Machine Learning Basics: Supervised, Unsupervised & Reinforcement

This video explains the paradigm shift from traditional algorithmic approaches to machine learning. In traditional computing, humans define explicit rules (recipes) that transform input data into outputs. Machine learning reverses this by providing data and desired outputs, then letting computers discover the optimal algorithm. This 'master algorithm' approach allows machines to compute the best way of combining data to create outputs, applicable across diverse domains from material extraction to solving complex social problems. The key insight is that instead of programming specific instructions, we let data-driven systems learn the optimal solution paths.

AI represents a fundamental paradigm shift from rule-based to data-driven systems. Traditional systems require explicit programming of rules for each scenario, while AI systems use constant algorithms that learn from data. In rule-based systems, new rules must be manually programmed for new situations. In AI systems, the same algorithm adapts to new data without manual modifications, enabling generalization across diverse scenarios.

Traditional programming requires explicit rule specification; machine learning inverts this by learning rules from data. We observe input-output pairs (X,Y) and seek to estimate the function f mapping X to Y. Mathematically, Y = f(X) + ε where ε is zero-mean noise. The target function equals the conditional expectation E[Y|X], which minimizes mean squared error. This paradigm enables solving problems too complex to encode manually, like recognizing objects in images or understanding natural language.

Machine learning represents a fundamental paradigm shift from traditional programming, where software engineers write explicit rules for every contingency, to systems that learn rules from data. This transformation enables breakthroughs like AlphaGo, which studied over a million games to develop its own strategies without explicit programming. Deep learning processes visual input through layered stages, extracting edges, shapes, and complex concepts. Applications span autonomous vehicles using camera and sensor data for navigation, and medical diagnosis where neural networks trained on 129,000 skin condition images match or exceed board-certified dermatologists' accuracy. Crowdsourcing combined with machine learning democratizes expertise, enabling solutions in days that would take corporations months.

AI represents a fundamental paradigm shift from rule-based programming to machine learning. Traditional systems used explicit rules (if this, then that), while modern AI learns by itself using machine learning techniques, allowing adaptation to unseen situations. AI processes data as numerical values: colors become unique numerical values, and words are assigned numerical values based on dictionaries and sentence positions. The AI builds neural arrays or matrices of patterns and assigns weights to each input. When AI generates novel ideas, it is called generative AI, which is the first indication of real intelligence.

This comprehensive section covers linear algebra and coordinate geometry: (1) Matrix operations: solving equations by equating corresponding elements, finding 2x2 matrix inverses using (1/det(A)) × [[d, -b], [-c, a]], calculating determinants as (a×d) - (b×c), (2) Matrix properties: symmetric matrices equal to their transpose, skew-symmetric matrices with zeros on diagonal, (3) Determinant properties: equals zero if any two rows/columns are identical or if any row/column consists entirely of zeros, (4) Coordinate geometry: midpoint formula, centroid of triangle, section formula for internal and external division, (5) Line equations: general form Ax + By + C = 0, slope-intercept form, intercept form, (6) Slope relationships: parallel lines have equal slopes, perpendicular lines have slopes whose product is -1. These concepts form the foundation for solving linear algebra and coordinate geometry problems.

This section covers the foundational concepts of coordinate systems and vectors. The global system uses fixed basis vectors e1=(1,0,0), e2=(0,1,0), e3=(0,0,1) with origin at (0,0,0). Local systems have their own origin and axis vectors that may not be perpendicular. Coordinate conversion expresses the same point in different systems using matrices. Vectors are created by subtracting points (q-p=v) and have magnitude/direction but no fixed location. Barycentric combinations express points as weighted averages with weights summing to 1, defining convex regions. Linear dependence means vectors are parallel (one is a scalar multiple of the other); independence means they point in different directions.

This comprehensive section covers three major mathematical domains: (1) Vector Algebra: Vectors have magnitude and direction, represented in Cartesian (x, y), Polar (r, θ), and Unit vector (i, j) forms. The norm is √(x² + y²). Operations include addition (triangle law), subtraction, and parallelism (a = kb). The section formula finds points dividing segments in given ratios. (2) Coordinate Geometry: Lines are described by general (Ax + By + C = 0), slope-intercept (y = mx + c), and parametric forms. Slope calculation methods include from equation, direction vector, and two points. Key concepts: parallel lines have equal slopes, perpendicular lines have negative reciprocal slopes (m₁m₂ = -1). Circle geometry includes sector area ((1/2)r²θ), segment area ((1/2)r²(θ - sin(θ))), and regular polygon area ((n × s²)/(4 × tan(π/n))). (3) Matrices: Order is rows × columns. Types include row, column, square, diagonal, identity, and zero matrices. Transpose swaps rows and columns. Symmetric matrices equal their transpose (aᵢⱼ = aⱼᵢ), skew-symmetric matrices equal negative transpose (aᵢⱼ = -aⱼᵢ, diagonal zeros). Multiplication requires matching inner dimensions. Properties include associativity and distributivity but not commutativity.

This lecture introduces Chapter 3 on Vector Spaces (Ruang Vektor) in elementary linear algebra. Scalars have magnitude but no direction, while vectors have both magnitude and direction. Vectors are denoted by bold lowercase letters with arrows, defined by initial and terminal points. Two vectors are equivalent if they share the same magnitude and direction. The zero vector has zero length and no direction. Vector spaces generalize 2D and 3D geometry, named after You Clip of Alexandria.

A vector is a directed line segment with origin and endpoint; zero vector has coinciding endpoints with no direction. Vectors are collinear (same or parallel lines) or coplanar (parallel to some plane). Linear operations include addition by triangle/polygon rules, subtraction as sum with opposite vector, and scalar multiplication. Linear combination expresses vectors as weighted sums; dependence means non-trivial combination equals zero. Basis on line/plane/space consists of 1/2/3 linearly independent vectors. Coordinates are coefficients in basis representation. Two vectors equal iff coordinates equal; collinear iff coordinates proportional. Right/left-handed systems determined by rotation direction. Point coordinates equal radius vector coordinates.

Training data is the data used to build a program that learns from data. To collect training data, one must gather samples from different sources, measure relevant features using appropriate tools (such as weight machines, height scales, or medical diagnostic equipment), and record both the feature values and the corresponding target labels. The data is typically stored in a structured format like an Excel file with columns representing each feature and a final column for the class label.

Data consists of various values collected for different objects, describing the properties of those objects. Data can be collected from various sources and experiments. The easiest way to store data is in terms of records, which consist of different attributes (columns). These columns are also called features. Data can have different types including numerical values and nominal (categorical) values, which describe one or more objects or their properties.

Features (also called attributes or properties) are the columns in a dataset that describe the data points. When data is presented in a table format, it becomes a tabular dataset. The technical term used throughout the course is 'features.' Tabular data is also called structured data because it has a defined structure with rows and columns, similar to tables in databases.

In Machine Learning, data is typically organized in a tabular format where each row represents a single item (like a patient, customer, or house) and each column represents a feature (attribute or characteristic). The last column is called the 'label' or 'target' - this is what the model is trying to predict. All other columns are called 'features' or 'input variables'. For example, in a diabetes prediction dataset, features might include age, blood glucose level, blood pressure, and BMI, while the label would be whether the patient has diabetes (yes/no).

Data points are units of information or objects that convey information. They are characterized by two types of properties: (1) Features - properties that can be easily measured or computed automatically, and (2) Labels - properties that are of interest but may be difficult to collect or infer. Machine learning aims to predict the label given the features of a data point. Examples include housing data (features: bedrooms, floor; label: price), text data (features: number of words; label: sentiment), images, and genomic sequences.