WBCHSE Applied AI cheat sheet

AI is the field of building systems that perform tasks associated with human intelligence, including perception, reasoning, learning, planning, and decision-making.

The Turing Test, proposed by Alan Turing in 1950, judges a machine by whether its text responses are indistinguishable from a human's.

Narrow AI performs a specific task; general AI would match humans across many tasks; deployed systems today are narrow AI.

AI is the broad field, machine learning is a subset that learns from data, and deep learning is a subset of ML using multilayer neural networks.

Symbolic AI uses explicit rules and knowledge bases; connectionist AI uses learned numerical weights, as in neural networks.

An expert system combines a knowledge base of rules with an inference engine; MYCIN was an early medical diagnosis example.

A rational agent perceives its environment through sensors and acts through actuators to maximise a defined objective.

Search algorithms such as breadth-first, depth-first, and A* explore states; A* uses cost-so-far plus an admissible heuristic.

Algorithmic bias occurs when training data or design choices systematically disadvantage a group; it can persist without malicious intent.

Fairness metrics include demographic parity, equal opportunity, and equalised odds; they can conflict, so the metric must match the context.

Transparency means disclosing how a system works; explainability means giving reasons a specific decision was made.

Data minimisation collects only data needed for a stated purpose; consent and purpose limitation are core privacy principles.

Pseudonymisation replaces identifiers but can be reversible with a key; anonymisation aims to make re-identification practically impossible.

India's Digital Personal Data Protection Act, 2023, uses terms such as data principal, data fiduciary, and consent manager.

Human-in-the-loop review is required for high-stakes decisions such as credit, health, and school admissions when AI informs them.

Deepfakes and synthetic media raise consent, defamation, and election-integrity risks; watermarking and provenance metadata help detection.

Accountability requires a named owner, audit logs, and a remedy path when an AI system causes harm.

Computational thinking decomposes a problem, recognises patterns, abstracts irrelevant detail, and designs an algorithm.

Python is dynamically typed; common types are int, float, str, bool, list, tuple, dict, and set.

Lists are mutable, tuples immutable, sets store unique elements, and dictionaries store key-value pairs with hashable keys.

// performs floor division, % returns the remainder, and ** raises to a power; multiplication and division bind before addition.

A for loop iterates over an iterable; while repeats while a condition is true; break exits and continue skips to the next iteration.

Python sequences are zero-indexed; s[start:stop:step] excludes stop, and negative indices count from the end.

input() returns a string, so numeric input needs int() or float(); range(n) yields 0 to n-1.

Common exceptions include SyntaxError, NameError, TypeError, ValueError, IndexError, and KeyError; each signals a distinct failure.

with open('file.txt', 'r') as f: closes the file automatically; modes include r, w, a, and b.

Structured data fits rows and columns; unstructured data includes free text, images, audio, and video.

Nominal categories have no order, ordinal categories have order; discrete data is countable, continuous data is measured.

Data cleaning handles missing values, duplicates, inconsistent formats, outliers, and wrong types before analysis.

Mean is sensitive to outliers, median is robust, and mode is the most frequent value; report the measure that fits the distribution.

Variance and standard deviation measure spread around the mean; range is maximum minus minimum.

A bar chart compares categories, a histogram shows a numeric distribution, a line chart shows trend over time, and a scatter plot shows association between two numeric variables.

Pearson correlation r ranges from -1 to 1; r near 0 means weak linear association, and correlation does not prove causation.

A truncated y-axis can exaggerate differences; label axes, units, and sample size, and avoid pie charts for many categories.

In pandas, read_csv() loads a file, head() previews rows, info() shows dtypes and non-null counts, and isnull().sum() counts missing values.

Machine learning fits a model from data to make predictions or find patterns rather than relying only on hand-written rules.

Supervised learning uses labelled examples, unsupervised learning uses unlabelled data, and reinforcement learning learns from reward signals.

Classification predicts a categorical label; regression predicts a continuous numeric value.

In supervised learning, X is the feature matrix and y is the target; training, validation, and test sets must remain separate.

Data leakage occurs when information from the test set or target leaks into training features, producing inflated evaluation scores.

Overfitting fits training noise and fails on new data; underfitting is too simple; regularization and more representative data reduce overfitting.

Accuracy is misleading on imbalanced classes; use precision, recall, F1, and the confusion matrix.

Precision = TP/(TP+FP), recall = TP/(TP+FN), and F1 is the harmonic mean of precision and recall.

Common algorithms include linear regression, logistic regression, decision trees, k-nearest neighbours, and k-means clustering.

An NLP pipeline commonly performs tokenisation, lowercasing, stop-word removal, stemming or lemmatisation, vectorisation, then modelling.

Stemming cuts words using crude rules; lemmatisation uses vocabulary and grammar to return a valid base form.

Bag-of-words counts term occurrences and ignores order; TF-IDF weights a term by frequency in a document and rarity across documents.

Word embeddings such as Word2Vec and GloVe map words to dense vectors so similar meanings lie close together.

NLP tasks include sentiment analysis, named-entity recognition, text classification, machine translation, summarisation, and chatbots.

A digital image is a grid of pixels; colour images usually have red, green, and blue channels with values from 0 to 255.

A CNN typically stacks convolution layers, pooling layers, and fully connected layers; convolution learns local visual features.

Image classification assigns one label to an image, object detection locates objects with bounding boxes, and segmentation assigns a label to each pixel.

Data augmentation such as flipping, rotating, cropping, and brightness changes increases training variety and can reduce overfitting.

In healthcare, AI supports medical imaging, triage, and drug discovery, but clinical responsibility remains with qualified professionals.

In agriculture, AI supports crop-disease detection from leaf images, yield forecasting, and precision irrigation.

In finance, AI supports fraud detection, credit scoring, and trading; credit decisions need explainability and fair-treatment checks.

In manufacturing, predictive maintenance estimates equipment failure, and computer vision inspects product defects.

In transport, AI supports route optimisation, traffic prediction, and driver-assistance features; autonomous driving still requires safety validation.

In retail, recommendation systems rank products, demand forecasting plans inventory, and dynamic pricing adjusts to demand.

In education, adaptive learning adjusts content difficulty, and automated feedback speeds practice, but bias and over-reliance need review.

In public services, chatbots handle routine queries and document processing speeds applications; accessibility and language coverage matter.

A domain fit test asks whether labelled data, a measurable objective, and a safe fallback exist before deploying AI.

NumPy provides fast n-dimensional arrays; pandas provides DataFrames; Matplotlib and Seaborn create plots.

scikit-learn provides classical ML algorithms and metrics; TensorFlow/Keras and PyTorch build deep neural networks.

NLTK and spaCy support NLP; OpenCV supports image and video processing.

Jupyter Notebook runs code in cells and keeps outputs; Google Colab provides cloud notebooks with optional GPU access.

An API exposes a service through requests and responses, often in JSON; API keys must not be hard-coded in public repositories.

Cloud AI services offer pretrained vision, speech, and language models on a pay-as-you-go basis, reducing initial infrastructure.

Large language models generate text using learned statistical patterns and can hallucinate; outputs need verification for factual use.

Edge AI runs models on local devices, reducing latency and data transfer, while cloud AI offers more scalable compute.

Federated learning trains across devices without centralising raw data; explainable AI methods help users understand model decisions.

Create a data dictionary that defines each field, its type, units, allowed range, and missing-value meaning.

Use a virtual environment and record dependencies in requirements.txt so the project can be rebuilt.

Set random seeds and pin library versions to make experiments reproducible.

Read a Python traceback from the last line upward to identify the exception, then inspect the referenced line.

Check shape, dtypes, unique values, ranges, duplicates, and missing counts before modelling.

Follow PEP 8 naming, keep functions small, and add docstrings that state inputs, outputs, and side effects.

Label every chart with a title, axis names, units, and a legend when multiple series appear.

Apply identical preprocessing to training and test data; fit scalers and encoders on training data only.

Save a trained model with joblib or pickle and record the preprocessing steps needed to use it.

A project moves through problem definition, data collection, cleaning, modelling, evaluation, documentation, and presentation.

A SMART goal is specific, measurable, achievable, relevant, and time-bound; a baseline gives a comparison point.

An MVP is the smallest version that tests the core value proposition with real users.

A README should state the problem, data sources, setup steps, results, limitations, and ethical considerations.

AI job roles include data analyst, data scientist, ML engineer, NLP engineer, computer vision engineer, and AI ethics reviewer.

A portfolio project is stronger when it shows a clear problem, reproducible code, evaluation metrics, and a demo.

A value proposition states who the customer is, what pain is solved, and why the solution is better than alternatives.

Market validation tests demand through interviews, landing-page tests, or pilot users before large investment.

A pitch deck normally covers problem, solution, market, product, traction, team, financials, and the ask.