IoT Analytics, AI and Intelligent Decision-Making

Introduction and Context

An Internet of Things (IoT) system becomes strategically valuable when it converts observations into reliable decisions. Sensors generate measurements, communication networks transport them, and analytics interprets them to support actions such as adjusting a process, predicting equipment failure, optimizing energy use, or detecting unsafe conditions. Consequently, IoT analytics is not merely the application of machine-learning algorithms to sensor data. It is an end-to-end discipline involving sensing, data engineering, statistical reasoning, artificial intelligence, decision theory, control, cybersecurity, and governance.

IoT environments differ from conventional information systems in several important respects. Data are generated continuously, often at high velocity and with heterogeneous formats. Devices may have limited energy, memory, processing capability, and connectivity. Measurements are affected by calibration errors, missing values, drift, interference, and changing operating conditions. Decisions may also be time-critical: a safety system cannot wait for a cloud batch job if an actuator must respond within milliseconds.

This chapter develops a framework for descriptive, diagnostic, predictive, and prescriptive analytics in IoT. It then examines machine learning, edge intelligence, TinyML, federated learning, anomaly detection, forecasting, optimization, and intelligent automation. Particular attention is given to the relationship between model performance and operational usefulness: a highly accurate offline model may still be unsuitable if it is too slow, opaque, energy-intensive, or poorly calibrated for deployment.

Learning Objectives

After studying this chapter, the learner should be able to:

1. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics in IoT applications.

2. Design an IoT analytics pipeline from data acquisition to operational decision-making.

3. Select appropriate statistical and machine-learning models for forecasting, classification, regression, and anomaly detection.

4. Explain edge AI, TinyML, and federated learning, including their benefits and limitations.

5. Evaluate IoT models using technically appropriate metrics, validation procedures, and deployment constraints.

6. Formulate intelligent decisions using prediction, optimization, constraints, uncertainty, and human oversight.

7. Critically assess reliability, privacy, fairness, explainability, security, and regulatory implications.

1. The Analytics Value Chain in IoT

A useful IoT analytics architecture contains the following logical stages:

Sensing and acquisition: Sensors measure physical variables such as temperature, vibration, pressure, current, location, or image intensity. Sampling frequency should reflect the dynamics of the monitored phenomenon. Sampling too slowly can conceal important events; sampling too rapidly increases storage, communication, and energy costs.

Data preparation: Data are time-stamped, synchronized, filtered, validated, transformed, and associated with device and location metadata. Preparation may include unit conversion, resampling, interpolation, outlier handling, feature extraction, and quality flags.

Data transport and storage: Data may move through gateways to cloud platforms, local servers, or edge nodes. Stream processing is suitable for immediate decisions, whereas historical storage supports trend analysis and model training.

Analytics and inference: Statistical models or machine-learning models estimate states, detect events, forecast future values, classify conditions, or infer hidden variables.

Decision and actuation: An inference becomes operationally meaningful only when it informs a decision. The decision may be automatic, recommended to an operator, or used to trigger a control policy.

Feedback and learning: Outcomes are recorded. These outcomes support model monitoring, recalibration, retraining, and assessment of whether the intervention produced the intended result.

A common design error is to optimize the analytics stage while ignoring the complete loop. For example, a predictive-maintenance model may achieve excellent accuracy in a laboratory dataset but have little value if sensors cannot capture the relevant vibration frequency or if maintenance teams cannot act before failure.

2. Four Forms of Analytics

2.1 Descriptive Analytics: What Has Happened?

Descriptive analytics summarizes observed or historical data. Typical outputs include dashboards, time-series plots, distributions, percentiles, correlation matrices, device-health indicators, and energy-consumption profiles. A mean summarizes central tendency, while a median is generally more robust to extreme values. The sample mean of observations x1, x2, ..., xn is

x̄ = (1/n) Σ xi.

The sample standard deviation is

s = √[(1/(n−1)) Σ(xi − x̄)²],

where n is the number of observations. The denominator n−1 is used when estimating population variability from a sample under the usual independent-sampling assumptions.

In IoT, descriptive analytics should include data-quality information. A dashboard reporting average temperature without indicating missing-data percentage, sensor identity, calibration status, and time window may encourage misleading conclusions. Aggregation can also hide short-duration peaks that are operationally important.

2.2 Diagnostic Analytics: Why Did It Happen?

Diagnostic analytics investigates causes and contributing conditions. Techniques include drill-down analysis, stratification by device or location, correlation analysis, change-point detection, causal diagrams, fault trees, and controlled experimentation.

Correlation does not establish causation. A rise in motor temperature may correlate with production rate, but the causal mechanism could be overload, inadequate cooling, ambient temperature, or a failing bearing. Diagnostic analysis should combine data-driven evidence with domain knowledge and, where possible, interventions or natural experiments.

A practical diagnostic workflow is:

1. Define the event and its time window.

2. Verify data integrity and sensor validity.

3. Compare affected and unaffected devices or periods.

4. Examine temporal precedence and operating context.

5. Generate competing causal hypotheses.

6. Test hypotheses using additional measurements, experiments, or engineering models.

7. Record confidence and unresolved uncertainty.

2.3 Predictive Analytics: What Is Likely to Happen?

Predictive analytics estimates an unknown or future quantity. Examples include forecasting electricity demand, estimating remaining useful life, predicting traffic congestion, and classifying whether a machine will fail within a specified horizon.

For a regression model, common metrics include mean absolute error (MAE),

MAE = (1/n) Σ |yi − ŷi|,

and root mean squared error (RMSE),

RMSE = √[(1/n) Σ(yi − ŷi)²],

where yi is the observed value and ŷi is the prediction. Both have the same units as the target. RMSE penalizes large errors more strongly than MAE and is therefore sensitive to rare but substantial misses.

For binary classification, accuracy can be misleading when classes are imbalanced. Precision is TP/(TP+FP), while recall or sensitivity is TP/(TP+FN), where TP, FP, and FN denote true positives, false positives, and false negatives. The F1 score is the harmonic mean of precision and recall:

F1 = 2(precision × recall)/(precision + recall),

provided precision and recall are nonzero. The appropriate metric depends on consequences. In fire detection, missing a fire may be much more costly than issuing a false alarm, so recall and decision costs deserve priority.

Time-series prediction requires chronological validation. Randomly mixing future observations into the training set produces leakage and overestimates performance. A rolling-origin evaluation trains on an earlier period and evaluates on a later period, repeating the process across several time windows.

2.4 Prescriptive Analytics: What Should Be Done?

Prescriptive analytics recommends or selects actions using predictions, objectives, constraints, and costs. If a demand forecast predicts high load, a system may schedule storage discharge, shift flexible loads, or purchase energy. If a model predicts failure risk, a maintenance planner may select an intervention date while considering spare parts, workforce, downtime, and safety.

A generic constrained decision problem can be expressed as

minimize over a: E[C(a, ξ)]

subject to: g(a, ξ) ≤ 0,

where a is an action, ξ represents uncertain conditions, C is cost, and g represents operational constraints. The expectation indicates that uncertainty may be considered explicitly. In deterministic optimization, ξ may be replaced by a forecast or scenario. A valid solution must be feasible, not merely statistically likely.

Prescriptive systems should distinguish prediction from decision. A prediction estimates what may occur; a decision changes what will occur and may consequently invalidate the original prediction. This feedback is particularly important in energy management, traffic control, and adaptive manufacturing.

3. Data Engineering for Reliable IoT Intelligence

3.1 Data Quality and Uncertainty

IoT data are often imperfect. Missingness may result from packet loss, battery depletion, or communication outages. Missing completely at random is uncommon; missingness may depend on temperature, location, workload, or device failure. Replacing all missing values with a mean can suppress variability and conceal failures. Alternatives include forward filling for slowly varying process states, interpolation for appropriate continuous signals, model-based imputation, or explicit missingness indicators.

Sensor drift is a gradual change in measurement behavior. Calibration against a reference, redundant sensing, physical constraints, and statistical monitoring can identify drift. A quality-aware pipeline should preserve the original measurement, corrected value, correction method, and confidence or quality flag.

3.2 Feature Engineering

Features may be raw values, rates of change, moving averages, variances, frequency-domain components, spectral energy, or domain-specific indicators. For vibration monitoring, root-mean-square amplitude and selected frequency-band energy can reveal mechanical changes. For buildings, features may include occupancy-adjusted energy use, indoor-outdoor temperature difference, and recent heating or cooling state.

Feature engineering must respect time. A feature calculated using observations after the prediction time is leakage, even if it appears harmless. Normalization parameters must also be learned from training data and then applied unchanged to validation and deployment data.

3.3 Data Fusion

Combining data from multiple sensors can improve observability. However, fusion requires temporal alignment, spatial registration, unit consistency, sensor-confidence assessment, and resolution management. A fast vibration sensor and a slow temperature sensor cannot be joined without deciding how each measurement represents a time interval. Fusion may be achieved at the raw-data, feature, or decision level.

4. Machine Learning for IoT

4.1 Learning Paradigms

Supervised learning uses labeled examples. Regression predicts numerical values; classification predicts categories. It is effective when representative labels exist, but industrial failure labels may be scarce because failures are deliberately avoided.

Unsupervised learning identifies structure without labels. Clustering can group operating regimes, while dimensionality reduction can reveal latent patterns. Unsupervised results require engineering interpretation.

Semi-supervised and self-supervised learning exploit abundant unlabeled data together with limited labels. A model may learn representations from normal operating sequences and later be adapted to a smaller labeled fault dataset.

Reinforcement learning learns actions through interaction and rewards. It can be useful for resource allocation or control, but unsafe exploration, delayed rewards, and changing environments make direct deployment difficult. A simulator, safety shield, offline training, or human approval may be required.

4.2 Model Selection

Linear regression is interpretable and computationally efficient, but it may fail for nonlinear relationships. Decision trees and ensembles handle mixed variables and nonlinear interactions. Support vector machines can perform well in moderate-dimensional settings. Neural networks are powerful for images, audio, and complex temporal patterns, but require careful regularization, data management, and monitoring.

For sequential IoT data, models include autoregressive methods, state-space models, recurrent networks, temporal convolutional networks, and transformers. Model choice should be based on forecast horizon, data volume, latency, interpretability, energy budget, and failure consequences—not only on benchmark accuracy.

4.3 Overfitting and Generalization

Overfitting occurs when a model learns noise or conditions specific to training data. Controls include simpler models, regularization, early stopping, data augmentation where valid, cross-validation designed for time and devices, and external testing. A model trained on one machine should be tested on other machines or operating regimes if deployment will span them. Device-level splitting is essential when measurements from the same device are highly correlated.

5. Edge AI, TinyML, and Federated Learning

5.1 Edge Intelligence

Edge AI executes inference near the data source rather than sending every measurement to a distant cloud. Advantages include reduced latency, lower bandwidth use, continued operation during intermittent connectivity, and improved data locality. Limitations include restricted compute capacity, difficult software maintenance, hardware heterogeneity, and limited visibility into field behavior.

A partitioned architecture may perform signal filtering and immediate anomaly detection on a microcontroller, aggregate features at a gateway, train models in the cloud, and return compressed model updates or parameters to the edge. The partition should reflect latency, privacy, energy, and reliability requirements.

5.2 TinyML

TinyML refers to machine learning on highly constrained embedded devices. Deployment commonly involves quantization, pruning, knowledge distillation, compact architectures, and hardware-aware optimization. Quantization represents parameters or activations with lower precision, such as 8-bit integers instead of 32-bit floating-point values. It reduces memory and may accelerate inference, but can reduce accuracy, especially when signal ranges are poorly calibrated.

A TinyML design should report memory footprint, inference latency, energy per inference, model accuracy under realistic sensor noise, and update procedures. A model that fits in memory but causes unacceptable battery drain is not a successful design.

5.3 Federated Learning

Federated learning trains models across devices or organizations while retaining raw data locally. In a basic synchronous procedure, a server distributes a global model, participating clients train locally for several steps, and the server aggregates updates. A common weighted aggregation is

w_{t+1} = Σ(k=1 to K) (nk/n) w_{t+1}^{(k)},

where w is the model parameter vector, K is the number of participating clients, nk is the local sample count of client k, and n = Σ nk. This weighted averaging is often called federated averaging.

The method does not automatically guarantee privacy. Model updates can leak information, and malicious clients may poison training. Secure aggregation, differential privacy, robust aggregation, client authentication, and anomaly monitoring can reduce risks. Non-identically distributed data are a major challenge: devices may observe different users, climates, machines, or operating regimes. Communication failures and intermittent participation also complicate convergence.

6. Anomaly Detection and Condition Monitoring

An anomaly is an observation or sequence that deviates from an accepted pattern. It may indicate a fault, attack, sensor error, unusual but legitimate operation, or a change in the environment. Anomaly detection is therefore not equivalent to fault diagnosis.

Statistical methods use thresholds, control charts, residual analysis, or probability models. If a residual is approximately normal with known mean μ and standard deviation σ, a standardized residual may be written as z = (r−μ)/σ. A threshold such as |z| > 3 is a heuristic under appropriate assumptions; it is not a universal law and must be adapted for autocorrelation, multiple testing, and operational costs.

Distance and density methods identify points far from normal clusters or in low-density regions. One-class methods learn a boundary around normal data. Autoencoders learn to reconstruct normal patterns; unusually large reconstruction error may indicate an anomaly. Forecast-residual methods compare observed values with model predictions. No method is universally superior.

A production detector requires hysteresis, persistence rules, alert suppression, severity levels, and a response plan. Triggering an alarm after one noisy sample can create alarm fatigue. Conversely, excessive smoothing can delay detection. Thresholds should be evaluated using false-alarm rate, missed-event rate, detection delay, and operational cost.

7. Intelligent Decision-Making Under Uncertainty

A decision system should represent uncertainty rather than presenting every prediction as certain. Prediction intervals, calibrated probabilities, ensembles, Bayesian models, and conformal prediction can provide uncertainty information. Calibration asks whether events assigned a probability of 0.7 occur approximately 70% of the time over comparable cases.

Decision thresholds should reflect asymmetric consequences. Suppose a predictive-maintenance model produces a probability p of failure before the next inspection. A simple expected-cost rule compares the cost of preventive maintenance with the expected cost of failure. If preventive maintenance costs Cp and failure costs Cf, maintenance is economically preferred when

Cp < p Cf,

under the simplified assumptions that these are the only relevant costs and that maintenance prevents the failure. Real systems must also include downtime, residual risk, spare-part availability, labor, and the possibility of imperfect maintenance.

Automation levels can be designed progressively:

1. Descriptive: display current conditions.

2. Advisory: recommend an action to an operator.

3. Supervised automation: execute routine actions subject to approval or limits.

4. Autonomous control: act automatically within verified safety constraints.

Human oversight is especially important when models operate outside their training distribution or when actions affect safety, health, essential services, or people’s rights.

8. Deployment, Monitoring, and MLOps

An IoT model is a software component embedded in a changing physical system. Deployment requires versioned data schemas, model artifacts, feature definitions, firmware, calibration records, and rollback mechanisms. Models should be tested against simulated failures and adversarial or abnormal inputs before field release.

Monitoring should include:

• input distribution and missingness drift;
• sensor calibration and communication health;
• prediction accuracy when labels become available;
• confidence and calibration;
• latency, memory, energy, and availability;
• alert volume and operator response;
• fairness or performance differences across sites and populations.

Concept drift occurs when the relationship between inputs and outcomes changes. Causes include aging equipment, changed production settings, weather, user behavior, or policy changes. Retraining should not be automatic without validation: a new model can be statistically better but operationally unsafe. A controlled champion–challenger process, staged deployment, and explicit acceptance criteria are recommended.

9. Ethical, Security, and Governance Considerations

IoT analytics can expose personal, industrial, and infrastructure information. Data minimization, purpose limitation, access control, encryption, retention limits, and privacy-preserving computation should be integrated from the design stage. Consent and lawful processing are relevant where personal data are collected, while industrial systems may require sector-specific safety and cybersecurity controls.

Models can reproduce bias if training data underrepresent particular locations, populations, or operating regimes. Evaluation should be stratified where legitimate and appropriate, while avoiding the careless use of sensitive attributes. Explainability should be matched to the decision: a maintenance engineer may require influential sensor channels and trends; a regulator may require traceability, validation evidence, and audit records.

Security threats include spoofed sensor values, replay attacks, adversarial inputs, compromised gateways, model theft, and poisoning of federated updates. Secure boot, signed firmware, device identity, least privilege, network segmentation, secure update mechanisms, anomaly monitoring, and incident response are essential. Safety-critical control should not depend on an unverified AI model alone; independent limits and fail-safe behavior are required.

10. Detailed Illustrative Case: Predictive Maintenance in a Pumping Station

Background

Consider a municipal pumping station containing electrically driven centrifugal pumps. Each pump has sensors for vibration, motor current, discharge pressure, temperature, and operating state. The station seeks to reduce unplanned outages while avoiding unnecessary maintenance.

Problem

Historical failure labels are limited. Operators report that bearing degradation is often preceded by increasing vibration, but vibration also changes with flow rate and valve position. A naive threshold generates many false alarms during high-demand periods.

Analysis

The engineering team first synchronizes sensor clocks and removes intervals marked by maintenance or sensor calibration. Data are divided chronologically, with entire pump units held out for testing. Features include vibration-band energy, vibration trend, motor-current statistics, pressure deviation, and operating regime. The team confirms that features are calculated only from information available before each prediction time.

A two-stage strategy is selected. A condition model estimates expected vibration for the current flow and speed. The residual between observed and expected vibration is then analyzed. A temporal classifier estimates the probability of a maintenance event within the next fourteen days. Because missed failures are costly, recall is emphasized, but an alert-persistence rule requires abnormal behavior over several observations.

Approach and Solution

Inference is performed at the station gateway so that alerts continue during cloud-network outages. Historical data and periodic model retraining occur in the cloud. Operators receive the predicted risk, influential signals, confidence information, and recommended inspection—not merely a binary alarm. The decision policy considers risk, pump redundancy, available technicians, and spare bearings.

Before deployment, the system is tested during controlled load changes. A safety rule prevents automated shutdown based solely on the model; shutdown remains governed by independent protection systems. Performance is monitored using false alarms per pump-month, missed failures, detection lead time, and avoided downtime.

Outcome

In this illustrative scenario, the system reduces nuisance alarms because vibration is interpreted relative to operating condition. It does not eliminate failures, since some failures have no measurable precursor and sensors can themselves fail. Its principal value is improved inspection prioritization and earlier human intervention rather than fully autonomous maintenance.

Learning Points

1. Contextual residuals are often more useful than fixed raw-signal thresholds.

2. Device-level and chronological validation are necessary to estimate deployment performance.

3. A prediction must be connected to a feasible maintenance decision.

4. Edge inference improves continuity but increases update and security responsibilities.

5. Independent safety mechanisms should remain outside the experimental AI loop.

11. Best Practices and Contemporary Developments

Effective IoT intelligence begins with a clearly defined decision, not an algorithm. Teams should specify the prediction horizon, action owner, cost of errors, latency requirement, data-quality limits, and acceptable operating envelope. Baseline rules and simple statistical models should be retained for comparison. Data and model documentation should record provenance, assumptions, training period, known limitations, and intended use.

Contemporary developments include multimodal foundation models for combining text, images, and sensor streams; digital twins for simulation and what-if analysis; neuromorphic and event-driven hardware for low-power sensing; privacy-enhancing technologies; graph-based models for connected infrastructure; and physics-informed machine learning. These developments are promising but should be evaluated against reproducibility, energy consumption, domain shift, security, and evidence of operational benefit. Novelty does not substitute for validation.

Chapter Summary

IoT analytics transforms sensor observations into knowledge and action through a complete pipeline of acquisition, preparation, inference, decision, actuation, and feedback. Descriptive analytics explains what happened, diagnostic analytics investigates why, predictive analytics estimates what may happen, and prescriptive analytics selects what should be done under objectives and constraints. Reliable analytics depends on data quality, leakage-free temporal validation, suitable metrics, uncertainty awareness, and domain knowledge.

Machine learning supports regression, classification, forecasting, representation learning, and anomaly detection, but model selection must consider latency, energy, interpretability, safety, and generalization. Edge AI and TinyML reduce latency and dependence on connectivity, while federated learning keeps raw data local but introduces communication, heterogeneity, privacy, and poisoning challenges. Intelligent decisions require calibrated predictions, cost-sensitive thresholds, feasible optimization, human oversight, and independent safety controls. Successful deployment also requires monitoring for drift, secure updates, governance, privacy protection, and continuous evaluation of operational outcomes.

Key Terms

Analytics: Systematic examination of data to produce information, insight, or action.

Descriptive analytics: Analysis that summarizes historical or current observations.

Diagnostic analytics: Analysis intended to identify causes or contributing factors.

Predictive analytics: Estimation of future or unknown outcomes.

Prescriptive analytics: Selection or recommendation of actions under objectives and constraints.

Feature engineering: Construction or transformation of variables used by an analytical model.

Data leakage: Use of information during training that would not be available at decision time.

Concept drift: Change over time in data distributions or input–outcome relationships.

Edge AI: Execution of AI inference near the data-generating device.

TinyML: Machine learning implemented on highly resource-constrained embedded hardware.

Federated learning: Collaborative model training in which raw client data remain local.

Anomaly: Observation or pattern that deviates from an accepted reference behavior.

Residual: Difference between an observed value and a model-based expected value.

Calibration: Agreement between predicted probabilities or measurements and their actual frequencies or reference values.

Precision: Fraction of predicted positive cases that are actually positive.

Recall: Fraction of actual positive cases detected by a classifier.

Model drift: Degradation or change in model behavior caused by changing data or conditions.

Digital twin: Computational representation of a physical asset or process linked to operational data.

Prescriptive optimization: Mathematical selection of actions that minimize cost or maximize utility while satisfying constraints.

Explainability: Ability to provide understandable reasons or evidence for a model output.

Review Questions

Short Answer Questions

1. Distinguish descriptive, diagnostic, predictive, and prescriptive analytics.

2. Why is chronological validation important for IoT time-series models?

3. State two advantages and two limitations of edge AI.

4. What is concept drift, and how can it affect an IoT model?

5. Why does federated learning not automatically guarantee privacy?

Long Answer Questions

1. Explain the complete IoT analytics value chain from sensing to feedback, including the role of data quality.

2. Compare statistical, distance-based, one-class, and autoencoder approaches to anomaly detection.

3. Discuss the design considerations for deploying a TinyML model on a battery-powered sensor.

4. Explain how uncertainty, asymmetric costs, and operational constraints influence prescriptive decision-making.

5. Discuss the ethical, cybersecurity, governance, and safety requirements for AI-enabled IoT systems.

Analytical / Application Questions

1. A classifier identifies 80 true positives, 20 false positives, and 10 false negatives. Calculate precision, recall, and F1 score, and explain which operational concern each metric represents.

2. Design a leakage-free evaluation strategy for a predictive-maintenance model deployed across ten machines operating under changing loads.

3. A smart-building controller predicts high electricity demand for the next hour. Formulate a prescriptive strategy that considers battery capacity, comfort limits, energy price, forecast uncertainty, and human override.

Multiple Choice Questions with Answers

Question 1

Which form of analytics recommends an action subject to costs and operational constraints?

A. Descriptive analytics

B. Diagnostic analytics

C. Predictive analytics

D. Prescriptive analytics

Correct Answer: D. Prescriptive analytics

Question 2

Which validation approach is most appropriate for forecasting future IoT measurements?

A. Randomly shuffle all observations before splitting

B. Chronological or rolling-origin validation

C. Use only the training error

D. Duplicate the most recent observations in every split

Correct Answer: B. Chronological or rolling-origin validation

Question 3

A classifier has 80 true positives and 20 false positives. Its precision is:

A. 0.20

B. 0.25

C. 0.80

D. 0.?

Correct Answer: C. 0.80

Question 4

What is a principal advantage of edge AI?

A. It guarantees perfect model accuracy

B. It can reduce latency and dependence on cloud connectivity

C. It eliminates the need for cybersecurity

D. It always provides unlimited memory

Correct Answer: B. It can reduce latency and dependence on cloud connectivity

Question 5

In federated averaging, client updates are commonly weighted according to:

A. The alphabetical order of clients

B. The color of the sensor enclosure

C. The number of local training samples

D. The network address only

Correct Answer: C. The number of local training samples

Question 6

Which metric penalizes large regression errors more strongly than MAE?

A. RMSE

B. Median absolute error only

C. Recall

D. Precision

Correct Answer: A. RMSE

Question 7

A residual is generally defined as:

A. The device serial number

B. Observed value minus model-expected value

C. The model memory size

D. The communication bandwidth

Correct Answer: B. Observed value minus model-expected value

Question 8

Why can random splitting be misleading for IoT time-series data?

A. It always reduces the number of features

B. It may place future information in the training set

C. It prevents all models from converging

D. It changes sensor units

Correct Answer: B. It may place future information in the training set

Question 9

Which technique commonly reduces TinyML memory and computation requirements?

A. Quantization

B. Increasing every layer width

C. Storing duplicate models

D. Increasing sensor sampling without limit

Correct Answer: A. Quantization

Question 10

What does concept drift describe?

A. A permanent loss of device identity

B. A change in the relationship between inputs and outcomes over time

C. A type of battery chemistry

D. A method for encrypting firmware

Correct Answer: B. A change in the relationship between inputs and outcomes over time