Machine Learning Guide

Try a walking desk to stay healthy while you study or work!

Full notes at ocdevel.com/mlg/13

Support Vector Machines (SVM)

• Purpose: Classification and regression.
• Mechanism: Establishes decision boundaries with maximum margin.
• Margin: The thickness of the decision boundary, large margin minimizes overfitting.
• Support Vectors: Data points that the margin directly affects.
• Kernel Trick: Projects non-linear data into higher dimensions to find a linear decision boundary.

Naive Bayes Classifiers

• Framework: Based on Bayes' Theorem, applies conditional probability.
• Naive Assumption: Assumes feature independence to simplify computation.
• Application: Effective for text classification using a "bag of words" method (e.g., spam detection).
• Comparison with Deep Learning: Faster and more memory efficient than recurrent neural networks for text data, though less precise in complex document understanding.

Choosing an Algorithm

• Assessment: Evaluate based on data type, memory constraints, and processing needs.
• Implementation Strategy: Apply multiple algorithms and select the best-performing model using evaluation metrics.

Links

• Andrew Ng Week 7
• Pros/cons table for algos
• Sci-Kit Learn's decision tree for algorithm selection.
• Machine Learning with R book for SVMs and Naive Bayes.
• "Mathematical Decision-Making" great courses series for Bayesian methods.

Podcast is still alive, working on next episode - sorry for the delay!

Sorry guys, had an intense last 3 weeks. Mid-move to Portland OR my employer folded so I had to hunt, stat (new expenses & all). Luckily I found a job in NLP, my favorite branch of ML! Big-time silver lining. BTW, everyone these days does whiteboarding. Used to be the top-dogs, now it's everyone. If you're ever on the hunt, read CtCI first (http://amzn.to/2nu6ya1). Essential.

Try a walking desk to stay healthy while you study or work!

Full notes at ocdevel.com/mlg/12

Topics

• Shallow vs. Deep Learning: Shallow learning can often solve problems more efficiently in time and resources compared to deep learning.

• Supervised Learning: Key algorithms include linear regression, logistic regression, neural networks, and K Nearest Neighbors (KNN). KNN is unique as it is instance-based and simple, categorizing new data based on proximity to known data points.

• Unsupervised Learning:

  • Clustering (K Means): Differentiates data points into clusters with no predefined labels, essential for discovering data structures without explicit supervision.
  • Association Rule Learning: Example includes the a priori algorithm, which deduces the likelihood of item co-occurrence, commonly used in market basket analysis.
  • Dimensionality Reduction (PCA): Condenses features into simplified forms, maintaining the essence of the data, crucial for managing high-dimensional datasets.

• Decision Trees: Utilized for both classification and regression, decision trees offer a visible, understandable model structure. Variants like Random Forests and Gradient Boosting Trees increase performance and reduce overfitting risks.

Links

• Focus material: Andrew Ng Week 8.
• A Tour of Machine Learning Algorithms for a comprehensive overview.
• Scikit Learn image: A decision tree infographic for selecting the appropriate algorithm based on your specific needs.
• Pros/cons table for various algorithms

Try a walking desk to stay healthy while you study or work!

Full notes at ocdevel.com/mlg/10

Topics:

• Recommended Languages and Frameworks:

  • Python and TensorFlow are top recommendations for machine learning.
  • Python's versatile libraries (NumPy, Pandas, Scikit-Learn) enable it to cover all areas of data science including data mining, analytics, and machine learning.

• Language Choices:

  • C/C++: High performance, suitable for GPU optimization but not recommended unless already familiar.
  • Math Languages (R, MATLAB, Octave, Julia): Optimized for mathematical operations, particularly R preferred for data analytics.
  • JVM Languages (Java, Scala): Suited for scalable data pipelines (Hadoop, Spark).

• Framework Details:

  • TensorFlow: Comprehensive tool supporting a wide range of ML tasks; notably improves Python's performance.
  • Theano: First in symbolic graph framework, but losing popularity compared to newer frameworks.
  • Torch: Initially favored for image recognition, now supports a Python API.
  • Keras: High-level API running on top of TensorFlow or Theano for easier neural network construction.
  • Scikit-learn: Good for shallow learning algorithms.

Comparisons:

• C++ vs Python in ML: C++ offers direct GPU access for performance, but Python streamlined performance with frameworks that auto-generate optimized C code.
• R and Python in Data Analytics: Python's Pandas and NumPy rival R with a strong general-purpose application beyond analytics.

Considerations:

• Python's Ecosystem Benefits: Single programming ecosystem spans full data science workflow, crucial for integrated projects.
• Emerging Trends: Keep an eye on Julia for future considerations in math-heavy operations and industry adoption.

Additional Notes:

• Hardware Recommendations:
  • Utilize Nvidia GPUs for machine learning due to superior support and integration with CUDA and cuDNN.
• Learning Resources:
  • TensorFlow's documentation and tutorials are highly recommended for learning due to their thoroughness and regular updates.
  • Suggested learning order: Learn Python fundamentals, then proceed to TensorFlow.

Links

• Other languages like Node, Go, Rust: why not to use them.
• Best Programming Language for Machine Learning
• Data Science Job Report 2017
• An Overview of Python Deep Learning Frameworks
• Evaluation of Deep Learning Toolkits
• Comparing Frameworks: Deeplearning4j, Torch, Theano, TensorFlow, Caffe, Paddle, MxNet, Keras & CNTK - grain of salt, it's super heavy DL4J propaganda (written by them)

Try a walking desk to stay healthy while you study or work!

Full notes at ocdevel.com/mlg/9

Key Concepts:

• Deep Learning vs. Shallow Learning: Machine learning is broken down hierarchically into AI, ML, and subfields like supervised/unsupervised learning. Deep learning is a specialized area within supervised learning distinct from shallow learning algorithms like linear regression.
• Neural Networks: Central to deep learning, artificial neural networks include models like multilayer perceptrons (MLPs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). Neural networks are composed of interconnected units or "neurons," which are mathematical representations inspired by biological neurons.

Unique Features of Neural Networks:

• Feature Learning: Neural networks learn to combine input features optimally, enabling them to address complex non-linear problems where traditional algorithms fall short.
• Hierarchical Representation: Data can be processed hierarchically through multiple layers, breaking down inputs into simpler components that can be reassembled to solve complex tasks.

Applications:

• Medical Cost Estimation: Neural networks can handle non-linear complexities such as feature interactions, e.g., age, smoking, obesity, impacting medical costs.
• Image Recognition: Neural networks leverage hierarchical data processing to discern patterns such as lines and edges, building up to recognizing complex structures like human faces.

Computational Considerations:

• Cost of Deep Learning: Deep learning's computational requirements make it expensive and resource-intensive compared to shallow learning algorithms. It's cost-effective to use when necessary for complex tasks but not for simpler linear problems.

Architectures & Optimization:

• Different Architectures for Different Tasks: Specialized neural networks like CNNs are suited for image tasks, RNNs for sequence data, and DQNs for planning.
• Neuron Types: Neurons in neural networks are referred to as activation functions (e.g., logistic sigmoid, relu) and differ based on tasks and architecture needs.

Mathematics essential for machine learning includes linear algebra, statistics, and calculus, each serving distinct purposes: linear algebra handles data representation and computation, statistics underpins the algorithms and evaluation, and calculus enables the optimization process. It is recommended to learn the necessary math alongside or after starting with practical machine learning tasks, using targeted resources as needed. In machine learning, linear algebra enables efficient manipulation of data structures like matrices and tensors, statistics informs model formulation and error evaluation, and calculus is applied in training models through processes such as gradient descent for optimization.

Links

• Notes and resources at ocdevel.com/mlg/8
• Try a walking desk - stay healthy & sharp while you learn & code

Come back here after you've finished Ng's course; or learn these resources in tandem with ML (say 1 day a week).

Recommended Approach to Learning Math

• Direct study of mathematics before beginning machine learning is not necessary; essential math concepts are introduced within most introductory courses.
• A top-down approach, where one starts building machine learning models and learns the underlying math as needed, is effective for retaining and appreciating mathematical concepts.
• Allocating a portion of learning time (such as one day per week or 20% of study time) to mathematics while pursuing machine learning is suggested for balanced progress.

Linear Algebra in Machine Learning

• Linear algebra is fundamental for representing and manipulating data as matrices (spreadsheets of features and examples) and vectors (parameter lists like theta).
• Every operation involving input features and learned parameters during model prediction and transformation leverages linear algebra, particularly matrix and vector multiplication.
• The concept of tensors generalizes vectors (1D), matrices (2D), and higher-dimensional arrays; tensor operations are central to frameworks like TensorFlow.
• Linear algebra enables operations that would otherwise require inefficient nested loops to be conducted quickly and efficiently via specialized computation (e.g., SIMD processing on CPUs/GPUs).

Statistics in Machine Learning

• Machine learning algorithms and error measurement techniques are derived from statistics, making it the most complex math branch applied.
• Hypothesis and loss functions, such as linear regression, logistic regression, and log-likelihood, originate from statistical formulas.
• Statistics provides both the probability framework (modelling distributions of data, e.g., housing prices in a city) and inference mechanisms (predicting values for new data).
• Statistics forms the set of "recipes" for model design and evaluation, dictating how data is analyzed and predictions are made.

Calculus and Optimization in Machine Learning

• Calculus is used in the training or "learning" step through differentiation of loss functions, enabling parameter updates via techniques such as gradient descent.
• The optimization process involves moving through the error space (visualized as valleys and peaks) to minimize prediction error, guided by derivative calculations indicating direction and magnitude of parameter updates.
• The particular application of calculus in machine learning is called optimization, more specifically convex optimization, which focuses on finding minima in "cup-shaped" error graphs.
• Calculus is generally conceptually accessible in this context, often relying on practical rules like the power rule or chain rule for finding derivatives of functions used in model training.

The Role of Mathematical Foundations Post-Practice

• Greater depth in mathematics, including advanced topics and the theoretical underpinnings of statistical models and linear algebra, can be pursued after practical familiarity with machine learning tasks.
• Revisiting math after hands-on machine learning experience leads to better contextual understanding and practical retention.

Resources for Learning Mathematics

• MOOCs, such as Khan Academy, provide video lessons and exercises in calculus, statistics, and linear algebra suitable for foundational knowledge.
• Textbooks recommended in academic and online communities cover each subject and are supplemented by concise primer PDFs focused on essentials relevant to machine learning.
• Supplementary resources like The Great Courses offer audio-friendly lectures for deeper or alternative exposure to mathematical concepts, although they may require adaptation for audio-only consumption.
• Audio courses are best used as supplementary material, with primary learning derived from video, textbooks, or interactive platforms.

Summary of Math Branches in Machine Learning Context

• Linear algebra: manipulates matrices and tensors, enabling data structure operations and parameter computation throughout the model workflow.
• Statistics: develops probability models and inference mechanisms, providing the basis for prediction functions and error assessments.
• Calculus: applies differentiation for optimization of model parameters, facilitating the learning or training phase of machine learning via gradient descent.
• Optimization: a direct application of calculus focused on minimizing error functions, generally incorporated alongside calculus learning.

The logistic regression algorithm is used for classification tasks in supervised machine learning, distinguishing items by class (such as "expensive" or "not expensive") rather than predicting continuous numerical values. Logistic regression applies a sigmoid or logistic function to a linear regression model to generate probabilities, which are then used to assign class labels through a process involving hypothesis prediction, error evaluation with a log likelihood function, and parameter optimization using gradient descent.

Links

• Notes and resources at ocdevel.com/mlg/7
• Try a walking desk - stay healthy & sharp while you learn & code

Classification versus Regression in Supervised Learning

• Supervised learning consists of two main tasks: regression and classification.
• Regression algorithms predict continuous values, while classification algorithms assign classes or categories to data points.

The Role and Nature of Logistic Regression

• Logistic regression is a classification algorithm, despite its historically confusing name.
• The algorithm determines the probability that an input belongs to a specific class, using outputs between zero and one.

How Logistic Regression Works

• The process starts by passing inputs through a linear regression function, then applying a logistic (sigmoid) function to produce a probability.
• For binary classification, results above 0.5 usually indicate a positive class (for example, "expensive"), and results below 0.5 indicate a negative class ("not expensive").
• Multiclass problems assign probabilities to each class, selecting the class with the highest probability using the arg max function.

Example Application: Housing Spreadsheet

• An example uses a spreadsheet of houses with features like square footage and number of bedrooms, labeling each as "expensive" (1) or "not expensive" (0).
• Logistic regression uses the spreadsheet data to learn the pattern that separates expensive houses from less expensive ones.

Steps in Logistic Regression

• The algorithm follows three steps: predict (infer a class), evaluate error (calculate how inaccurate the guesses were), and train (refine the underlying parameters).
• Predictions are compared to actual data, and the difference (error) is calculated via a log likelihood function, which accounts for how confident the prediction was compared to the true value.
• Model parameters (theta values) are updated using gradient descent, which iteratively reduces the error by adjusting these values based on the derivative of the error function.

The Mathematical Foundation

• The hypothesis function is the sigmoid or logistic function, with the formula: 1 / (1 + e^(-theta^T x)), where theta represents the parameters and x the input features.
• The error function (cost function) for logistic regression uses log likelihood, aggregating errors over all data points to guide model learning.

Practical Considerations

• Logistic regression finds a "decision boundary" on the graph (S-curve) that best separates classes such as "expensive" versus "not expensive."
• When the architecture requires a proper probability distribution (sum of probabilities equals one), a softmax function is applied to the outputs, but softmax is not covered in this episode.

Composability in Machine Learning

• Machine learning architectures are highly compositional, with functions nested within other functions - logistic regression itself is a function of linear regression.
• This composability underpins more complex systems like neural networks, where each "neuron" can be seen as a logistic regression unit powered by linear regression.

Building Toward Advanced Topics

• Understanding logistic and linear regression forms the foundation for approaching advanced areas of machine learning such as deep learning and neural networks.
• The concepts of prediction, error measurement, and iterative training recur in more sophisticated models.

Resource Recommendations

• The episode recommends the Andrew Ng Coursera course for deeper study into these concepts and details, especially for further exploration of multivariate regression and error functions.

People interested in machine learning can choose between self-guided learning, online certification programs such as MOOCs, accredited university degrees, and doctoral research, with industry acceptance and personal goals influencing which path is most appropriate. Industry employers currently prioritize a strong project portfolio over non-accredited certificates, and while master's degrees carry more weight for job applications, PhD programs are primarily suited for research interests rather than industry roles.

Links

• Notes and resources at ocdevel.com/mlg/6
• Try a walking desk - stay healthy & sharp while you learn & code

Learner Types and Self-Guided Education

• Individuals interested in machine learning may be hobbyists, aspiring professionals, or scientists wishing to contribute to research in artificial intelligence.
• Hobbyists can rely on structured resources, including curated syllabi and recommended online materials, to guide their self-motivated studies.
• The "Andrew Ng Coursera" course is frequently recommended as an initial step for self-learners, and advanced resources such as "Artificial Intelligence: A Modern Approach" and "Deep Learning" textbooks are valuable later.

MOOCs and Online Certificates

• MOOCs (Massive Open Online Courses) are widely available from platforms such as Coursera, Udacity, edX, and Khan Academy, but only Coursera and Udacity are commonly recognized for machine learning and data science content.
• Coursera is typically recommended for individual courses; its specializations are less prominent in professional discussions.
• Udacity offers both free courses and paid "nano degrees" which include structured mentoring, peer interaction, and project-based learning.
• Although Udacity certificates demonstrate completion and the development of practical projects, they lack widespread recognition or acceptance from employers.
• Hiring managers and recruiters consistently emphasize the value of a substantial project portfolio over non-accredited certificates for job-seekers.

University Degrees and Industry Recognition

• Master's degrees in machine learning or computer science remain the most respected credentials for job applications in the industry, with requirements often officially listed in job postings.
• The Georgia Tech OMSCS program provides an accredited, fully online Master's degree in Computer Science at a much lower cost than traditional programs, reportedly leveraging Udacity's course infrastructure.
• In some cases, a strong portfolio can substitute for formal educational requirements, particularly if the applicant demonstrates practical and scalable machine learning project experience.
• Portfolio strength is considered analogous to web development hiring, where demonstrated skills and personal projects can compensate for missing degree credentials.

PhD Pathways and Research Careers

• A PhD is generally unnecessary for industry positions in machine learning; a master's degree or an exceptional portfolio is usually adequate.
• Doctoral degrees are most useful for those seeking research roles or wishing to investigate complex theoretical questions in artificial intelligence, rather than working in standard industry applications.
• PhD programs pay a stipend to students, though the compensation is much less than typical industry salaries, which should factor into an individual's decision-making process.

Considerations and Resources

• Choosing an educational path depends on individual goals, available resources, and desired career trajectory; a portfolio of significant machine learning projects is universally beneficial regardless of the chosen route.
• Community discussions and recruiter perspectives suggest that practical skills, proven through real-world projects, are highly valued in addition to or in place of formal degrees.
• Interested individuals can review ongoing discussions and perspectives:
  • Self-Guided Data Science Guide (canyon289)
  • Hacker News – Are credentials required?
  • Cole MacLean's Self-Taught AI Blog
  • Hacker News: Self-Study Paths

Linear regression is introduced as the foundational supervised learning algorithm for predicting continuous numeric values, using cost estimation of Portland houses as an example. The episode explains the three-step process of machine learning - prediction via a hypothesis function, error calculation with a cost function (mean squared error), and parameter optimization through gradient descent - and details both the univariate linear regression model and its extension to multiple features.

Links

• Notes and resources at ocdevel.com/mlg/5
• Try a walking desk - stay healthy & sharp while you learn & code
• Generate a podcast - use my voice to listen to any AI generated content you want

Linear Regression Overview of Machine Learning Structure

• Machine learning is a branch of artificial intelligence, alongside statistics, operations research, and control theory.
• Within machine learning, supervised learning involves training with labeled examples and is further divided into classification (predicting discrete classes) and regression (predicting continuous values).

Linear Regression and Problem Framing

• Linear regression is the simplest and most commonly taught supervised learning algorithm for regression problems, where the goal is to predict a continuous number from input features.
• The episode example focuses on predicting the cost of houses in Portland, using square footage and possibly other features as inputs.

The Three Steps of Machine Learning in Linear Regression

• Machine learning in the context of linear regression follows a standard three-step loop: make a prediction, measure how far off the prediction is, and update the prediction method to reduce mistakes.
• Predicting uses a hypothesis function (also called objective or estimate) that maps input features to a predicted value.

The Hypothesis Function

• The hypothesis function is a formula that multiplies input features by coefficients (weights) and sums them to make a prediction; in mathematical terms, for one feature, it is: h(x) = theta_1 * x_1 + theta_0
  • Here, theta_1 is the weight for the feature (e.g., square footage), and theta_0 is the bias (an average baseline).
• With only one feature, the model tries to fit a straight line to a scatterplot of the input feature versus the actual target value.

Bias and Multiple Features

• The bias term acts as the starting value when all features are zero, representing an average baseline cost.
• In practice, using only one feature limits accuracy; including more features (like number of bedrooms, bathrooms, location) results in multivariate linear regression: h(x) = theta_0 + theta_1 * x_1 + theta_2 * x_2 + ... for each feature x_n.

Visualization and Model Fitting

• Visualizing the problem involves plotting data points in a scatterplot: feature values on the x-axis, actual prices on the y-axis.
• The goal is to find the line (in the univariate case) that best fits the data, ideally passing through the "center" of the data cloud.

The Cost Function (Mean Squared Error)

• The cost function, or mean squared error (MSE), measures model performance by averaging squared differences between predictions and actual labels across all training examples.
• Squaring ensures positive and negative errors do not cancel each other, and dividing by twice the number of examples (2m) simplifies the calculus in the next step.

Parameter Learning via Gradient Descent

• Gradient descent is an iterative algorithm that uses calculus (specifically derivatives) to find the best values for the coefficients (thetas) by minimizing the cost function.
• The cost function's surface can be imagined as a bowl in three dimensions, where each point represents a set of parameter values and the height represents the error.
• The algorithm computes the slope at the current set of parameters and takes a proportional step (controlled by the learning rate alpha) toward the direction of the steepest decrease.
• This process is repeated until reaching the lowest point in the bowl, where error is minimized and the model best fits the data.
• Training will not produce a perfect zero error in practice, but it will yield the lowest achievable average error for the data given.

Extension to Multiple Variables

• Multivariate linear regression extends all concepts above to datasets with multiple input features, with the same process for making predictions, measuring error, and performing gradient descent.
• Technical details are essentially the same though visualization becomes complex as the number of features grows.

Essential Learning Resources

• The episode strongly directs listeners to the Andrew Ng course on Coursera as the primary recommended starting point for studying machine learning and gaining practical experience with linear regression and related concepts.

Machine learning consists of three steps: prediction, error evaluation, and learning, implemented by training algorithms on large datasets to build models that can make decisions or classifications. The primary categories of machine learning algorithms are supervised, unsupervised, and reinforcement learning, each with distinct methodologies for learning from data or experience.

Links

• Notes and resources at ocdevel.com/mlg/4
• Try a walking desk stay healthy & sharp while you learn & code

The Role of Machine Learning in Artificial Intelligence

• Artificial intelligence includes subfields such as reasoning, knowledge representation, search, planning, and learning.
• Learning connects to other AI subfields by enabling systems to improve from mistakes and past actions.

The Core Machine Learning Process

• The machine learning process follows three steps: prediction (or inference), error evaluation (or loss calculation), and training (or learning).
• In an example such as predicting chess moves, a move is made (prediction), the error or effectiveness of that move is measured (error function), and the underlying model is updated based on that error (learning).
• This process generalizes to real-world applications like predicting house prices, where a model is trained on a large dataset with many features.

Data, Features, and Models

• Datasets used for machine learning are typically structured as spreadsheets with rows as examples (e.g., individual houses) and columns as features (e.g., number of bedrooms, bathrooms, square footage).
• Features are variables used by algorithms to make predictions and can be numerical (such as square footage) or categorical (such as "is downtown" yes/no).
• The algorithm processes input data, learns the appropriate coefficients or weights for each feature through algebraic equations, and forms a model.
• The combination of the algorithm (such as code in Python or TensorFlow) and the learned weights forms the model, which is then used to make future predictions.

Online Learning and Model Updates

• After the initial training on a dataset, models can be updated incrementally with new data (called online learning).
• When new outcomes are observed that differ from predictions, this new information is used to further train and improve the model.

Categories of Machine Learning Algorithms

• Machine learning algorithms are broadly grouped into three categories: supervised, unsupervised, and reinforcement learning.
  • Supervised learning uses labeled data, where the model is trained with known inputs and outputs, such as predicting prices (continuous values) or classes (like cat/dog/tree).
  • Unsupervised learning finds similarities within data without labeled outcomes, often used for clustering or segmentation tasks such as organizing users for advertising.
  • Reinforcement learning involves an agent taking actions in an environment to achieve a goal, receiving rewards or penalties, and learning the best strategies (policies) over time.

Examples and Mathematical Foundations

• Regression algorithms like linear regression are commonly used supervised learning techniques to predict numeric outcomes.
• The process is rooted in algebra and particularly linear algebra, where matrices represent datasets and the algorithm solves for optimal coefficient values.
• The model's equation generated during training is used for making future predictions, and errors from predictions guide further learning.

Recommended Resources

• MachineLearningMastery.com: Accessible articles on ML basics.
• Podcast's own curated learning paths: ocdevel.com/mlg/resources.
• The book "The Master Algorithm" offers an introductory and audio format overview of foundational machine learning algorithms and concepts.