Frequently Asked Questions About Edureka ML Foundations & Math Skill
22 answers covering everything from basics to advanced usage.
// Basics
What is supervised learning in simple terms?
Supervised learning is a paradigm where each training instance has input attributes X and an expected output Y, and the algorithm learns the mapping function Y = f(X) so it can predict Y for never-seen inputs. Use it when your data is labelled with the answers — for example, historical customer records tagged as delinquent or non-delinquent.
What is unsupervised learning and what can't it do?
Unsupervised learning detects patterns and groups similar instances into clusters based on inherent characteristics, without any expected output. Its key limitation: it groups similar instances but cannot label the meaning of a cluster. A domain expert must name clusters post-hoc. Algorithms include K-Means Clustering, Apriori, Hierarchical Clustering, and Autoencoders.
What is reinforcement learning with a real example?
Reinforcement learning is where an agent interacts with an environment, earns rewards for correct actions and penalties for wrong ones, and refines its policy to maximise cumulative reward. Example: a gaming AI learning a platformer level — actions are move/jump, rewards are progress, penalties are falling. It parallels Pavlov's conditioning: repeated reward signals shape future behaviour.
What is the difference between a scalar, a vector, and a matrix?
A scalar is a single numeric value like price or temperature. A vector is an ordered list of numbers representing magnitude and direction — a row or column of a dataset, or pixel/audio data. A matrix is a rectangular array of numbers used to encode equations and data so computational operations like scaling and rotation can run systematically.
// How To
How do I represent my data as matrices for ML?
Convert every equation into matrix form: extract coefficients into rows, put variables (x, y, z) into a column vector, and outputs into a result vector. Before any multiplication, confirm matrix orders are compatible — the number of columns in matrix A must equal the number of rows in matrix B. This makes operations computable.
How do I apply gradient descent step by step?
Define an error function like (predicted − target)². Differentiate it using the Power, Sum, Product, or Chain Rule. For multi-variable models, use partial differentiation — change one variable, hold others constant. Then update: weight = weight − learning_rate × gradient, and repeat for enough steps until output converges to the target.
How do I use PCA in scikit-learn?
Import PCA from sklearn.decomposition, fit it on your dataset, then inspect explained_variance_ and components_. The longer the eigenvector, the more information it carries. Reduce to the minimum number of components that preserve the information you need, then inverse-transform to visualise the reduced representation. The goal is extracting information, not keeping all the data.
How do I handle a dataset with only a few labelled examples?
Use semi-supervised learning when labels are scarce but unlabelled data is abundant. Train on the labelled data to learn feature patterns, then extend that learning to the unlabelled pool with a hybrid model combining labelled and unlabelled loss. For high-dimensional data like MRI scans, apply PCA to reduce dimensionality before training to cut cost and noise.
// Troubleshooting
Why is my matrix multiplication failing?
Matrix multiplication fails when the orders are incompatible. The number of columns in matrix A must equal the number of rows in matrix B before you can multiply them. Always verify order compatibility first. If dimensions don't match, use a transpose to flip rows into columns and align the shapes.
Why is my model's accuracy not improving?
Common causes: too few gradient descent iterations (10 versus 100,000 steps changes convergence dramatically), the wrong learning paradigm for your data, or insufficient training data. Follow the loop: train → evaluate → augment data → retrain. Also check your learning rate — too high overshoots the target, too low converges too slowly.
Why is my image transformation pipeline corrupting features?
Use eigenvector analysis to diagnose it. Identify vectors that change direction under the transformation — these are unreliable and should be discarded or fixed. Identify vectors that preserve direction but scale — these are eigenvectors, the most sensitive and trustworthy parts of the dataset. Use those eigenvectors as the basis for further analysis or PCA components.
My inverse matrix doesn't exist — what went wrong?
An inverse does not exist when the vector lacks sufficient information, which shows up as a zero determinant. Always compute the determinant first — it quantifies the matrix's weight and sensitivity — before attempting to invert. If the determinant is zero, the system Ax = b can't be solved directly via x = A⁻¹b; you may need more data or a different method like row echelon.
// Comparisons
How does this roadmap compare to a generic 'just try algorithms' approach?
A generic try-everything approach skips paradigm selection, so you might run supervised learning on unlabelled data and get meaningless results. This roadmap forces you to classify the problem first, then match algorithm families, verify matrix compatibility, and optimise deliberately. It makes failures debuggable because you understand what each step is doing rather than trusting a black box.
What is the difference between the Jacobian and the Hessian?
The Jacobian is the matrix of first-order partial derivatives of a vector function; it points to global maxima and linearises nonlinear functions at a point. The Hessian is the matrix of second-order partial derivatives — the derivative of the Jacobian — used to minimise errors in deep learning and optimise gradient descent. Jacobian is first derivative, Hessian is second.
How does exploration differ from exploitation in reinforcement learning?
Exploration is the agent acting on trial and error to discover the environment — it hasn't learned yet, so it tries random actions. Exploitation is acting on accumulated knowledge to maximise known rewards. Training runs the exploration phase until the policy stabilises via accumulated rewards, then the exploitation phase begins. Both are needed for the agent to learn effectively.
Is classification or regression right for my problem?
Choose based on your output type: use classification when the output is discrete (categories like default/no-default), and regression when the output is continuous (a numeric value like price). Both fall under supervised learning. This choice determines which specific algorithm you pull from the supervised family — Decision Trees and SVM lean toward classification, Linear Regression toward continuous prediction.
// Advanced
Why can't delta x be exactly zero in differentiation?
In differentiation, delta x approaches zero but is never exactly zero. Treating it as absolute zero destroys the function because you'd be dividing by zero. This is a common pitfall — the limit concept means you get infinitely close without ever touching zero, which is what makes derivatives (and therefore gradient descent) mathematically valid.
What is projection and how is it different from subtraction?
Projection of vector v1 onto v2 extracts the shadow of v1's information onto v2 — it's not arithmetic subtraction of values. Confusing the two is a common pitfall. Projection captures how much of one vector's information aligns with another's direction, which matters for decomposing data along meaningful axes like principal components.
How do partial derivatives help in multi-variable ML models?
Partial differentiation lets you isolate the effect of each input on the model's output by differentiating with respect to one variable while holding all others constant. In models with many weights, this tells you how much each weight contributes to the error, which is exactly what gradient descent needs to update each weight in the right direction.
What MLOps tools should I use to deploy an ML model?
Use TensorFlow for model building and serving, and AWS Machine Learning or Azure Machine Learning for scalable cloud deployment. Apply MLOps practices to version, monitor, and retrain models in production as new data arrives. Deployment isn't the end — you continuously monitor performance and retrain as data drifts to keep accuracy acceptable.
Do all probabilities in a model need to sum to one?
Yes — all probabilities within a sample space must sum to 1. Probability itself is the ratio of desired outcomes to total outcomes. Before applying any probabilistic ML method, define the random experiment (the uncertain process), the sample space (all possible outcomes), and the events (specific outcomes of interest). This foundation validates model confidence and sets decision thresholds.
When is deep learning worth using over classical ML?
Deep learning is worth it when classical ML can't reach the accuracy you need, especially on unstructured data like images, audio, or text with complex patterns. Deep learning uses deep neural networks and relies heavily on the Hessian to minimise errors. But it's more compute-hungry and data-hungry, so start with classical ML unless the problem demands more.