Who this is aimed at

Route A: Career changersYou can code or work with data already but you want the maths that makes ML training & evaluation feel logical rather than mystical.

Route B: Students & recent graduatesYou have seen some maths at university but you want job-ready fluency: translating maths into code plus explaining decisions clearly.

Same topics for both routes. Route A learns best by building first. Route B often benefits from connecting concepts to model behaviour & evaluation.

What “good at maths” really means in ML interviews

Interviewers are rarely checking whether you can reproduce proofs. They are checking whether you can:

• Keep shapes & dimensions correct

• Explain why a model is underfitting or overfitting

• Choose a sensible metric & validation method

• Reason about uncertainty & class imbalance

• Understand what gradients are doing during training

• Make trade-offs between accuracy, latency, stability, interpretability & cost

So we will focus exactly on that.

1) Linear algebra essentials for ML

Most ML is linear algebra in disguise. Features are vectors. Datasets are matrices. Weights are vectors or matrices. Embeddings are vectors in high-dimensional space.

What you actually need

Vectors & matrices

• Vector addition, scalar multiplication

• Dot product

• Norms as “size”

• Matrix multiplication & shape rules

• Transpose

Geometry intuition

• Distance vs similarity

• Cosine similarity for embeddings & retrieval

Projections & PCA intuition

• What it means to project data onto a direction

• PCA as “find directions that capture most variance”

Lightweight SVD or eigen intuitionYou do not need deep proofs. You do need to understand why PCA produces components & what “variance explained” means.

Where it shows up at work

• Linear regression, logistic regression

• Neural nets as repeated matrix multiplies with non-linearities

• Embeddings & nearest neighbour search

• PCA as a sanity check on datasets & batch effects

How to learn it without getting stuck

If you want intuition fast, 3Blue1Brown’s linear algebra series is a strong visual route. YouTube

Mini exercises that build job skill

• Implement cosine similarity & test it on simple vectors

• Take an embeddings dataset or a toy dataset, run PCA, interpret PC1 & PC2

• Write a “shape checker” function that asserts matrix dimensions in a small model pipeline

2) Probability & statistics you will actually use

ML is uncertainty management. Your data is noisy. Labels are imperfect. Sampling introduces randomness. Metrics wobble from run to run. Probability & stats is how you stop guessing.

What you actually need

Core probability

• Conditional probability intuition

• Independence vs dependence

• Base rates and why rare events break naive thinking

Distributions you see in practice

• Bernoulli and binomial for conversion events and churn

• Normal as an approximation for aggregated metrics

• Poisson for counts per time period

Statistics for evaluation

• Mean, variance, standard deviation

• Confidence interval intuition

• Sampling variability

• Correlation vs causation awareness

Metrics

• Classification: precision, recall, F1, ROC AUC, PR AUC

• Regression: MAE, RMSE

• Calibration awareness if you use probabilities

A solid free pathway for the basics is Khan Academy’s statistics & probability content. khanacademy.org

Where it shows up at work

• Choosing thresholds for classifiers

• Handling class imbalance

• Explaining whether improvements are real or noise

• Designing experiments & offline evaluations

Mini exercises that build job skill

• Build a confusion matrix & compute precision & recall

• Create an imbalanced dataset where accuracy looks good but recall is poor then fix it

• Show how metric estimates vary across different train test splits

3) Calculus essentials for gradients & backprop

Most working ML engineers do not do calculus on paper every day. But you must understand gradients conceptually because that is how models learn.

What you actually need

• Derivative as rate of change

• Partial derivatives conceptually

• Chain rule idea

• Gradient as direction of steepest increase

• Gradient descent as repeated small steps downhill

For intuition, 3Blue1Brown’s calculus series is a strong option. YouTubeFor clear step-by-step explanations of ML ideas including gradient descent, StatQuest is widely used by learners. YouTube

Where it shows up at work

• Understanding why loss goes down or fails to go down

• Diagnosing exploding or vanishing gradients

• Choosing learning rates & schedules

• Explaining backprop in plain English

Mini exercises that build job skill

• Plot a simple loss function then implement one step of gradient descent

• Train a tiny model then log gradient norms to see instability

• Explain autograd in your own words

PyTorch’s autograd tutorial is a reliable reference for how automatic differentiation supports training. docs.pytorch.org

4) Basic optimisation for training & tuning

Training is optimisation. You choose a loss function then you minimise it under constraints.

What you actually need

• Loss functions: MSE, cross-entropy

• SGD & mini-batch training intuition

• Learning rate as a primary control knob

• Regularisation: L2, dropout intuition

• Early stopping

• Hyperparameter search basics

A practical way to connect maths to code is to build a full training loop once. PyTorch’s training loop tutorial is a good reference. docs.pytorch.org

Where it shows up at work

• Model training, fine-tuning

• Debugging non-convergence

• Preventing overfitting

• Explaining trade-offs between model size, speed & performance

Mini exercises that build job skill

• Train the same model with three learning rates & interpret the curves

• Compare regularisation strengths & show how it affects validation performance

• Run a small hyperparameter search & document findings

What you can ignore for most ML jobs

If your goal is an applied ML role, you can often postpone:

• Proof-heavy linear algebra

• Measure theory and advanced probability

• Advanced calculus beyond chain rule & gradients

• Abstract optimisation theory

• Advanced information theory

These become relevant for research-heavy roles but they are not the fastest path to job readiness for most candidates.

A six-week maths plan for ML jobs

This plan is designed so you build confidence & output at the same time. Aim for 4 to 5 sessions per week of 45 to 60 minutes. Each week ends with something you can publish.

Week 1: Linear algebra for shapes & similarity

Learn

• vectors, dot product, norms, matrix shapesBuild

• cosine similarity notebook

• simple matrix shape checks in NumPyOutput

• repo with a short README explaining cosine similarity & why it matters

Use 3Blue1Brown for intuition support if needed. YouTube

Week 2: Linear models as your foundation

Learn

• linear regression & logistic regression conceptually

• link to vectors & matricesBuild

• logistic regression from scratch in NumPy

• plot decision boundary on a toy datasetOutput

• repo showing training loop, loss curve & evaluation

Week 3: Probability, metrics & class imbalance

Learn

• base rates, confusion matrix, precision recall trade-offsBuild

• imbalanced classification example

• threshold tuning based on a simple cost storyOutput

• report notebook that explains metric choice in plain English

Khan Academy is a straightforward refresher. khanacademy.org

Week 4: Calculus intuition & autograd

Learn

• derivative, chain rule, gradient descent intuitionBuild

• show gradient descent on a simple function

• replicate using PyTorch autogradOutput

• notebook explaining gradients plus a short “what autograd is doing” section

PyTorch autograd tutorial is a good anchor reference. docs.pytorch.org

Week 5: Model evaluation the professional way

Learn

• train validation test split

• cross-validation

• leakage pitfallsBuild

• cross-validated evaluation for a model

• compare two models properlyOutput

• a model evaluation pack including cross-validation results & metric discussion

scikit-learn’s cross-validation guide is a reliable reference. scikit-learn.orgIts metrics documentation is useful when choosing scoring methods. scikit-learn.org

Week 6: Optimisation & a mini deep learning capstone

Learn

• SGD, batch size, learning rate schedules, regularisationBuild

• small PyTorch model end to end

• track loss curves, p95 inference latency on your machine if relevantOutput

• a clean repo with README explaining data, model, loss, optimiser, evaluation & next improvements

PyTorch training loop tutorial is useful for structure. docs.pytorch.orgTensorFlow’s beginner quickstart is an alternative if you prefer Keras. TensorFlow

Portfolio projects that prove the maths

You want projects that show you can translate maths into working systems.

Project 1: From-scratch logistic regression

Shows

• linear algebra, probability intuition, optimisation basicsDeliver

• vectorised implementation

• loss curve

• confusion matrix

• threshold tuning explanation

Project 2: PCA plus clustering insight report

Shows

• linear algebra intuition plus good analysis practiceDeliver

• scaling rationale

• PCA plot with interpretation

• note on what could mislead you such as leakage or batch effects

Project 3: Imbalanced classification decision note

Shows

• probability thinking plus metric maturityDeliver

• PR curve discussion

• chosen threshold based on costs

• monitoring plan for base rate shift

Project 4: Tiny neural net training loop with diagnostics

Shows

• gradients, optimisation, debugging maturityDeliver

• training curves for multiple learning rates

• gradient norm logging

• short note on overfitting signs & fixes

How to write this on your CV

Avoid “strong maths” as a claim. Use outcomes.

• Implemented logistic regression using vectorised NumPy operations & gradient descent with documented loss curves

• Evaluated imbalanced classifiers using precision recall trade-offs & threshold tuning aligned to business costs

• Applied cross-validation to compare models & prevent overfitting with clear metric justification scikit-learn.org

• Built a PyTorch training loop using autograd plus optimiser configuration & training diagnostics docs.pytorch.org

Resources section

Linear algebra

• 3Blue1Brown Essence of Linear Algebra playlist. YouTube

Probability & statistics

• Khan Academy Statistics & Probability. khanacademy.org

Calculus, gradients & intuition

• 3Blue1Brown Essence of Calculus. YouTube

• StatQuest playlists including gradient descent explanations. YouTube

Core ML maths reference

• Stanford CS229 lecture notes PDF. cs229.stanford.edu

Model evaluation & selection

• scikit-learn cross-validation guide. scikit-learn.org

• scikit-learn model evaluation metrics documentation. scikit-learn.org

Deep learning frameworks for building projects

• PyTorch autograd tutorial. docs.pytorch.org

• PyTorch training loop tutorial. docs.pytorch.org

• TensorFlow beginner quickstart & tutorials hub. TensorFlow