eduKateSG · Why Science?
Turn data into accountable intelligence—connecting models, computing, scientific discovery, privacy and real decisions
A fact-checked guide to A*STAR IAIC and the evidence practices behind advanced AI, secure analytics and computational science.
Singapore’s government directory currently lists the Institute of Advanced Intelligence and Computing as an A*STAR research entity. Recent A*STAR Research reports identify IAIC work in multivariate time-series graph learning and privacy-preserving federated health analytics.
A current IAIC recruitment page also describes AI-driven materials modelling that connects first-principles calculations, machine-learning interatomic potentials, high-throughput workflows, continuum models and experimental validation. These are concrete examples, not a claim that every IAIC project has the same scope.
Did you know? A model can be accurate on yesterday’s data and still fail tomorrow when sensors, populations, products or operating conditions change.
Official sources and current entity status were checked 11 November 2026. This guide explains research evidence and offers no access, admission or deployment approval.
Chapter index
Chapters 1–9: advanced intelligence, computing, datasets, baselines, time-series AI, graph neural networks, multimodal learning, uncertainty, evaluation
Chapters 10–17: high-performance computing, AI for science, materials modelling, privacy-preserving analytics, trustworthy AI, deployment, monitoring and evidence receipts
An intelligence problem links observations to a prediction, representation, plan or decision. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 1 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Define the target, user, consequence, time horizon and baseline. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A vague request for AI cannot define success. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Frame an intelligence question” as a safe evidence exercise. Write a decision statement. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 1 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
A data receipt records origin, meaning, permissions, sampling, transformations and known gaps. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 2 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Create a schema, provenance map and versioned data card. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Large data can still be biased, stale or unlawful to use. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Build a data receipt” as a safe evidence exercise. Draft a fictional data card. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 2 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Baselines reveal whether complexity improves on simple rules, statistics or current practice. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 3 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Compare naive, linear and domain-informed alternatives under the same split. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A better average score may hide worse outcomes for critical cases. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Start with baselines” as a safe evidence exercise. Design a baseline ladder. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 3 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Reliable evaluation matches metrics to error costs, prevalence, uncertainty and intended use. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 4 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Use held-out tests, calibration, subgroup checks and confidence intervals. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
One benchmark cannot establish general intelligence or safety. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Evaluate beyond accuracy” as a safe evidence exercise. Build an evaluation matrix. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 4 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
| Evidence move | What to record | What it cannot prove alone |
|---|---|---|
| Intelligence question question | Population or system, variable, unit, comparator and time window | A definition does not establish a result |
| Quality check | Calibration, controls, uncertainty and version | One successful run does not prove general reliability |
| Decision boundary | Owner, trigger, review and escalation | A model or measurement does not replace accountable judgement |
Time-series models use order, trends, cycles and lagged relationships across measurements. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 5 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Preserve chronology and test across future periods and operating regimes. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Random shuffling can leak future information into training. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Learn from time series” as a safe evidence exercise. Create a time-aware split. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 5 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Graph models represent entities and changing links such as sensors, molecules or interactions. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 6 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Define nodes, edges, timestamps, direction and missing links. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A learned edge weight does not prove a causal connection. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Model graph relationships” as a safe evidence exercise. Draw a temporal graph. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 6 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Multimodal AI integrates text, audio, images, sensors or structured records for a stated task. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 7 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Align samples, timestamps and labels and test each modality’s contribution. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
More modalities can add noise, bias or privacy risk. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Combine modalities” as a safe evidence exercise. Plan a modality ablation. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 7 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Uncertainty describes limits in data, parameters, predictions and distribution shifts. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 8 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Calibrate probabilities and separate known from out-of-scope cases. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A confident probability is not automatically trustworthy. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Quantify uncertainty” as a safe evidence exercise. Draft an abstention rule. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 8 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
| Evidence move | What to record | What it cannot prove alone |
|---|---|---|
| Model evidence question | Population or system, variable, unit, comparator and time window | A definition does not establish a result |
| Quality check | Calibration, controls, uncertainty and version | One successful run does not prove general reliability |
| Decision boundary | Owner, trigger, review and escalation | A model or measurement does not replace accountable judgement |
HPC enables large simulations, models and searches by coordinating processors, memory and storage. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 9 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Record hardware, software, precision, scaling and energy use. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
More compute cannot correct a flawed target or invalid data. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Use high-performance computing” as a safe evidence exercise. Create a reproducibility manifest. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 9 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
AI for science can emulate simulations, prioritise experiments and search large design spaces. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 10 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Join models to physical constraints, domain knowledge and confirmatory tests. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A generated candidate is not a discovered material or medicine. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Accelerate science with AI” as a safe evidence exercise. Build a hypothesis loop. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 10 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Materials modelling can connect electronic structure, atoms, microstructure and component behaviour. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 11 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Validate each scale and the transfer of properties between them. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A potential trained on limited structures may fail outside its domain. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Model materials across scales” as a safe evidence exercise. Map a multiscale workflow. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 11 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Federated and cryptographic analytics aim to extract approved results without centralising raw records. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 12 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Define participants, threat model, leakage tests and governance. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Keeping data local does not eliminate privacy or security risk. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Protect distributed data” as a safe evidence exercise. Draft a secure-computation map. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 12 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
| Evidence move | What to record | What it cannot prove alone |
|---|---|---|
| Scientific computing question | Population or system, variable, unit, comparator and time window | A definition does not establish a result |
| Quality check | Calibration, controls, uncertainty and version | One successful run does not prove general reliability |
| Decision boundary | Owner, trigger, review and escalation | A model or measurement does not replace accountable judgement |
Trustworthy AI combines technical performance with robustness, safety, security, fairness and accountability. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 13 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Use red teams, misuse cases, audits and escalation paths. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A principles document cannot prove a system behaves responsibly. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Engineer trustworthy systems” as a safe evidence exercise. Create an assurance case. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 13 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Contextual validation tests the complete workflow with representative users, interfaces and constraints. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 14 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Measure human-AI performance, latency, workload and fallback behaviour. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Offline model metrics cannot establish operational benefit. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Validate in context” as a safe evidence exercise. Plan a shadow evaluation. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 14 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Deployed systems need checks for data drift, concept drift, incidents, feedback loops and changing goals. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 15 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Set thresholds, owners, logs, rollback and review intervals. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Retraining on new data can repeat or amplify new bias. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Monitor change” as a safe evidence exercise. Build a monitoring dashboard. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 15 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
Translation connects research prototypes to maintained products, public services or scientific platforms. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 16 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Define ownership, interoperability, security, documentation and lifecycle cost. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
A paper or demo cannot guarantee sustainable adoption. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Translate research responsibly” as a safe evidence exercise. Draft a translation gate. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 16 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
The receipt links question, data, code, compute, evaluation, limitations and accountable decisions. Advanced intelligence and computing turns raw observations into models, predictions and decisions while keeping the evidence chain visible. Chapter 17 asks what kind of intelligence is needed, what computing makes it possible and which real-world outcome is being served. A*STAR currently lists IAIC as a research entity, and its recent work spans time-series modelling, privacy-preserving analytics and AI-driven materials discovery.
Archive versions, tests, failures, reviews and monitoring changes. A defensible project defines the population or physical system, input variables, target, time window, comparator, data provenance, compute environment and decision owner before training. Teams separate development, validation and test data; record versions and random seeds; measure uncertainty; and challenge a useful-looking result with baselines, ablations, failure cases and independent evidence.
Documentation cannot make an unsafe use acceptable by itself. A high benchmark score does not establish causation, safety, fairness, scientific truth or readiness for deployment. Models can learn leakage, historical bias, shortcuts or unstable correlations. Sensitive data requires lawful governance and privacy protection, while scientific and industrial decisions need domain experts, security review, human oversight and evidence matched to the consequence of error.
Use “Keep an intelligence evidence receipt” as a safe evidence exercise. Assemble a one-page receipt. Work with an open or fictional dataset, make a data card, choose one simple baseline and define an error analysis before trying a complex model. Record who could be missed or harmed, which assumptions could fail, and what evidence would trigger retraining, escalation or a decision not to deploy.
Did you know? More parameters do not automatically mean more intelligence. A compact model with appropriate data, a clear target and a calibrated uncertainty estimate can be more useful than a larger system whose behaviour cannot be tested in the intended setting. Chapter 17 shows why computing is valuable when it makes reasoning more reproducible, inspectable and connected to measurement.
The capability pathway includes mathematics, statistics, computer science, software engineering, data governance, high-performance computing, cryptography, human-computer interaction and domain science. Students can begin with graphs, probability, algorithms, experimental design and careful writing. Current courses and roles set their own prerequisites; no article can guarantee admission, qualification, employment or model approval.
| Evidence move | What to record | What it cannot prove alone |
|---|---|---|
| Intelligence evidence question | Population or system, variable, unit, comparator and time window | A definition does not establish a result |
| Quality check | Calibration, controls, uncertainty and version | One successful run does not prove general reliability |
| Decision boundary | Owner, trigger, review and escalation | A model or measurement does not replace accountable judgement |
Contents · Previous chapter · Continue in the Science Learning Hub
Frequently asked questions
What is IAIC?
Singapore’s government directory lists the Institute of Advanced Intelligence and Computing as an A*STAR research entity. Current A*STAR publications show work in time-series AI, secure federated analytics and AI-driven materials modelling.
What is multivariate time-series learning?
It studies several measurements recorded across time, including delayed and cross-sensor relationships. A useful evaluation preserves time order and checks performance under changing conditions.
What is privacy-preserving federated analytics?
It allows parties to compute approved analyses across distributed data while aiming to avoid centralising raw records. Privacy still depends on protocol, threat model, governance, implementation and access controls.
Can AI prove a scientific theory?
No. AI can find patterns, approximate models and suggest hypotheses, but scientific claims still require appropriate experiments, mechanisms, uncertainty and independent evidence.
Which school subjects support advanced intelligence and computing?
Mathematics and computing are central, while physics, chemistry, biology, engineering, languages and social science provide domain questions and ways to test impact.
Primary sources and verification date
Source status for Why Science? | Institute of Advanced Intelligence and Computing was checked 11 November 2026. Research priorities, programmes, facilities and participation routes can change; use the linked official owner for the current edition.
- Singapore Government Directory — IAIC — Current official directory entry and leadership for the A*STAR research entity.
- A*STAR Research — Catching cross-sensor linkups — Current IAIC example in multivariate time-series and graph neural learning.
- A*STAR Research — Crunching numbers behind closed doors — Current IAIC example in privacy-preserving federated health analytics.
- A*STAR Careers — AI-driven Materials Modelling at IAIC — Current official description of DFT, machine-learning potentials, high-throughput workflows, HPC and validation.
- A*STAR Scientific Equipment & Services Finder — Current official listing of IAIC among A*STAR science and engineering research entities.
Continue the evidence journey
Advanced intelligence becomes useful when data, algorithms, computing, domain evidence, privacy and accountable decisions stay connected from question to monitoring. Families considering science tuition or a science tutor can use the same evidence questions: what will the learner practise, what feedback will they receive, and how will progress be shown?