Software Automation (Year 12)

AI — Artificial intelligence is the broad field of making machines perform tasks that normally need human intelligence (reasoning, perception, language, decisions).

ML — Machine learning is a SUBSET of AI: algorithms that learn patterns from data and improve with experience instead of being explicitly programmed with every rule.

Distinguish — AI = the goal/broad concept (includes rule-based systems). ML = one method of achieving AI by learning from data. All ML is AI; not all AI is ML.

Automation — Using software to perform tasks with little or no human input. ML lets automation handle variable, data-driven tasks, not just fixed rules.

DevOps — Combines development and operations to deliver software faster and more reliably through CI/CD pipelines. 2025 Q11 principles: rapid feedback, automation of repetitive tasks, collaboration and shared ownership. ML supports it by predicting build failures, detecting anomalies in logs and prioritising tests.

RPA — Robotic process automation uses software 'bots' to copy what a person does on a user interface (rule-based, repetitive, structured tasks such as copying data between systems or processing invoices). Adding ML lets bots read unstructured data (e.g. emails, scanned forms).

BPA — Business process automation automates an entire end-to-end workflow across systems (e.g. staff onboarding, approvals). ML supports decisions in the workflow (e.g. fraud detection, routing, demand forecasting).

RPA v BPA — RPA automates individual repetitive tasks by mimicking user actions. BPA automates whole processes and the flow of information between them.

MLOps (course spec) — The automated process of designing, training and deploying ML models, borrowing DevOps principles and bringing together model developers and operations teams. THREE stages:

MLOps - Design — Define the business problem; refactor it into a machine learning problem; define success metrics; research available data.

MLOps - Model development — Data wrangling; feature engineering; model training; model testing and validation.

MLOps - Operations — Model deployment; supporting operations/use; monitoring model performance.

Supervised — Trains on LABELLED data (inputs with known correct outputs) to predict labels for new data. Used for classification and regression: spam detection, house price prediction. 2025 Q3 answer: supervised learning.

Unsupervised — Finds patterns/structure in UNLABELLED data without being told the answers. Clustering (customer segments) and anomaly detection. (K-means is the standard unsupervised clustering example.)

Semi-supervised — A small amount of labelled data plus a large amount of unlabelled data. Useful when labelling is expensive (e.g. medical images).

Reinforcement — An agent learns by trial and error in an environment, receiving rewards or penalties, to maximise total reward (game-playing AI, robotics, self-driving).

Remember — Labelled = supervised. No labels = unsupervised. Few labels = semi-supervised. Rewards = reinforcement.

FLAG — The school notes group K-nearest neighbour under unsupervised learning. KNN is normally taught as a supervised algorithm (it uses labelled neighbours). K-means is the unsupervised one. If a question asks, write supervised.

Data analysis and forecasting — ML finds patterns in historical data to predict future values (sales, weather, demand, stock levels). Regression and time-series models are common.

Virtual personal assistants — Siri, Alexa and Google Assistant use ML for speech recognition, natural language processing (understand intent), personalised responses and learning user preferences. Neural networks are central.

Image recognition — ML (usually convolutional neural networks) classifies images and detects objects: face unlock, medical scans, quality inspection, self-driving cars. Needs large, diverse, correctly labelled training sets.

Decision tree — A flowchart-like model: each internal node tests a feature, each branch is an outcome of the test, and each leaf is a final decision/class. Reads from the root down to a leaf.

Advantages — Easy to understand and explain, handles numeric and categorical data, little data preparation.

Disadvantages — Can overfit and become too complex (pruning/limiting depth helps), unstable to small data changes.

Use — Classification and regression, e.g. loan approval, medical diagnosis.

Drawing tip — Label the decision at each node, label each branch (Yes/No or value ranges), end with clear outcomes. One path from root to leaf per case.

Inspired by the brain — A neural network is made of connected nodes (neurons) in layers that learn patterns from data. It can model non-linear relationships and works with structured and unstructured data (images, text, speech).

Layers — Input layer (receives the data), hidden layers (process it; more hidden layers = 'deep'), output layer (the prediction).

Weights and bias — Each connection has a weight showing how important that input is. Learning means adjusting weights (and biases).

Activation function — Each neuron takes a weighted sum of inputs and passes it through an activation function (e.g. ReLU, sigmoid, tanh) to add non-linearity so the network can learn complex patterns.

Forward propagation — Data flows from the input layer through the hidden layers to the output layer, producing a prediction.

Loss — A loss (cost) function measures the difference between the prediction and the correct answer.

Backpropagation — The error is sent backwards through the network to work out how much each weight contributed, then weights are updated (gradient descent) to reduce the loss.

Course-spec wording — A neural network mimics the processing in the human brain: interconnected nodes (artificial neurones), each able to accept a binary input signal and potentially output a signal to connected nodes. Internal weightings and threshold values for each node are set in the training cycle (exposed to inputs with known responses; regular exposure improves accuracy and pattern matching). In diagrams, thicker lines = stronger weightings = higher priority in the final output. The execution cycle follows the training cycle and uses the values from training to determine the output.

Training cycle — Repeat forward propagation, loss calculation, backpropagation over many epochs until loss is low. Then the trained model is used in execution (prediction on new data).

Strengths and limits — Strengths: handles complex patterns, improves with data. Limits: needs large datasets, heavy computation (GPUs), risk of overfitting, hard to explain ('black box').

Types — Feedforward (FNN), convolutional (CNN, images), recurrent (RNN, sequences such as speech and time series), deep (DNN, many hidden layers).

2025 Q25 model (5 marks) — Explain HOW the neural network supports the VPA using the scenario (recognises language patterns to categorise queries such as lost luggage; nodes process; assesses tone for frustration; weights refine recommendations from customer ratings and turnaround times). THEN one positive AND one negative impact (faster, more accurate support v staff over-rely on recommendations / lose familiarity with trends / raised customer expectations).

Linear regression — Predicts a continuous numeric value from one or more inputs by fitting a straight line: y = mx + c. e.g. house price from size.

Polynomial regression — Fits a curve (y = ax^2 + bx + c) when the relationship is non-linear, e.g. growth or temperature over the year. Too high a degree overfits.

Logistic regression — Despite its name, used for CLASSIFICATION: predicts the probability (0 to 1, via the sigmoid function) of a category, usually yes/no (binary). 2025 Q12: spam detection, two-party election outcome and university admission are logistic regression problems (binary outcomes); predicting a numeric value is not.

K-nearest neighbour (KNN) — Classifies (or predicts) a new point by looking at the K closest data points and using the majority class (or average). Simple, no training phase, but slow on big data and sensitive to the choice of K and to scaling.

Choosing — Number to predict = linear/polynomial. Category/yes-no = logistic or KNN. Straight-line trend = linear; curved trend = polynomial.

Course-spec regression code (linear regression, NumPy and scikit-learn) —

import numpy as np
from sklearn.linear_model import LinearRegression
x = np.array([[2], [4], [6], [8], [10], [12], [14], [16]])
y = np.array([1, 3, 5, 7, 9, 11, 13, 15])
model = LinearRegression()
model.fit(x, y)                      # find line of best fit
print(model.predict([[4]]))          # prints 3
print(model.predict([[4.5]]))        # prints 3.5 (new data)

Spec note — Students must know how to DESIGN programs that use and apply regression but are not expected to implement the algorithms themselves. Pattern: import, create data (x as 2D array), create model, fit, predict.

Regression class (original example, not from NESA) —

from sklearn.linear_model import LinearRegression

class PricePredictor:
    def __init__(self):
        self.model = LinearRegression()

    def train(self, sizes, prices):
        # sizes: [[50], [80], [120]]   prices: [200, 310, 450]
        self.model.fit(sizes, prices)

    def predict(self, size):
        return self.model.predict([[size]])[0]

p = PricePredictor()
p.train([[50], [80], [120]], [200, 310, 450])
print(p.predict(100))

Notes — Uses a class with a constructor, a train method (fit) and a predict method. Logistic regression is the same pattern with LogisticRegression. Polynomial regression adds PolynomialFeatures. (scikit-learn is a common library; the HSC syllabus does not name a library, so write the logic clearly.)

Neural network use — Create the model object, train it on training data, then call predict on new data. Explain what the inputs and outputs are.

Safety of workers — Positive: robots and monitoring remove people from dangerous tasks; sensors and AI detect hazards in real time. Negative: monitoring can invade privacy, false alerts cause stress, and over-reliance can reduce vigilance. (2025 Q24 safety monitoring wearables.)

People with disability — Positive: voice assistants, automatic captioning, smart home control, assistive devices increase independence and inclusion. Negative: poor accessibility or bias can exclude people; cost.

Nature and skills of employment — Routine jobs reduced or changed; new roles (data scientist, ML engineer, automation maintenance) appear; workers need reskilling and digital skills; people shift to oversight and creative or interpersonal tasks.

Production efficiency, waste and environment — Automation increases speed, accuracy and consistency, reduces defects and material waste, optimises energy use (smart grids). But training large models and running data centres uses lots of energy and e-waste.

Economy and wealth distribution — Productivity and growth rise and costs fall; but wealth can concentrate with owners of technology while displaced workers lose income, widening inequality unless supported by training and policy.

Answer pattern for 5 marks — Both positive AND negative ways, each tied to the scenario. 2025 Q24 MG: 'comprehensive discussion... with reference to the scenario' = positive (reduced human error, real-time alerts, PPE compliance, risky behaviour identified), negative (privacy at breaks, misidentified safe behaviour causing stress and distraction, worry about data storage/use).

Psychological responses — Users adopt AI if they trust it, so build transparency and explanations (e.g. a medical AI justifies its diagnosis). Fear of AI creates resistance, so design intuitive, friendly systems. Cognitive bias such as confirmation bias means users favour outputs that match their beliefs.

Acute stress response — Under high stress people have high cognitive load, so AI should give quick, clear, concise outputs (visual alerts, short instructions) not complex data-heavy ones. Example: disaster management systems for first responders.

Cultural protocols — Design to respect cultural practices: language, dialect, gender roles, formal address, traditional values. Example: voice assistants adapted for each country.

Belief systems — Avoid ethical conflict; ensure fairness and inclusivity so the AI does not discriminate by belief, race or culture. Example: facial recognition tested across demographics.

Bias — When data, models or developer decisions unintentionally favour certain outcomes, groups or perspectives.

Human bias — Explicit bias is intentional (a researcher only collects pictures of male physicists). Implicit bias is unconscious, shaped by social and cultural factors. It can enter at design, data collection, labelling and testing.

Dataset (source) bias — Training data that is not diverse or representative (too few examples of some groups, historical prejudice in records, data collected from one region) so the model performs poorly or unfairly for under-represented groups.

Labelling bias — People may interpret the same data differently when labelling it, so inconsistent or biased labels teach the AI the wrong pattern (video example: if most people draw shoes in one style, other shoes are not recognised as shoes).

Effect on usefulness — Inaccurate, unfair or unhelpful results; discrimination (hiring, loans, policing, facial recognition); loss of trust and legal or reputational damage; AI may amplify existing biases.

Mitigation — Diverse and representative training data; audit datasets for balance; diverse development teams; clear labelling guidelines with several reviewers; test performance across groups; transparency and human oversight; continual monitoring and retraining.

2025 Q27 (6 marks) — Needed: EXPLAIN how human bias affects usefulness AND support with reference to the video (male physicists in a photo library; shoes drawn in particular ways). 5 marks = describe + scenario; 4 = outline + refer to video. Always connect bias to the AI's learning from data, then to the effect.

Link to AI learning — One sentence: AI learns from data, so what goes into it shapes what comes out.

Source — Name the source of bias or the automation change, with an example from the stimulus.

Effect — Explain the effect (who is harmed or helped and how).

Both sides — Include a positive AND a negative (for impact questions).

Second example — Another stage or group (labelling, wealth, disability, environment) with the stimulus.

Mitigation or judgement — Finish with a mitigation (diverse data, oversight) or an overall judgement.