ASSEB Class 10 AI & Robotics cheat sheet

AI is the broad goal of machines performing tasks that normally require human intelligence; ML is the subset where models learn patterns from data; deep learning is ML using multi-layer neural networks.

Narrow (weak) AI is built for a specific task such as spam filtering or face unlock; general AI (human-level across domains) and superintelligence remain hypothetical, not deployed.

Rule-based systems encode hand-written if-then logic and are transparent but brittle; learning-based systems infer rules from examples and generalize but need data and are harder to explain.

Supervised learning uses labelled input-output pairs, unsupervised learning finds structure in unlabelled data, and reinforcement learning learns from reward signals through trial and error.

A model's output is a prediction, not a fact: confidence scores express probability, so a 0.7 score means the model is often wrong roughly 3 times in 10 on similar inputs.

Common application categories: classification (which class?), regression (what number?), clustering (what groups?), generation (what content?), and control (what action next?).

Training data, features, algorithm choice, and evaluation metric jointly determine performance; changing the metric can make the same model look good or bad.

AI systems fail predictably when deployment data differs from training data (distribution shift), so monitoring inputs in production is a design requirement, not an afterthought.

Data types: nominal (unordered labels), ordinal (ordered categories), interval (equal spacing, no true zero), ratio (true zero, ratios meaningful); the type restricts which statistics are valid.

Mean is sensitive to outliers; median is the middle value and robust; mode is the most frequent value and the only valid average for nominal data.

Standard deviation measures spread around the mean; variance is its square. A high SD relative to the mean signals heterogeneous data.

Correlation coefficient r ranges from -1 to +1; r near 0 means no linear relationship, and correlation never establishes causation.

Data pipeline order: collection, cleaning (missing values, duplicates, outliers), transformation/normalization, exploration, modelling, evaluation, reporting.

Train/validation/test split is typically 60/20/20 or 70/15/15; the test set is touched once, at the end, to estimate generalization.

Data leakage occurs when information from the target or the test set enters training features, producing unrealistically high scores that collapse in deployment.

Visualization choice rule: bar charts for comparing categories, line charts for trends over time, histograms for distributions, scatter plots for relationships between two numeric variables.

Underfitting = high bias, poor on both training and test data; overfitting = low bias, high variance, good on training but poor on test data.

Regularization (L1/L2), more training data, simpler models, early stopping, and dropout reduce overfitting; a more expressive model or better features reduce underfitting.

Classification metrics: accuracy (misleading on imbalanced classes), precision = TP/(TP+FP), recall = TP/(TP+FN), F1 = harmonic mean of precision and recall.

Confusion matrix cells are true positives, true negatives, false positives (type I), and false negatives (type II); the cost of each error type should drive metric choice.

k-Nearest Neighbours classifies by majority vote of the k closest points; small k gives low bias and high variance, large k the reverse.

Decision trees split on the feature that most reduces impurity (Gini or entropy); deep trees overfit, so pruning and depth limits are used.

Linear regression fits y = wx + b by minimizing mean squared error; logistic regression outputs a probability via the sigmoid and is used for classification.

k-means clustering alternates assignment and centroid update until convergence; k must be chosen in advance, and results depend on initialization (k-means++ mitigates this).

An image is a grid of pixels with numeric channel values (0-255 per channel in 8-bit RGB); convolution slides a small kernel over the grid to detect local patterns such as edges.

CNN building blocks: convolution (feature extraction), pooling (downsampling, translation tolerance), fully connected layers (decision), and softmax (class probabilities).

Image classification assigns one label per image; object detection adds bounding boxes; semantic segmentation labels every pixel by class; instance segmentation separates individual objects.

Data augmentation (flip, crop, rotate, colour jitter) enlarges effective training data and improves invariance without new collection.

NLP pipeline: tokenization, normalization (lowercasing, stemming/lemmatization), stop-word handling, vectorization (bag-of-words, TF-IDF, embeddings), then modelling.

TF-IDF weights a term by frequency in a document divided by how common it is across the corpus, so frequent-but-uninformative words score low.

Word embeddings map words to dense vectors where similar meanings are close, enabling analogy arithmetic; transformers use attention to weigh context across a sequence.

Evaluation: precision/recall/F1 and confusion matrices for classification; BLEU for machine translation; perplexity for language models; IoU (intersection over union) for detection and segmentation.

Robot anatomy: controller (decision), actuators (motion), sensors (perception), power supply, and end effector; degrees of freedom equal the number of independent joints.

Sensor taxonomy: proprioceptive (internal - encoders, IMU) versus exteroceptive (external - ultrasonic, camera, LiDAR); active sensors emit energy, passive sensors receive it.

Ultrasonic distance = (speed of sound x time of flight) / 2; the division by 2 accounts for the round trip, with sound near 343 m/s in air at 20 °C.

IR sensors suit short-range proximity and line following but are affected by ambient light and surface reflectivity; ultrasonic is longer-range but has a wide cone and soft-object errors.

LiDAR gives accurate range and angular resolution for mapping; cameras give rich semantic information but no direct depth without stereo or structured light.

Encoder counts convert to angle or distance via counts-per-revolution and wheel circumference; IMUs drift over time, so sensor fusion (e.g., Kalman filter) combines sources.

Forward kinematics computes end-effector position from joint angles; inverse kinematics solves joint angles for a target pose and may have zero, one, or many solutions.

Control loop: sense, compare to setpoint, compute error, apply correction; a PID controller sums proportional, integral (removes steady-state error), and derivative (damps overshoot) terms.

Microcontroller = CPU + memory (flash for program, RAM for data) + I/O peripherals on one chip; it runs one program repeatedly, unlike a general-purpose computer with an OS.

Arduino sketch structure: setup() runs once for configuration; loop() repeats forever and holds the control logic.

GPIO pins are digital (HIGH/LOW) or analog input via ADC; PWM output approximates analog values by varying duty cycle and is used for motor speed and LED brightness.

DC motors spin continuously and need an H-bridge (e.g., L298N) to reverse direction; servo motors hold a commanded angle using internal feedback; steppers move in discrete steps for precise positioning.

Motor drivers are required because microcontroller pins supply milliamps while motors draw amps; the driver isolates logic power from motor power and handles inductive kickback.

I2C uses two wires (SDA, SCL) with device addresses; SPI uses four wires (MOSI, MISO, SCK, CS) and is faster; UART is asynchronous point-to-point with TX/RX.

Reading a sensor: analogRead returns a value scaled to ADC resolution (0-1023 on a 10-bit ADC); convert to voltage as raw x Vref / 1023.

Power budget rule: sum actuator stall currents, choose a supply with margin, share a common ground between controller and driver, and never power motors from the microcontroller's regulator.

Design cycle: define the problem and success criteria, gather data, build a baseline, iterate, evaluate against the criteria, document, and deploy with monitoring.

A workable problem statement names the user, the input, the output, the environment, and a measurable success threshold (e.g., accuracy or error tolerance).

Baseline first: a simple rule or majority-class model establishes the reference score that any complex model must beat to justify its cost.

Version control and reproducibility: fix random seeds, record dataset versions, hyperparameters, and library versions so results can be re-run.

Hardware-software integration order: test each component alone, then interfaces (sensor read, motor command), then the full loop; isolate faults by substitution.

Iteration log should record hypothesis, change, measured result, and decision - this is the evidence for the project report and viva.

Evaluation must use a held-out set and task-relevant metrics; for a robot, also measure latency, reliability over repeated runs, and failure modes.

Documentation set: block diagram, wiring/pin map, flowchart or pseudocode, dataset description, results table, limitations, and future work.

Bias enters through biased training data, biased labelling, and proxy features; it produces disparate error rates across groups and must be measured, not assumed absent.

Transparency and explainability: stakeholders should be able to learn why a decision was made; high-stakes domains (health, credit, hiring, justice) demand stronger justification.

Privacy principles: data minimization, purpose limitation, informed consent, anonymization, and secure storage; personal data includes images, voice, and location traces.

Accountability requires a named human owner for each automated system, plus an escalation path for contested decisions and a right to human review.

Safety engineering: fail-safe defaults, emergency stop, speed and force limits, guarded workspaces, and watchdog timers for lost communication.

Automation can displace tasks and create new roles; the societal debate covers job transition, accessibility gains, and concentration of data power.

Deepfakes and synthetic media raise consent and misinformation risks; provenance labelling and detection tools are partial mitigations.

Deployment checklist: intended use and misuse analysis, bias and safety testing, human oversight plan, incident reporting, and periodic audit after release.