Quick Summary: Machine Learning is the single largest subject in GATE DA — roughly 18–22 marks, more than any other core section. The official IIT Madras GATE 2027 syllabus splits it into supervised learning (regression, classifiers, SVM, decision trees, bias-variance, cross-validation, neural networks) and unsupervised learning (k-means, hierarchical clustering, PCA). This guide explains every listed topic, shows how GATE frames its questions, and gives you a 6-week plan to master the section.
By Piyush Wairale — Instructor, BS Data Science program at IIT Madras · IIT Madras alumnus · 10,000+ GATE students mentored · Last updated: August 2026 · Verified against the official IIT Madras GATE 2027 DA syllabus
Key Takeaways
- Machine Learning carries approximately 18–22 of the 85 core marks in GATE DA — about one mark in every four — making it the highest-priority subject in the paper.
- The official syllabus lists supervised learning (simple/multiple linear regression, ridge regression, logistic regression, k-NN, naive Bayes, LDA, SVM, decision trees, bias-variance trade-off, LOO and k-fold cross-validation, MLP and feed-forward neural networks) and unsupervised learning (k-means/k-medoid, hierarchical clustering with single and multiple linkage, PCA).
- High bias means underfitting (poor performance on both training and test data); high variance means overfitting (excellent training performance, poor test performance). Total expected error decomposes as bias² + variance + irreducible noise.
- k-fold cross-validation trains k models, each holding out one fold; leave-one-out (LOO) is the extreme case where k equals the number of data points — nearly unbiased but expensive.
- PCA finds orthogonal directions of maximum variance — the eigenvectors of the covariance matrix ranked by eigenvalue — which ties Machine Learning directly to the Linear Algebra section of GATE DA.
On this page
Why ML decides your rank · The syllabus, mapped · Regression family · The five classifiers · Bias-variance & cross-validation · Neural networks · Clustering & PCA · How GATE tests ML · 6-week plan · FAQs
Ask any GATE DA topper where their rank came from and the answer is almost always the same subject: Machine Learning. It is the largest block of the paper, it connects to three other sections (probability, linear algebra and calculus all feed into it), and its questions are refreshingly predictable — numericals on regression and distance-based classifiers, conceptual questions on overfitting and validation, and computation questions on clustering and PCA. This article walks through every topic named in the official IIT Madras GATE 2027 syllabus, in teaching order, with the diagrams and exam patterns you need. Everything here is taught in full depth — with derivations, solved PYQs and tests — in the Machine Learning for GATE DA Course & Test Series, and the free video lectures below.
Watch Free: Machine Learning for GATE DA — Full Playlist
Every algorithm in this article, taught on the whiteboard with GATE-style numericals by Piyush Wairale (IIT Madras):
Subscribe to Piyush Wairale IITM on YouTube for new GATE DA lectures, PYQ solutions and strategy sessions.
Why Machine Learning Decides Your GATE DA Rank
Do the arithmetic once and the priority becomes obvious. The core (non-aptitude) portion of GATE DA is 85 marks across seven subjects. Machine Learning has consistently contributed the largest share since the paper’s debut in 2024 — typically 18–22 marks once you count the direct ML questions plus the questions that dress ML in probability or linear-algebra clothing. A candidate who is strong in ML and average everywhere else routinely outscores a candidate who is uniformly decent. There are three structural reasons for this weightage: ML is the paper’s namesake discipline (this is a Data Science and AI exam), its topics compose naturally into 2-mark numericals, and it lets examiners test probability, calculus and linear algebra indirectly — a PCA question is secretly an eigenvector question; a naive Bayes question is secretly a conditional-probability question.
The flip side: ML is the worst subject to study from scattered internet notes, because its GATE syllabus is specific. The official IIT Madras GATE 2027 syllabus names exact algorithms and exact validation techniques. Anything outside that list — boosting, random forests, deep CNNs, transformers — is not in the syllabus, and time spent there is time lost. Study the list, the whole list, and nothing but the list.
The Official Syllabus, Mapped
The Regression Family
Simple and multiple linear regression
Simple linear regression fits y = β₀ + β₁x by minimising the sum of squared errors. The two results you must know cold: the least-squares slope is β₁ = Σ(xᵢ−x̄)(yᵢ−ȳ) / Σ(xᵢ−x̄)² — equivalently Cov(x,y)/Var(x) — and the fitted line always passes through the point (x̄, ȳ). GATE numericals hand you five or six data points and ask for the slope, the intercept, a prediction at a new x, or the residual at a given point. A second recurring fact: the slope is related to the correlation coefficient by β₁ = r·(sy/sx), which lets examiners connect this topic to the correlation material in Probability & Statistics. Multiple linear regression generalizes to several predictors, y = Xβ + ε, with the closed-form normal-equation solution β̂ = (XᵀX)⁻¹Xᵀy — know the formula, its derivation sketch (set the gradient of the squared error to zero), and the condition for it to exist (XᵀX must be invertible, i.e., no perfect multicollinearity).
Ridge regression
Ridge regression adds an L2 penalty to the loss: minimise ‖y − Xβ‖² + λ‖β‖². The closed form becomes β̂ = (XᵀX + λI)⁻¹Xᵀy. Three exam-ready facts: the penalty shrinks coefficients toward zero but never exactly to zero; adding λI makes the matrix invertible even under multicollinearity (a practical motivation GATE has asked directly); and λ controls the bias-variance trade-off — larger λ means more bias, less variance. When λ = 0 ridge reduces to ordinary least squares; as λ → ∞ all coefficients shrink toward zero.
Logistic regression — the misnamed classifier
Despite the name, logistic regression is a classification algorithm. It models the probability of the positive class as P(y=1|x) = σ(wᵀx + b), where σ(z) = 1/(1+e⁻ᶻ) is the sigmoid. Facts GATE tests: the sigmoid maps any real number to (0,1); the decision boundary (where P = 0.5, i.e., wᵀx + b = 0) is linear; the model is trained by maximising log-likelihood (equivalently minimising cross-entropy), not squared error; and σ′(z) = σ(z)(1−σ(z)) — a one-mark derivative question that has appeared in multiple GATE papers. The odds and log-odds formulation (log-odds = wᵀx + b) is also fair game.
The Five Classifiers You Must Know
k-nearest neighbours (k-NN)
The simplest classifier in the syllabus: to label a query point, find the k closest training points (usually Euclidean distance) and take a majority vote. k-NN has no training phase — it is a “lazy learner” that memorises the data — and its prediction cost grows with the size of the training set. Small k → flexible, jagged boundary, high variance; large k → smooth boundary, high bias. GATE numericals give you a small 2-D dataset and ask for the predicted class at a query point for k = 1 and k = 3 — practise computing squared Euclidean distances quickly (skip the square root; ordering is unchanged).
Naive Bayes
A direct application of Bayes’ theorem with one big assumption: features are conditionally independent given the class. Classify x by choosing the class c maximising P(c)·∏ᵢP(xᵢ|c). The numerator product is exactly why it is “naive” — real features are rarely independent — yet the classifier works surprisingly well. GATE questions give a small frequency table and ask you to classify a new instance; the only trap is forgetting the prior P(c) or normalising when the question only asks which class wins (you don’t need to normalise to compare). Note the direct bridge to the reasoning-under-uncertainty topic: naive Bayes is a tiny Bayesian network with the class as the single parent of every feature.
Linear discriminant analysis (LDA)
LDA models each class as a Gaussian with its own mean but a shared covariance matrix; the resulting decision boundary between two classes is linear. Contrast with logistic regression (which models P(y|x) directly rather than modelling the class densities) — “generative vs discriminative” is a classic conceptual question. LDA is also used as a supervised dimensionality-reduction technique that maximises between-class separation relative to within-class scatter; keep that one-line summary ready.
Support vector machines (SVM)
The SVM finds the separating hyperplane wᵀx + b = 0 that maximises the margin — the distance to the nearest training points. The margin equals 2/‖w‖, so maximising margin is minimising ‖w‖²/2 subject to yᵢ(wᵀxᵢ + b) ≥ 1. The points on the margin boundaries are the support vectors — they alone determine the solution; removing any non-support vector changes nothing (a repeated GATE MCQ). For non-separable data, soft-margin SVM introduces slack variables with penalty C: large C → narrow margin, few violations (low bias, high variance); small C → wide margin, more violations. Know the kernel idea in one line — kernels compute inner products in an implicit high-dimensional feature space, letting a linear method learn non-linear boundaries — with the RBF and polynomial kernels as named examples.
Decision trees
A decision tree splits the feature space by asking one question per internal node, chosen greedily to maximise impurity reduction. The two impurity measures to memorise for a node with class proportions pᵢ: entropy H = −Σpᵢlog₂pᵢ (maximum 1 bit for a balanced binary node) and Gini index G = 1 − Σpᵢ² (maximum 0.5). Information gain = parent entropy − weighted average child entropy; GATE’s favourite ML numerical is computing the information gain of a candidate split from a small table. Also know: deep unpruned trees overfit (high variance); pruning or depth limits trade variance for bias; trees natively handle non-linear boundaries and mixed feature types.
Bias-Variance Trade-off and Cross-Validation
This is the conceptual heart of the section, and GATE tests it every single year in some form. Bias is error from a model too simple to capture the true pattern — it underfits, performing poorly on training and test data. Variance is error from a model too sensitive to the particular training sample — it overfits, performing brilliantly on training data and poorly on test data. Expected test error decomposes as bias² + variance + irreducible noise. Every regularisation choice you have met so far is a knob on this trade-off: ridge’s λ, k in k-NN, tree depth, SVM’s C.
Cross-validation: k-fold and leave-one-out
How do you estimate test error without a separate test set? k-fold cross-validation: split the n training points into k equal folds; train k times, each time holding out one fold for validation and training on the other k−1; average the k validation errors. Each data point is validated on exactly once and trained on k−1 times. Leave-one-out (LOO) is the special case k = n: n models, each trained on n−1 points. LOO’s estimate is nearly unbiased (each training set is almost the full data) but expensive (n trainings) and can have high variance; k = 5 or 10 is the practical default. GATE asks: how many models are trained (k, or n for LOO)? How many times is each point used for training (k−1)? What trade-off does small vs large k control? All three have appeared as direct questions.
Neural Networks: MLP and Feed-Forward
The syllabus stops at multi-layer perceptrons (MLP) and feed-forward networks — no CNNs, no RNNs, no transformers. A feed-forward network is layers of units, each computing an activation of a weighted sum: a = f(Wx + b), with information flowing input → hidden → output with no cycles. What GATE tests: (1) parameter counting — a network with layer sizes n₀-n₁-n₂ has n₀·n₁ + n₁ weights-plus-biases into the hidden layer and n₁·n₂ + n₂ into the output, so count weights = Σ nᵢ·nᵢ₊₁ and biases = Σ nᵢ₊₁; (2) why non-linear activations matter — stacking linear layers without non-linearity collapses to a single linear map, so an MLP with identity activations is no more powerful than linear regression; (3) the classic XOR fact — a single perceptron cannot represent XOR because XOR is not linearly separable, but one hidden layer fixes it; (4) activation functions and derivatives — sigmoid (σ′ = σ(1−σ)), tanh, ReLU (max(0, z)); and (5) the one-line description of backpropagation: gradients of the loss computed layer by layer via the chain rule. Deep networks overfit small datasets — connecting straight back to the bias-variance story.
Unsupervised Learning: Clustering and PCA
k-means and k-medoid
k-means alternates two steps until nothing changes: assign each point to its nearest centroid, then recompute each centroid as the mean of its assigned points. It minimises the within-cluster sum of squared distances, is guaranteed to converge (the objective decreases every step and has finitely many configurations), but only to a local optimum — the final clustering depends on initialisation. GATE numericals give you a handful of 1-D or 2-D points and initial centroids and ask for the assignments or centroids after one or two iterations — practise these until they take three minutes. k-medoid differs in one word: the cluster representative must be an actual data point (the medoid), which makes it robust to outliers and usable with arbitrary distance measures, at higher computational cost.
Hierarchical clustering
The syllabus names both directions: bottom-up (agglomerative) — start with every point as its own cluster and repeatedly merge the two closest clusters — and top-down (divisive) — start with one cluster and recursively split. “Closest clusters” depends on the linkage: single linkage uses the minimum distance between any pair of points across the two clusters (produces elongated, chain-like clusters — the “chaining effect”), while complete/multiple linkage uses the maximum pairwise distance (produces compact, spherical clusters). The output is a dendrogram; cutting it at a chosen height yields a flat clustering. GATE gives a small distance matrix and asks which clusters merge first under a given linkage, or for the final clustering after all merges above some threshold — mechanical marks with practice.
Principal component analysis (PCA)
PCA finds the orthogonal directions along which the (centred) data varies most. Procedure: centre the data, compute the covariance matrix, take its eigenvectors and eigenvalues; the eigenvector with the largest eigenvalue is the first principal component, and each eigenvalue equals the variance captured along its component. The proportion of variance explained by the top k components is (λ₁+…+λk)/(λ₁+…+λd) — a standard GATE numerical: “given eigenvalues 4, 2, 1, 1, how many components capture at least 75% of the variance?” (Answer: two — 6/8 = 75%.) Because principal components are eigenvectors of a symmetric matrix, they are orthogonal — and this is where GATE loves to fuse ML with the Linear Algebra section: a PCA question is an eigenvalue question wearing a data-science costume. Also connect PCA to SVD: the principal components are the right singular vectors of the centred data matrix.
How GATE Actually Tests Machine Learning
| Question pattern | What it looks like | Typical marks |
|---|---|---|
| Regression numerical | Compute slope/intercept/prediction from 4–6 data points | 2 (NAT) |
| k-NN / k-means numerical | Distances on a small 2-D dataset; classify a point or run one k-means iteration | 1–2 |
| Entropy / information gain | Compute the best split from a frequency table | 2 |
| Bias-variance concepts | Which change increases/decreases overfitting; interpret train-vs-test error | 1 |
| Cross-validation counting | Models trained / evaluations in k-fold and LOO | 1 |
| PCA / eigen numerical | Variance explained from eigenvalues; first principal component of a 2×2 covariance | 2 |
| Neural network counting | Number of parameters in an MLP; XOR separability; activation derivatives | 1–2 |
| Naive Bayes / MSQ concepts | Classify from a frequency table; which statements about SVM/LDA/logistic are true | 1–2 |
The 6-Week Machine Learning Study Plan
- Week 1 — Regression: simple, multiple, ridge, logistic; derive the least-squares slope once by hand; 25 numericals.
- Week 2 — Distance & probabilistic classifiers: k-NN, naive Bayes, LDA; connect naive Bayes to Bayes’ theorem PYQs.
- Week 3 — SVM and decision trees: margins and support vectors; entropy, Gini and information gain drills.
- Week 4 — Validation and neural networks: bias-variance scenarios, k-fold/LOO counting questions, MLP parameter counting, XOR.
- Week 5 — Unsupervised: k-means iterations by hand, linkage exercises on distance matrices, PCA variance-explained numericals.
- Week 6 — Integration: full ML sectional tests, GATE DA + GATE CS ML PYQs, error-log review; revisit whichever algorithm cost you marks.
Study Machine Learning the Structured Way
Every algorithm above — with derivations, solved GATE PYQs, topic-wise tests and doubt support — is covered in the dedicated course:
- Machine Learning for GATE DA Course & Test Series — the complete supervised + unsupervised syllabus, taught by IIT Madras alumnus Piyush Wairale.
- Complete GATE DA 2027 Course — all 7 subjects, 250+ hours, 4,000+ questions and full-length AI mock tests in one package.
FAQs on Machine Learning for GATE DA
How many marks does Machine Learning carry in GATE DA?
Approximately 18–22 of the 85 core marks — the highest of any subject. Counting probability and linear-algebra questions framed in ML language, its effective footprint is even larger.
Are deep learning topics like CNNs and transformers in the GATE DA syllabus?
No. The official IIT Madras GATE 2027 syllabus stops at multi-layer perceptrons and feed-forward networks. CNNs, RNNs, transformers, boosting and random forests are all outside the syllabus — don’t spend preparation time there.
What should I study before starting Machine Learning?
Probability & Statistics (Bayes’ theorem, distributions, expectation) and Linear Algebra (eigenvalues, projections) — every ML algorithm in the syllabus sits on one or both. See the 6-month preparation plan for the full sequencing.
Is logistic regression a regression or classification algorithm?
Classification — despite the name. It models the probability of a class through the sigmoid function and draws a linear decision boundary. This exact trap has appeared in GATE MCQs.
Which book is best for GATE DA Machine Learning?
“An Introduction to Statistical Learning” (ISLR) matches the syllabus almost one-to-one, with Tom Mitchell for decision trees and neural network basics. Full list in our best books for GATE DA guide.
Machine Learning rewards structured preparation more than any other GATE DA subject: fifteen named algorithms, each with two or three examiner-favourite facts, and numericals that repeat their shape year after year. Master the syllabus map above, drill the numerical patterns, and this quarter of the paper becomes your scoring engine. For the complete subject sequence, see the GATE DA Syllabus 2027 breakdown and the 6-month study plan.
Master GATE DA Machine Learning
Every algorithm in this guide — regression to PCA — with derivations, solved PYQs, sectional tests and mentorship, by Piyush Wairale (IIT Madras).
Join the ML Course & Test Series Get the Complete GATE DA CourseRecent Post
Probability and Statistics carries ~12–16 marks in GATE DA and underpins Machine Learning. This complete guide covers the full official syllabus — counting, Bayes' theorem, expectation, every named distribution, CLT, confidence intervals and the z/t/chi-squared tests — with diagrams, a worked Bayes example, GATE patterns, a study plan and FAQs, by IIT Madras instructor Piyush Wairale.
Master reasoning under uncertainty for GATE DA 2027 — Bayesian networks, conditional independence, d-separation, variable elimination with a fully worked alarm-network example, and sampling-based approximate inference. With diagrams, GATE question patterns, a 10-day plan and FAQs, by IIT Madras instructor Piyush Wairale.
Complete GATE RA 2027 guide to servo motors: closed-loop principle, AC/DC/BLDC types, continuous vs intermittent torque-speed zones, nested PID control, sizing, diagrams and worked example.
Complete GATE RA 2027 guide to stepper motors: VR/PM/hybrid types, step angle, full/half/microstepping, pull-in and pull-out torque-speed characteristics, drives, diagrams and worked examples.
Complete GATE RA 2027 guide to DC motors: principle of operation, back-EMF, torque equation, shunt/series/compound torque-speed characteristics, speed control, performance, diagrams and worked examples.
Complete GATE RA 2027 guide to hydraulic and pneumatic actuators: principle of operation, cylinders and fluid motors, force/torque-speed characteristics, performance, worked examples and diagrams.
Learn Daily, Wherever You Are
Free lectures, exam updates, PYQ discussions, and job alerts — delivered through our YouTube channel and Telegram communities.
