15-288 - Spring 2021

Machine Learning
in a Nutshell




Key Information

Lectures: MW 10:30am - 11:50am - Online

Labs/Recitations: R 4:30pm - 5:50pm, Room 1030

9.0

35% In-class assessments (Quizzes, Labs), 35% Homework, 30% Project (Two Tasks)

15-112 passed with a C or a higher letter grade


Overview

This course is about the application of machine learning (ML) concepts and models to solve challenging real-world problems.
The emphasis of the course is on the methodological and practical aspects of designing, implementing, and using ML solutions.
Course topics develop around the notion of ML process pipeline, that identifies the multi-staged process of building and deploying an ML solution. An ML pipeline includes:

  • definition of the problem, objectives, and performance metrics;
  • collection and management of relevant operational data;
  • data wrangling (transforming, cleaning, filtering, scaling);
  • perform feature engineering on the available data in terms of feature selection, feature extraction, feature processing;
  • selection of appropriate ML models based on problem requirements and available data;
  • implementation, application, testing, and evaluation of the selected model(s);
  • deployment of the final ML model.

The process proceeds both forward and backward, iterating each stage until a satisfactory solution model is built.
The workflow of an ML pipeline is illustrated in the figure below (source: Practical ML with Python).

The course tackles all the stages of the ML pipeline, presenting conceptual insights and providing algorithmic and software tools to select and implement effective ways of proceeding and dealing with the challenges of the different stages.
The python ecosystem for data science and ML pandas, numpy, matplotlib, scikit-learn, keras, notebooks is introduced and used to retrieve, store, manipulate, visualize, and perform exploratory analysis of the data.

Course workflow:

The first part of the course addresses the data part of the pipeline, from data mining and collection, to data filtering and processing, to feature engineering for different types of data (numeric, categorical, textual, image, temporal).
Next, unsupervised learning (UL) techniques are introduced to further operating on the data by learning effective representations, perform dimensionality reduction and data compression. UL techniques include: Clustering models, Principal Component Analysis (PCA), Autoencoders.
Moving from data to techniques for classification and regression, a number supervised ML models are presented, including:

  • Decision Trees,
  • k-Nearest Neighbors,
  • Naive Bayes,
  • Logistic Regression,
  • Support Vector Machines (SVMs),
  • Least Squares Linear Regression,
  • Regularization,
  • Feature maps,
  • Kernelization,
  • Deep / Convolutional Neural Networks.

The different models are introduced by a conceptualization of the main underlying ideas and by providing the algorithmic and software tools necessary to experiment with the model on different datasets.
A discussion of the aspects of generalization, bias and variance, model evaluation and selection using cross-validation techniques, completes the ML pipeline.
The different techniques are tested and evaluated in problem scenarios from different domains and based on different data types. Selected problem domains include: natural language processing, machine vision, financial forecasting, logistics, production planning, diagnosis and prediction for bio-medical data.

Learning Objectives

Students who successfully complete the course will have acquired a general knowledge of the main concepts and techniques of data science and ML, and will be adept to use ML in different fields of application.
The course will provide the students with a toolkit of different skills needed to effectively go through the entire ML pipeline. The students will acquire conceptual and practical knowledge about:

  • collecting, handling, exploring, and wrangling data in different formats and originating from different sources;
  • selecting, extracting and engineering data features using both manual and learning techiques;
  • identifying the most appropriate ML techniques for the problem and the data at hand;
  • implementing and using a set of core ML models;
  • testing and evaluating ML models;
  • using the Python ecosystem for ML and data science;
  • applying ML to problems from a range of different application domains.

Course Layout

  • Course is based on two lectures per week where the different problems, solution models, and algorithms are formally introduced. The introduction of a new concept is always accompanied by the presentation of practical use cases.

  • Each week, a third class is used as a laboratory or for recitation. Laboratory classes let the students aswering graded assignments that require both programming hands-on and conceptual understanding of course subjects. Recitation classes are aimed to revise the concepts introduced in the lecture classes and profile the use of the different software tools.

Prerequisites

Having passed 15-112 with a C (minimum).

The basic notions of linear algebra, calculus, and probability theory that are necessary for the understanding of the formal concepts will be explained assuming no or little previous knowledge.

Assignments and Grading

  • Laboratory assessments: Students take laboratory classes (LabTests) where they have to answer questions involving both programming hands-on and conceptual aspects. LabTests will be typically done weekly, with a few exceptions and will require to prepare a Jupyter notebook integrating code, plots, and discussions.
  • Homework: Outside of the classroom, students practice with four homework consisting of programming tasks integrated in Jupyter notebooks. In the homework, students implement and experiment with the different algorithmic solutions, are confronted with different types of data, answer to conceptual questions, learn how to present material and results combining text, data, images, and code.
  • Project: Students have to complete a project that addresses the full ML pipeline and iteration cycle. Project work is staged in two tasks and comprises four deliverables. The final report is in the form of a Jupyter notebook implementing a Query Answering Machine for an application domain selected by the student. The project is done in small groups and the results are presented and discussed at the end of the course.

Grade: 35% Laboratory Assessments, 35% Homework, 30% Project

Readings

In addition to the lecture handouts and python notebooks (that will be made available after each lecture), during the course additional material will be provided by the instructor to cover specific parts of the course.

A number of (optional) textbooks can be consulted to ease the understanding of the different topics (the relevant chapters will be pointed out by the teacher), these include, but are not restricted to:

  • Machine Learning, Tom Mitchell (in the library)
  • Machine Learning: The Art and Science of Algorithms that Make Sense of Data, P. Flach
  • A Course in Machine Learning, Hal Daume', available online

Schedule



Date Topics Handouts References
1/18 General concepts, ML pipeline: Machine learning for data-driven decision making, extracting information from data, finding structures; basic ML concepts and applications; ML pipeline: from data sources to final model learning and deployment; course information and organization pdf
1/20 ML tasks and application problems: Taxonomy of ML tasks and problems; Supervised Learning (Classification, Regression); feature spaces; geometric view; workflow of SL; Unsupervised Learning (Finding patterns and relations, clustering, compression, dimensionality reduction); Reinforcement Learning (Sequential decision-making); advantages and issues learning with a teacher / supervisor; data labels and error quantification; preparing a labeled dataset; practical examples. pdf
1/21 Introduction to python's ecosystem for ML: Introduction to the use of Jupyter Notebooks; format of and typical opearations on ML datasets; CSV format for tabular data; basic concepts behind numpy arrays for vector/matrix operations; pandas data frames; basic pandas methods for reading and inspecting data; first examples of data display using matplotlib.
LabTest 1: Basic concepts of ML, use of Jupyter Notebooks and Python tools (15 minutes)
Notebook

1/25 SL task flow and design choices, model hypothesis, loss functions: A complete example of SL task flow; design choices: hypothesis class, parametric model functions; design choices: how to evaluate a model, loss functions; examples and properties of basic loss functions for classification and for regression. pdf
1/27 Empirical and Generalization errors, Canonical SL problem, ML workflow: Optimization problem for SL; empirical error; model complexity and overfitting; examples using regression; expected generalization (out-of-sample) error; validation sets and estimation of the generalization error; canonical SL problem; SL workflow; building a model in the ML pipeline; promoting generalization. pdf
1/28 LabTest 2: Core ML concepts, Practice with python's tools

2/1 From data to models, a complete regression example: A complete, step-by-step example of how to proceed in practice from the data available in a dataset file and a general problem statement to the definition and validation of sound regression models: data ingestion, data preparation, EDA, model hypotheses, loss function, model testing, looping over the models, model selection. Pandas, numpy, scipy, and matplot software tools at work. Notebook Web server traffic dataset
2/3 Regression with linear models, OLS: Regression and linear regression concepts and models; linearity of models in prediction and in training; role of feature weights; Ordinary Least Squares (OLS); analytic solution of OLS using linearity in coefficients, partial derivatives and linear system of equations; matricial form; solution using sklearn methods; issues with matrix inversion; numeric examples with generation of instances.
OPTIONAL: rank of a matrix and related concepts of linear algebra, using numpy matrix manipulation methods for solving OLS.
Notebook Notebook with optional material on matrix rank and singularity issues
2/4 LabTest 3: Regression pipeline

2/8 Classification tasks, k-NN classifier: Example of classification task using the Iris dataset; use of scikit-learn datasets; data visualization; analysis by visual inspection; ML classification using k-Nearest Neighbors; general concepts behind k-NN; decision regions and decision boundaries; effect of k on the classification boundaries. Notebook
2/10 k-NN for classification tasks: Use of k-NN for classification tasks; effect of k on empirical error; plotting decision regions with meshgrids; use of scikit-learn methods (.fit(), .predict()); measure performance; finding the best k? Notebook
2/11 LabTest 4: Classification pipeline using k-NN

2/15 Model Validation and Model Selection: Model validation and expected generalization error; hold-out and cross-validation methods for validation and for model selection; general operational scheme for using a dataset (training, validation, testing, model optimization and selection); sciki-learn methods and iterators for model selection, dataset splitting, CV-based grid search for parameter setting. Notebook
2/17 Data cleaning, Missing values: Introduction to data wrangling; data cleaning: dealing with missing data; types of missingness: MCAR, MAR, MNAR; general strategies, pros and cons; discard data: listwise, pairwise, dropping; imputation techniques: statistical estimates, common point, frequent category, category adding, arbritrary value, adding variable, random sampling, multiple imputation, last and next observation in time series, interpolation in time series, predictive models; sklearn methods for data removal and imputation. Notebook Diabetes, corrupted dataset
2/18 Dealing with Outliers: Concept of outlier in data; reasons for an outlier; types of outliers (global, contextual, collective); detection of outliers; removing or keeping? parametric vs. non-parametric statistics; parametric approaches: Gaussians, 3σ rule of thumb, z-scores, univariate vs. multivariate; non-parametric approaches: median and quantiles, IQR, box plots, 1.5·IQR rule of thumb. Notebook

2/22 Scaling data: need for scaling / normalizing features; standardization (z-transform); scaling to a range; normalization; robuts scaling; sklearn methods and classes for scaling. Notebook
2/24 Feature engineering 1: Correlations, polynomial transformations: correlations among data; use of correlation analysis for selecting/removing features; correlation coefficients; correlation matrix and heatmap visualization; case studies, application to regression problems; beyond linear features: interaction among features; why and how to define feature interaction; linear (regression) models using non-linear features; polynomial feature transformations: advantages and limits; sklearn methods for polynomial transformations; first use of pipeline methods for automating transformation and fitting processes. general concepts on the utility of transforming the linear features into high(er) dimensional feature spaces. Notebook
2/25 LabTest 5: Numeric feature engineering, feature transformations, correlations

3/1 Spring break
3/3 Spring break
3/4 Spring break

3/8 Feature engineering 2, Types of data, Numeric data: Notion and importance of feature enginering; different feature data types; pandas and numpy methods to deal with data types; from categorical data to ordinal data and to numeric values; numeric data types: values, counts, frequencies, percentages; binarization and rounding feature transformation and creation; features transformation to account for interaction (check last lecture). Notebook Pokemon games dataset
Song views dataset
Item popularity dataset
3/10 Feature engineering 3, Image data: Properties of image data; RGB enconding; skimage methods for handling and processing images; raw pixel intensities as features; grayscale transformation; feature extraction by binning, properties and issues using histograms in rgb and grayscale domains, pie charts; features extracted by aggregation statistics. Notebook Cat image
Dog image
Panda image
Sea image
Another cat image
Coder survey dataset
3/11 Feature engineering 4, Image data:Edge extraction with Canny's algorithm; Gaussian filters; Edge operators: Sobel and Roberts filters; gradients and derivatives, image gradients, function gradients; Histogram of Oriented Gradients (HOG) as image feature descriptor; step-by-step process for computing the HOG; skimage methods; use of HOG in image classification and object detection. Notebook Sliding window image
Sitting dog image

3/15 Feature engineering 5, Image data: Convolution operator; convolution examples: sharpening, embossing, blurring / smoothing, Gaussian smoothing; combination of filters; Laplacian filter for computing 2nd order derivatives; non-linear filters, median example; localized feature extraction: SIFT, SURF, ORB; main ideas and processing flow in SIFT; example of use with OpenCV; example of use of ORB with skimage. Notebook Lena in grayscale
Lena, salt & pepper noise
Colosseum image
3/17 Feature engineering 6, Image data, PCA (Unsupervised learning 1): Unsupervised learning tasks; high-dimensionality of data and related issues; need for dimensionality reduction; feature extraction by identifying latent features; dimensionality reduction and compression; Principal Component Analysis (PCA); PCA for learning representations; PCA for dimensionality reduction; key ideas: linearity, directions of maximal variance; mathematical details (optional); PCA at work; limitations; application to image data using sklearn; use of PCA to extract image features in classification. PDF
Notebook
Andrew Ng notes on Principal Component Analysis
3/18 Clustering 1 (Unsupervised learning 2): Clustering ideas and models: partitional, hierarchical, hard vs. soft clustering; similarity measures; partitional clustering and K-Means problem; naive K-means algorithm; phases of the algorithm; linear cluster boundary regions and Voronoi tasselation of feature space; K-Means at work; impact of initial cluster centers; converence and complexity properties; use of sklearn for K-Means clustering, with application examples. PDF
Notebook
Mall customers
Daume', Chapter 15.2

3/22 Clustering 2, Properties of K-Means: K-Means examples; generating synthetic datasets; step-by-step K-Means; K-Means as an instance of Expectation-Maximization; assumptions and limitations of K-Means; ideal data for K-Means: balanced, spherical cluster of data generated by Gaussian populations with equal variances; number of clusters: effects of wrong choices, how to select k (elbow method); local minima in the distortion function; minimization of the Euclidean distance and linear cluster boundaries; Voronoi tassellation of the feature space; effects of imbalanced cluster data; effects of unequal variances and presence of covariances; computational issues; comparison with other algorithms Notebook
3/24 Clustering 3, K-Means for image data, classification metrics: Signal compression and Vector Quantization (VQ); relationship between K-Means and VQ; use of K-Means / VQ for image compression in the color space; cluster centers as vector prototypes; data visualization in the RGB space; histogram information; compression ratio; prototype / unsupervised classification of a dataset; NIST image dataset; comparison with supervised classification; confusion matrix; performance maeasures for classification: confusion matrix, rate/accuracy, recall ratio, precision, F-measure; clustering for image segmentation; use of image segmentation; phases of image segmentation; examples. Notebook
3/25 Feature Engineering, Time series data: Examples and properties of time series; time series and supervised learning: the need for defining good features; date time features, pandas methods; time lagged features, notion of autocorrelation, pandas methods; features based on rolling window statistics; features based on expanding window statistics; linear regression on engineered features; cross-validation for time series, sklearn methods for splitting and cross-validating the dataset; components of a time series: level, trend, seasonality, random residuals; importance of stationarity; subtraction of components to get stationarity; techniques to check stationarity; STL method from statsmodels; statistical tests. Notebook Daily min temperatures
Train passengers
Airline passengers

3/29 Decision-trees 1: Supervised Learning and Query Answering Machines; learning and posing/answering questions; defining sequences of questions; decision trees; divid-and-conquer concept; properties and structure of a decision tree; function represention by DT; example of boolean functions; intractability of exahustive search; NP-hardness of finding the optimal decision tree; axis-parallel decision boundaries; overfitting. PDF
3/31 Decision-trees 2: Construction of a DT; decision stumps; recursive procedure; effects and goals of attribute splitting; purity and uncertainty; greedy, top-down heuristics; ID3; selection of best attributes based on different criteria; entropy of a random variable; entropy as a measure of purity of a labeled set; information gain; numeric examples. PDF
4/1 Decision-trees 3: Computing information gains; properties if ID3 ID3; overfitting issues; pruning approaches, C4.5; dealing with continuous attributes; thresholding with binary branching; axis-parallel boundaries; regression with decision trees; purity of a set: discrete vs. continuous labels; use of variance / standard deviation as measure of purity; prediction in regression scenarions; (extras) other measures of dispersion / purity in labeled sets (e.g., Gini index); criteria to decide to stop splitting a partition; examples of regression trees vs. max depth constraint; practice problems; notebook with skelearn methods for decision trees for classification and regression, visualization of the tree and of the decision regions, cross-validated model selection based on multiple input parameters. PDF
Notebook
Pima dataset

4/5 Ensemble models, Bagging, Boosting, Random forests: General ideas behind combining models; voting/averaging vs. stacking models; bagging and boosting as forms of combining different experts; bagging: construction of the datasets by bootstrapping, properties of the base model, variance reduction goals, aggregation by averaging; random forests as bagging with randomization of the features of each model; boosting: sequential generation of the weighted datasets, base model as a weak learner, goals of combining multiple weak learners, how to compute voting weights; sklearn methods for ensemble models. PDF Notebook
4/7 Linear models, Support Vector Machines (SMV) 1: General form and properties of linear models for classification and for regression; formulation of a linear model, bias, scalar product; basic geometrical properties; linear decision boundaries; from a linear function to discrete labels for classification; feature transformations and linear models; score and behavior of a linear classifier; notion of classifier margin; SVMs as max-margin classifiers; functional and geometric margin; classifier margin; SVM optimization problem; SVM hard-margin formulation and properties; support vectors. PDF
4/8 Support Vector Machines 2, Kernelization: Support vectors; formulation of primal and dual problems and their relation; soft-margin SVM for non linearly separable data; slack variables and penalty factor; support vectors in soft-margin SVM; Support Vector Regression (SVR); general concepts and basic formulation; loss function; ideas behind kernelization; kernelization at work in classification tasks; effect of using different kernels on the resulting non-linear decision boundaries; methods from scikit-learn; functions for problem generation, visualization, and analysis; hard-margin and soft-margin SVM at work; inspection of support vectors; SVC vs. other classifiers; performance comparison and decision boundaries; SVR at work: linear vs. non-linear kernels; effect of parameters. PDF Notebook

4/12 Neural networks 1; LabTest 6, Question Mix: From linear classifiers to the Perceptron model; abstraction of neuron processing, threshold models; Perceptron algorithm: iterative adjustments of weights, use of gradients to miminize quadratic loss; from single perceptrons to multi-layer perceptrons (MLP); neural networks as non-linear parametric models for function approximation; perceptron units with non-linear activations; hidden and visible layers; feed-forward (FF) multi-layer architectures; activation functions; hidden layer as feature extractor / feature map; NN and automatic feature learning through the hidden layers. PDF
4/14 Neural networks 2: Representation and visualization of the non-linear function encoded by a network; loss function and NN optimization problem; basic concepts behind gradient descent; full, stochastic, and batch gradient descent; idea of backpropagation for computing the gradients; choice of the activation function and optimization problem; design choices; overfitting issues; general theoretical properties. PDF
4/15 Neural networks 3: Use of keras for implementing and testing neural networks: creating a sequential layout, compile a model, test and visualize learning evolution, inspect/visualize the activations from the layers and relate to feature extractions; MLP examples with numeric and image data Notebook

4/19 Neural networks 4, Deep learning models: MLPs vs. Convolutional Neural Networks (CNNs); issues with fully connected networks; core reasons behind the success of CNNs; recap of convolution operator and SIFT feature extraction in images; role and rationale behind convolutional and pooling layers; locality of processing, hierarchical dimensionality reduction; number of trainable parameters; typical CNN architectures: sequence of (convolution filters, activations, pooling); constructing convolutional layers; role of stride; feature maps and filter banks; constructing pooling layers; max pooling; soft-max output layer; examples; visualization of features extracted at the different layers; notes about optimization, transfer learning, autoencoders; keras for CNNs (check the notebook!). PDF Notebook
4/21 LabTest 7: Neural networks
4/22 Regularization techniques: Explicit feature control and selection for minimizing risk of overfitting; implicit feature control and selection using regularized loss functions; bias-variance of a model; effect of large weights on variance / overfitting; regularized loss function; use of L-norms; Ridge regression; Lasso regression; Ridge regression and Lasso regression at work; comparative analysis over a number of test scenarios with linear and polynomial features; Elastic Net regression; a rel-world regression scenario from data wrangling to model selection. Notebook Mart sales dataset

Homework Assignments

Topic Files Due Dates
Homework 1: A Classification task: from data ingestion to model selection and testing
Homework 2: Data cleaning and model selection for regression tasks
Homework 3: Feature engineering and model selection for supervised image classification
Homework 4: Unsupervised image classification, Time series analysis


LabTests

Topic Files Due Dates
LabTest 1: Basic concepts of ML, use of Jupyter Notebooks and of Python tools
LabTest 2: Core ML concepts, Practice with python's tools
LabTest 3: Regression pipeline
LabTest 4: Classification pipeline using k-NN
LabTest 5: Numeric feature engineering, feature transformations, correlations
LabTest 6: Question Mix on decicsion trees, ensemble methods, support vector machines
LabTest 7: Neural networks


Project

A Query Answering Machine (QuAM)

Deliverables Files Due Dates
D1 [Report]: Initial Proposal and Dataset
D2 [Dataset and Notebook]: Final Dataset
D3 [Software and Notebook]: Query Answering Machine (QuAM)
D4 [Presentation]: Final report


Policies for Assignments

  • Homework is due on Gradescope by the posted deadline. Assignments submitted past the deadline will incur the use of late days.

  • You have 6 late days in total, but cannot use more than 2 late days per homework. No credit will be given for homework submitted more than 2 days after the due date. After your 6 late days have been used you will receive 20% off for each additional day late.

  • You can discuss the exercises with your classmates, but you should write up your own solutions. If you find a solution in any source other than the material provided on the course website or the textbook, you must mention the source.

  • In general, for all types of assignments and tests, CMU's directives for academic integrity apply and must be duly followed.

Office Hours

Name Email Hours Location
Gianni Di Caro gdicaro@cmu.edu By appointment / Zoom M 1007
Eduardo Feo-Flushing efeoflus@andrew.cmu.edu TBD M 1009
Mohammad Shahmeer Ahmad