Emerging technologies

Emerging technologies

A family of related ideas

• Artificial intelligence (AI): broad field of systems performing tasks associated with intelligent behaviour.
• Machine learning (ML): methods that learn patterns from data to make predictions or decisions rather than relying only on hand-written rules.
• Neural network: an ML model made from connected layers of weighted processing units.
• Deep learning: ML using neural networks with multiple processing layers to learn complex representations.
• Natural language processing (NLP): methods for analysing, generating or interacting through human language; these may use rules, statistical ML or deep learning.
• Reinforcement learning: an agent learns a policy by interacting with an environment and receiving rewards or penalties.

Classify a description

A system learns to label road images using thousands of examples and a multi-layer neural network. Recognise the most specific terms that apply.

Check your answer

It is machine learning and, because it uses a multi-layer neural network, deep learning. Image classification is not NLP because the input is not human language.

Supervised and unsupervised learning

Approach	Learning signal	Example
Supervised	Examples include target labels or values	Predict whether a sensor record is faulty from labelled history
Unsupervised	No supplied target label; the method seeks structure	Group similar travel patterns
Reinforcement	Rewards from actions in an environment	Learn a signal-control policy in a simulation

These labels describe how learning occurs, not whether a system is accurate, fair or safe.

Training and inference

1. Define the task: specify input, desired output and success criteria.
2. Collect and prepare data: clean, label, encode and split records appropriately.
3. Train: adjust model parameters to reduce error on training examples.
4. Validate: tune decisions using separate evidence.
5. Test: estimate performance on unseen examples.
6. Infer: apply the trained model to new input.
7. Monitor: detect changing data, failures and unequal outcomes.

Training or inference?

A model adjusts weights using labelled pothole images, then later assigns a severity category to a new image. Determine which event is training and which is inference.

Check your answer

Adjusting weights from labelled examples is training. Applying the resulting model to the new image is inference.

Data shapes model behaviour

A model learns statistical patterns present in its data. Problems include:

• missing or incorrect labels
• under-representation of relevant conditions or groups
• historical decisions that encode unfairness
• features that indirectly reveal protected or sensitive information
• training data that differs from real operating conditions
• data drift as behaviour or environments change.

“More data†is not automatically better. Relevance, quality, coverage, lawful handling and representativeness matter.

Analyse dataset coverage

A road-damage model is trained only on clear daytime photos from one city. It will be used across Queensland at night, in rain and on unsealed roads.

Analyse the risk and Recommend a refinement.

View a sample solution

The training data does not represent lighting, weather, surface and regional variation in deployment, so test accuracy may not transfer to real use. Collect and label representative examples across those conditions, keep a separate test set and report performance by condition rather than only one overall value.

Neural networks and deep learning

A neural network combines inputs through weighted connections and transformations. Training changes weights to reduce a defined error. Hidden layers can learn intermediate representations useful for the output.

• Input layer receives encoded features.
• Hidden layers transform combinations of values.
• Output layer produces scores, categories or numerical predictions.

Deep networks can model complex image, audio and language patterns but may require substantial data and computation. Their internal reasoning can be difficult to explain, which matters in high-consequence decisions.

Natural language processing

NLP systems process language tasks such as classification, search, translation, summarisation, speech interaction and response generation.

Language is ambiguous and context-dependent. A support classifier may misread slang, negation, multilingual text or an uncommon request.

Input	Possible NLP process	Output
“I can't access my bookingâ€	Intent classification	account_access
Long incident report	Information extraction	location, time, category

Design for uncertainty

Recommend one interface response when a chatbot has low confidence in the user's intent.

View a sample solution

State that the request was not understood, present a small set of likely options and provide a clear path to a human or standard form. Do not confidently perform a high-impact action from an uncertain classification.

Reinforcement learning

An agent observes a state, chooses an action, receives a reward and updates its policy to increase expected future reward.

state → policy chooses action → environment changes → reward → learning

The reward function matters. A traffic agent rewarded only for vehicle throughput might create long pedestrian waits. Simulation, safety constraints and evaluation of unintended strategies are essential.

Evaluate more than accuracy

• Performance: appropriate accuracy, error rates, latency and reliability.
• Fairness: outcomes and errors across relevant groups and conditions.
• Privacy: collection, consent, inference, retention and disclosure of personal data.
• Transparency: users understand when automation is used and what it can affect.
• Accountability: responsibility, review, logging and challenge processes exist.
• Safety: failures are detected, limited and recoverable.
• Security: training data, model, interface and output are protected.
• Environmental and economic cost: computation, infrastructure and ongoing monitoring are justified.

An overall accuracy can hide unequal error types. In a hazard detector, a missed hazard and a false alarm have different consequences.

Use a confusion matrix

	Predicted hazard	Predicted clear
Actual hazard	True positive	False negative
Actual clear	False positive	True negative

Explain why false negatives may deserve special attention in a pedestrian-hazard detector.

Check your answer

A false negative labels an actual hazard as clear, so the system may fail to trigger a protective action. Its safety consequence may be greater than the inconvenience caused by some false alarms.

Human oversight

Human involvement should be designed, not added as a slogan.

• Show the relevant input, output and uncertainty to the reviewer.
• Give reviewers time and authority to disagree.
• Record overrides and outcomes for monitoring.
• Provide a path for affected people to question or correct data.
• Do not overload one person with more alerts than can be meaningfully reviewed.

Independent practice: event safety classifier

An organisation proposes an NLP system that classifies incoming event messages as routine, urgent or emergency. Historical labels were created by one team; messages in languages other than English are rare in the training set; emergency messages are only 1% of examples; and an emergency classification automatically locks venue gates.

1. Analyse four risks in the data, output or automated action.
2. Evaluate whether overall accuracy alone is a suitable criterion.
3. Recommend three controls covering testing, interface design and human oversight.

View a sample solution

• One team's labels may be inconsistent or encode narrow interpretations; language coverage is unrepresentative; rare emergencies can be missed while overall accuracy stays high; and automatic locking could create a physical safety risk from a false positive.
• Overall accuracy is insufficient because routine examples dominate. Report the confusion matrix, emergency recall, false-lock rate and performance by language and scenario.
• Review and diversify labels and test data; show category, confidence and original message to a trained reviewer; require confirmation or a safety-constrained procedure before gate control, with logging and a manual override.