines Algebraic data types // Discriminated Union Pattern matching Type Providers Units of measure Linear Regression How to predict? 1. Make a guess. 2. Measure how wrong the guess is. 3. Fix the error. Make a guess! MATH Make a guess? What does it mean?... Hypothesis /guess : weights Find out our mistake… Cost function/ Mistake function: … and minimize it: Mistake function looks like… Global minimums How to reduce the mistake? Update each slope parameter until Mistake Function minimum is reached: Alpha Learning rate Simultaneously Derivative Direction of moving Fix the error Multiple Linear Regression X [ ] – Predictors: Statistical data about bike rentals for previous years or months. Y – Output: Amount of bike rentals we should expect today or some other day in the future. * Y is not nominal, here it’s numerical continuous range. Make a guess! Fix the error “Talk is cheap. Show me the code.” Multiple linear regression: Bike rentals demand What to remember? Linear Regression 1. Simplest regression algorithm 2. Very fast, runs in constant time 3. Good at numerical data with lots of features 4. Output from numerical continuous range 5. Linear hypothesis 6. Uses gradient descent Logistic Regression Hypothesis function Estimated probability that Y = 1 on input X Mistake function Mistake function is the cost for a single training data example h(x) Full mistake function 1. Uses the principle of maximum likelihood estimation. 2. We minimize it same way as with Linear Regression “Talk is cheap. Show me the code.” Logistic Regression Classification Example What to remember? Logistic Regression • Classification algorithm • Relatively small number of predictors • Uses logistics function for hypothesis • Has the cost function that is convex • Uses gradient descent for correcting the mistake At this point… Machine Learning What society thinks I do… What other programmers think I do… What I really do is… K-Means Clustering What’s next? I’m Lena @lenadroid Thank you! What if it doesn’t work? Algorithm debugging tips • • • • • • Try more data Try more features Try less features Try feature combinations Try polynomial features … What else can go wrong? Ideally... the hypothesis will … just fit the model Underfitting … Overfitting Try out different values for the regularization parameter. • Regularization…? • Too big regularization parameter? -> underfitting - the line is over-smoothed • Too small regularization parameter? -> overfitting - too optimized for train data