MAI study plan

Define AI, machine learning, and deep learning precisely, and classify ten given systems (rule-based expert system, spam filter, LLM chatbot, recommendation engine, and so on) into the correct category with a one-sentence justification for each.

Differentiate supervised, unsupervised, self-supervised, and reinforcement learning, and for each paradigm name two realistic business use cases plus the label or reward signal each one requires.

Explain the bias-variance tradeoff, overfitting, and underfitting in your own words, and sketch how training versus validation error curves reveal each failure mode.

Compare at least six classical algorithms (linear and logistic regression, decision trees, random forests, gradient boosting, k-means, k-nearest neighbors) on interpretability, data requirements, and typical failure modes, and match each to a plausible business problem.

Work through the core probability and linear algebra concepts that recur in later domains: conditional probability, Bayes' theorem, expectation, vectors, matrices, dot products, and gradients, solving at least fifteen short practice problems.

Distinguish classification, regression, clustering, ranking, and generation tasks, and select the correct task type and target metric for five described business scenarios.

Identify and remediate data quality problems including missing values, duplicates, outliers, leakage, and class imbalance, and justify the chosen imputation or resampling strategy for each case.

Engineer at least eight features from a raw tabular dataset, covering encoding of categorical variables, scaling, binning, date-part extraction, and one interaction term, and document why each transformation helps.

Explain the purpose of training, validation, and test splits and of k-fold cross-validation, and describe how each prevents an optimistic performance estimate.

Compute and interpret a confusion matrix, precision, recall, F1, ROC-AUC, and PR-AUC for a binary classifier, and explain when accuracy is a misleading metric.

Interpret regression diagnostics including MAE, RMSE, R-squared, and residual plots, and explain to a non-technical stakeholder what each one does and does not tell you.

Describe the structure and forward-pass computation of a feedforward neural network, including weights, biases, activation functions, and loss, and hand-calculate the output of a small two-layer network.

Explain backpropagation and gradient descent conceptually, distinguishing batch, stochastic, and mini-batch variants, and describe the roles of learning rate, epochs, and batch size.

Diagnose common training problems such as vanishing gradients, exploding gradients, and dead ReLU units, and match each to at least one standard remedy.

Explain regularization techniques including dropout, weight decay, early stopping, and batch normalization, and state the problem each one addresses.

Compare CNN, RNN, and transformer architectures on the data types they suit, their inductive biases, and their computational costs.

Explain the transformer attention mechanism, including queries, keys, values, and multi-head attention, at a conceptual level sufficient to describe why it enables long-range context.

Define tokenization, embeddings, context window, temperature, top-p sampling, and hallucination, and explain how each affects LLM output quality and cost.

Distinguish pretraining, supervised fine-tuning, parameter-efficient fine-tuning such as LoRA, and reinforcement learning from human feedback, and state when each approach is appropriate.

Describe retrieval-augmented generation end to end, including chunking, embedding, vector search, and prompt assembly, and identify the failure points at each stage.

Write and iteratively refine at least five prompts that use role specification, few-shot examples, structured output constraints, and chain-of-thought instructions, and compare outputs across versions.

Evaluate LLM outputs systematically by defining task-specific rubrics, building a small labeled test set, and computing agreement between human and model-graded scores.

Identify concrete LLM risks including hallucination, prompt injection, data leakage through prompts, copyright exposure, and toxic output, and propose one mitigation for each.

Explain core fairness concepts including disparate impact, equal opportunity, and proxy variables, and describe how bias enters data, model design, and deployment.

Map the practices of transparency, explainability, human oversight, and accountability to a described AI deployment, and name the documentation artifacts that would evidence each.

Summarize the obligations common to major AI governance frameworks and regulations, including risk-tiered approaches, impact assessments, and documentation duties, without overstating any single jurisdiction's requirements.

Conduct a structured risk assessment for a hypothetical AI system, classifying it by risk tier and listing required controls, monitoring, and approval gates.

Translate a business problem into an AI solution design covering data sources, model choice, integration points, latency and cost constraints, and a fallback path for model failure.

Compare deployment patterns including batch, real-time API, edge, and embedded inference, and select an appropriate pattern for three scenarios with different latency and privacy needs.

Draft the components of a model card and a system architecture diagram description that a reviewer could use to assess fitness for purpose.

Define operational monitoring signals for a deployed model, including input drift, prediction drift, data quality checks, latency, error rates, and cost per inference, and specify a threshold and alert for each.

Distinguish data drift, concept drift, and upstream pipeline breakage, and describe a detection method and a response playbook for each.

Explain shadow deployment, canary release, and A/B testing for model updates, and state the rollback criteria you would set for each.

Describe a retraining policy covering triggers, cadence, validation gates, and versioning, and justify why calendar-based retraining alone is insufficient.

Outline an incident response process for a model producing harmful or badly degraded outputs, including detection, containment, communication, and post-incident review.

Build a business case for an AI initiative that states the problem, baseline, expected value, cost drivers, key risks, and success metrics in business terms rather than model terms.

Sequence an AI project across discovery, data readiness, modeling, deployment, and monitoring phases, assigning realistic durations, owners, and go or no-go decision points.

Identify stakeholder, change-management, and adoption risks for an AI rollout, and propose concrete mitigations for each.

Complete a full-length timed practice exam under test conditions and categorize every missed question by domain.

Close knowledge gaps by re-studying your two weakest domains and rewriting the explanations for every missed question in your own words.