Home »
Trending Technologies MCQs
TinyML MCQs (Multiple-Choice Questions)
Practice TinyML MCQs to test your understanding of machine learning techniques that enable AI models to run on resource-constrained devices such as microcontrollers, sensors, and embedded systems. These multiple-choice questions cover model optimization, quantization, pruning, edge inference, hardware limitations, and low-power machine learning applications. They are useful for students, embedded systems developers, AI engineers, and candidates preparing for technical interviews. The collection includes foundational and practical questions to strengthen your understanding of TinyML.
TinyML MCQs
These TinyML multiple-choice questions cover important concepts such as embedded machine learning, microcontrollers, edge AI, TensorFlow Lite for Microcontrollers, model quantization, pruning, knowledge distillation, memory optimization, energy efficiency, and on-device inference. The questions explore how machine learning models operate within limited memory, processing power, and energy budgets through conceptual, technical, and scenario-based questions.
These TinyML MCQs help learners understand how to develop, optimize, deploy, and evaluate machine learning models on small embedded devices. Each question includes an answer and explanation.
List of TinyML MCQs
Explore the following 50 MCQs covering TinyML fundamentals, model compression, embedded hardware, development tools, deployment techniques, real-world applications, and practical implementation challenges.
1. What is TinyML?
- A technique used exclusively to train large language models in cloud data centers
- A programming language designed only for desktop applications
- An approach that enables machine learning models to run on resource-constrained devices
- A database system for storing large multimedia files
Answer: C) An approach that enables machine learning models to run on resource-constrained devices
Explanation:
TinyML focuses on deploying machine learning models on devices with limited memory, processing capacity, and power. Common targets include microcontrollers, embedded sensors, and small edge devices.
2. What is a primary goal of TinyML?
- To enable low-power machine learning inference directly on small devices
- To require every prediction to be processed by a cloud server
- To increase the memory requirements of embedded applications
- To eliminate all sensors from machine learning systems
Answer: A) To enable low-power machine learning inference directly on small devices
Explanation:
TinyML aims to bring machine learning capabilities to devices that have limited computational and energy resources. Local inference can reduce communication costs, improve response time, and support operation without a continuous internet connection.
3. Which device is commonly used for TinyML deployment?
- A large cloud computing cluster
- A high-capacity database server
- A desktop workstation with multiple GPUs
- A microcontroller-based development board
Answer: D) A microcontroller-based development board
Explanation:
Microcontroller boards are common TinyML platforms because they can operate with limited memory and power. Depending on the workload, developers may also use small embedded processors and other low-power edge devices.
4. What is edge inference in TinyML?
- Training every model in a remote data center
- Running a trained model locally on or near the device that collects the data
- Storing all model outputs in a cloud database before prediction
- Manually classifying every sensor reading
Answer: B) Running a trained model locally on or near the device that collects the data
Explanation:
Edge inference processes input data close to where it is generated. In TinyML, the model may run directly on a microcontroller, allowing the device to make predictions without sending every input to a remote server.
5. Which resource is typically constrained on a TinyML device?
- RAM and flash storage
- The number of internet websites available
- The number of programming languages in existence
- The size of a remote data center
Answer: A) RAM and flash storage
Explanation:
Many TinyML devices have limited RAM for intermediate computations and limited flash memory for model weights and application code. Model size, tensor allocation, and runtime overhead must be considered during deployment.
6. What is model quantization in TinyML?
- A process that increases every model weight to 64 bits
- A method for removing all input features from a model
- A technique that represents model values using lower-precision numerical formats
- A process that converts sensor readings into filenames
Answer: C) A technique that represents model values using lower-precision numerical formats
Explanation:
Quantization reduces the precision used to represent model weights and, in some cases, activations. For example, converting floating-point values to 8-bit integers can reduce storage requirements and enable efficient integer arithmetic on supported hardware.
7. What is a common benefit of 8-bit integer quantization?
- It guarantees that model accuracy will increase
- It can reduce model storage and support faster, lower-power inference on compatible hardware
- It eliminates the need for input data
- It makes the model independent of hardware capabilities
Answer: B) It can reduce model storage and support faster, lower-power inference on compatible hardware
Explanation:
Int8 quantization can reduce the memory footprint of weights compared with 32-bit floating-point representations. It may also improve inference speed and energy use, depending on the processor, operator support, and implementation.
8. What is model pruning?
- A technique for increasing the number of layers without retraining
- A method for recording sensor temperatures
- A process that converts every model weight into a text string
- A technique that removes or reduces the importance of selected model parameters or structures
Answer: D) A technique that removes or reduces the importance of selected model parameters or structures
Explanation:
Pruning removes or suppresses less useful parameters, connections, channels, or other structures. Depending on the pruning method and hardware, it can reduce model size or computation, although sparse weights do not automatically translate into faster inference on every device.
9. What is knowledge distillation in TinyML?
- Training a smaller student model to learn from a larger teacher model
- Removing all training data before model development
- Increasing the physical memory of a microcontroller
- Converting a machine learning model into a database table
Answer: A) Training a smaller student model to learn from a larger teacher model
Explanation:
Knowledge distillation transfers useful behavior from a larger teacher model to a smaller student model. The student can be designed for limited hardware while attempting to retain acceptable predictive performance.
10. Which framework is commonly used to deploy machine learning models on microcontrollers?
- Apache Hadoop
- TensorFlow Lite for Microcontrollers
- Microsoft Excel
- Apache Spark SQL alone
Answer: B) TensorFlow Lite for Microcontrollers
Explanation:
TensorFlow Lite for Microcontrollers is designed to run machine learning models on resource-constrained microcontrollers and similar devices. It provides a small inference runtime and supports selected operators suitable for embedded deployment.
11. What is the main purpose of a microcontroller in a TinyML application?
- To host a large distributed cloud database
- To replace every sensor with a remote server
- To execute embedded control logic and, when supported, local machine learning inference
- To provide unlimited RAM for every model
Answer: C) To execute embedded control logic and, when supported, local machine learning inference
Explanation:
A microcontroller integrates processing, memory, and peripheral interfaces for embedded applications. TinyML models can run on compatible microcontrollers to analyze sensor readings and generate predictions locally.
12. Why is RAM usage important in TinyML?
- RAM is used only to store remote cloud backups
- RAM has no effect on inference
- All model data must always fit into processor registers
- RAM stores temporary data, input tensors, intermediate activations, and other runtime information
Answer: D) RAM stores temporary data, input tensors, intermediate activations, and other runtime information
Explanation:
TinyML inference often requires RAM for input buffers, intermediate tensors, operator workspaces, and runtime structures. A model may fit in flash storage but still fail to run if its runtime memory requirements exceed available RAM.
13. What is flash memory commonly used for in a TinyML device?
- Storing firmware, model weights, and other persistent program data
- Performing all floating-point operations directly
- Replacing the processor's instruction execution unit
- Increasing the wireless network's bandwidth automatically
Answer: A) Storing firmware, model weights, and other persistent program data
Explanation:
Flash memory retains data without continuous power and commonly stores embedded firmware and model parameters. The exact allocation depends on the device architecture and deployment framework.
14. Which operation is commonly performed locally in a TinyML predictive maintenance system?
- Training a massive language model from scratch on every sensor reading
- Analyzing vibration signals to detect possible equipment abnormalities
- Storing all industrial data on a remote server before every calculation
- Replacing sensor measurements with random values
Answer: B) Analyzing vibration signals to detect possible equipment abnormalities
Explanation:
A TinyML device can analyze vibration or acoustic data to detect patterns associated with equipment problems. Local inference can support timely alerts while reducing the need to transmit every raw sensor sample.
15. What is the purpose of feature extraction in a TinyML workflow?
- To increase the number of unrelated sensor readings
- To eliminate the need for model inputs
- To derive informative representations from raw input data
- To store predictions in flash memory without processing
Answer: C) To derive informative representations from raw input data
Explanation:
Feature extraction transforms raw inputs into useful representations. For example, an audio application may calculate spectral features from a signal, while a vibration application may extract statistical or frequency-domain features before classification.
16. What is the main advantage of on-device inference in TinyML?
- It guarantees unlimited model capacity
- It eliminates all hardware constraints
- It requires every prediction to be sent to a cloud server
- It can reduce network dependence, response time, and the amount of raw data transmitted
Answer: D) It can reduce network dependence, response time, and the amount of raw data transmitted
Explanation:
On-device inference can process data locally without waiting for a remote server. This can improve responsiveness and privacy while reducing communication overhead, although the device still needs adequate compute, memory, and power.
17. What does quantization-aware training do?
- Simulates quantization effects during training to help a model adapt to lower-precision deployment
- Removes the need to train or validate a model
- Guarantees that every quantized model is more accurate than its original version
- Converts every input feature into a binary class label
Answer: A) Simulates quantization effects during training to help a model adapt to lower-precision deployment
Explanation:
Quantization-aware training introduces simulated quantization effects during training. This allows model parameters to adapt to numerical precision constraints and may preserve accuracy better than applying post-training quantization alone in some applications.
18. What is post-training quantization?
- A technique that requires training every model from scratch
- A process of converting a trained model to lower-precision representations after training
- A method for increasing the number of hidden layers automatically
- A process that removes all model parameters
Answer: B) A process of converting a trained model to lower-precision representations after training
Explanation:
Post-training quantization converts an already trained model to a lower-precision format. Depending on the method, representative calibration data may be used to estimate activation ranges and improve quantization quality.
19. Which type of model is often suitable for keyword spotting on a low-power microcontroller?
- A large transformer requiring several high-memory GPUs for every inference
- A model that accepts no audio input
- A compact neural network designed for small audio inputs and limited computation
- A cloud-only model that cannot run locally
Answer: C) A compact neural network designed for small audio inputs and limited computation
Explanation:
Keyword spotting identifies predefined words or commands in audio. Small convolutional or other lightweight neural networks can be designed to run on microcontrollers, provided their computation and memory requirements fit the target hardware.
20. What is the purpose of a tensor arena in TensorFlow Lite for Microcontrollers?
- To store all training datasets permanently
- To provide wireless communication between cloud servers
- To replace the device's processor
- To provide a memory region used for tensors and intermediate inference allocations
Answer: D) To provide a memory region used for tensors and intermediate inference allocations
Explanation:
TensorFlow Lite for Microcontrollers uses a tensor arena to manage memory needed for inference tensors and related allocations. Developers must allocate enough memory for the model's operators and runtime requirements on the target device.
21. Which sensor is commonly used in TinyML gesture recognition applications?
- An accelerometer
- A printer cartridge sensor used only for ink monitoring
- A database query counter
- A document page-number detector
Answer: A) An accelerometer
Explanation:
An accelerometer measures acceleration along one or more axes. TinyML models can analyze accelerometer readings to recognize gestures, activity patterns, or motion-related events.
22. Why is energy consumption an important consideration in TinyML?
- All TinyML devices operate with unlimited power
- Many devices run on batteries or must meet strict power budgets
- Energy use affects only the appearance of the device
- Machine learning inference never consumes electrical power
Answer: B) Many devices run on batteries or must meet strict power budgets
Explanation:
TinyML devices may need to operate for long periods without battery replacement. Efficient model execution, sensor sampling, processor selection, and sleep scheduling can help reduce overall energy consumption.
23. What is duty cycling in low-power embedded systems?
- Keeping every component active at full power continuously
- Increasing sensor sampling regardless of battery capacity
- Repeatedly enabling components when needed and placing them in low-power states when idle
- Removing all timers from the firmware
Answer: C) Repeatedly enabling components when needed and placing them in low-power states when idle
Explanation:
Duty cycling reduces energy consumption by limiting how long a device or component remains active. A TinyML system may collect sensor data, perform inference, transmit an alert if necessary, and then return to a low-power state.
24. Which statement best describes integer-only inference in TinyML?
- It requires every model operation to use 64-bit floating-point arithmetic
- It prevents the model from receiving numerical input
- It can only be used for sorting operations
- It uses integer arithmetic for supported inference operations to reduce dependence on floating-point computation
Answer: D) It uses integer arithmetic for supported inference operations to reduce dependence on floating-point computation
Explanation:
Integer-only inference can improve compatibility and efficiency on hardware designed for integer operations. Quantized models use scale and zero-point parameters to represent real values approximately in integer form, and the supported operators must match the deployment runtime.
25. What is the role of a representative dataset in full integer post-training quantization?
- To help estimate activation ranges during calibration
- To replace the trained model with a new dataset
- To guarantee zero prediction errors
- To determine the physical dimensions of the microcontroller
Answer: A) To help estimate activation ranges during calibration
Explanation:
Representative samples help calibration procedures estimate the ranges of activations in a trained model. These estimates support the conversion of floating-point computations into integer representations and can influence the accuracy of the quantized model.
26. What is the purpose of operator support in a TinyML inference framework?
- To ensure that every possible neural network operation runs on every microcontroller
- To determine which model operations the runtime and target hardware can execute
- To remove the need for model conversion
- To automatically increase available RAM
Answer: B) To determine which model operations the runtime and target hardware can execute
Explanation:
Embedded inference frameworks support a defined set of operators and implementations. A model containing unsupported operations may require conversion, replacement of certain layers, a different runtime, or additional implementation work.
27. What is model conversion in a TinyML workflow?
- A process that changes the physical shape of the microcontroller
- A technique for deleting all learned weights
- The transformation of a trained model into a format suitable for the target inference runtime
- A method for increasing the sampling frequency of every sensor
Answer: C) The transformation of a trained model into a format suitable for the target inference runtime
Explanation:
Model conversion transforms a trained model into a representation supported by the deployment framework. The process may include operator conversion and quantization, followed by packaging the model for use in embedded firmware.
28. Why should a TinyML model be tested on the actual target hardware?
- To ensure that its training dataset has the correct filename
- To guarantee that it performs identically on every processor
- To eliminate the need to check prediction accuracy
- To verify real memory usage, inference latency, operator compatibility, and power requirements
Answer: D) To verify real memory usage, inference latency, operator compatibility, and power requirements
Explanation:
Desktop tests cannot fully reproduce the constraints of a microcontroller. Testing on target hardware reveals actual runtime behavior, memory requirements, numerical differences, and compatibility issues that may not be apparent during model development.
29. Which metric is particularly important for a time-sensitive TinyML application?
- Inference latency
- The number of comments in the source code
- The size of the development team's email inbox
- The number of files in an unrelated directory
Answer: A) Inference latency
Explanation:
Inference latency measures the time required to produce a prediction. Applications such as wake-word detection, gesture recognition, and safety monitoring may need predictions within a specified response-time budget.
30. What is a common trade-off when reducing a TinyML model's size?
- The model automatically becomes more accurate
- Lower memory usage may come at the cost of some predictive accuracy
- The model no longer requires input data
- The device gains unlimited processing power
Answer: B) Lower memory usage may come at the cost of some predictive accuracy
Explanation:
Quantization, pruning, and architectural simplification can reduce model size, but they may also reduce accuracy. Developers should measure the trade-off and choose a model that meets both resource and task-performance requirements.
31. Which development environment is commonly associated with programming microcontroller-based TinyML devices?
- A web browser used only to display static pages
- A spreadsheet application used without embedded development tools
- Arduino IDE
- A PDF viewer
Answer: C) Arduino IDE
Explanation:
Arduino IDE is commonly used to develop and upload firmware to supported microcontroller boards. TinyML projects can integrate sensor code, embedded libraries, and an inference runtime into a firmware application.
32. What is an embedded inference engine?
- A cloud database for storing training labels
- A hardware component used only for battery charging
- A programming tool that generates images without computation
- A software runtime that executes a trained machine learning model on an embedded device
Answer: D) A software runtime that executes a trained machine learning model on an embedded device
Explanation:
An embedded inference engine loads or accesses a supported model and executes its operations on the device. TinyML runtimes are designed to work within constrained environments and may support only a selected set of operators.
33. How can TinyML improve privacy in sensor-based applications?
- By processing data locally so that less raw sensor information needs to be transmitted
- By automatically encrypting every device without configuration
- By storing all personal data in a publicly accessible server
- By eliminating all security risks from embedded systems
Answer: A) By processing data locally so that less raw sensor information needs to be transmitted
Explanation:
Local inference can reduce the amount of sensitive raw data sent to remote systems. However, privacy still depends on data handling, device security, access controls, retention policies, and the specific application design.
34. What is the purpose of sensor calibration in a TinyML application?
- To increase the number of neural network layers
- To improve the reliability of sensor measurements by accounting for systematic errors or offsets
- To eliminate all input measurements
- To guarantee that every machine learning prediction is correct
Answer: B) To improve the reliability of sensor measurements by accounting for systematic errors or offsets
Explanation:
Sensor calibration helps align measurements with expected physical values. Since TinyML models often depend on sensor input distributions, consistent and calibrated measurements can improve reliability when they match the model's training assumptions.
35. Why is input preprocessing important in TinyML deployment?
- It makes the input data independent of the trained model
- It eliminates the need to collect representative training data
- It ensures that device inputs are transformed into the format and representation expected by the model
- It automatically increases the available flash memory
Answer: C) It ensures that device inputs are transformed into the format and representation expected by the model
Explanation:
Input preprocessing may include scaling, normalization, windowing, or feature extraction. The embedded implementation must match the preprocessing used during training closely enough to avoid prediction errors caused by inconsistent input representations.
36. What is a common use of TinyML in smart agriculture?
- Running large-scale cloud simulations on every sensor
- Replacing all environmental measurements with random values
- Eliminating the need for agricultural monitoring
- Analyzing soil, temperature, humidity, or other sensor data to detect conditions that require attention
Answer: D) Analyzing soil, temperature, humidity, or other sensor data to detect conditions that require attention
Explanation:
TinyML can process sensor readings locally to identify conditions such as unusual temperature patterns or irrigation needs. Local predictions can reduce communication overhead in agricultural environments with limited connectivity.
37. What is the purpose of a sliding window in TinyML time-series processing?
- To divide a continuous stream into segments suitable for feature extraction or model inference
- To increase the flash memory capacity of a microcontroller
- To remove all timestamps from the input
- To force every sensor measurement to have the same value
Answer: A) To divide a continuous stream into segments suitable for feature extraction or model inference
Explanation:
A sliding window extracts a fixed-length segment from a time series, often advancing by a selected stride. This allows a model to analyze recent sensor measurements for tasks such as activity recognition or anomaly detection.
38. What is an important challenge when deploying TinyML models across different hardware platforms?
- Every platform supports identical memory sizes and operators
- Differences in processor capabilities, memory limits, supported operations, and toolchains can require platform-specific optimization
- Hardware has no effect on inference performance
- All models automatically run without conversion or testing
Answer: B) Differences in processor capabilities, memory limits, supported operations, and toolchains can require platform-specific optimization
Explanation:
Embedded devices differ in instruction sets, accelerators, memory architecture, and software support. A model that works on one board may need conversion, operator changes, or additional optimization before it can run on another platform.
39. What is anomaly detection in TinyML?
- A technique that forces every input into the same known class
- A method for measuring only the size of model files
- A process of identifying observations or patterns that differ significantly from expected behavior
- A procedure for removing all unusual sensor measurements before analysis
Answer: C) A process of identifying observations or patterns that differ significantly from expected behavior
Explanation:
Anomaly detection identifies unusual patterns in data, such as unexpected vibration or abnormal equipment behavior. TinyML can support local anomaly detection, provided the model and input processing fit the device's resource constraints.
40. Why should a TinyML model be evaluated with data collected under realistic operating conditions?
- To ensure that the model sees only training examples
- To eliminate the need for deployment testing
- To guarantee identical results across every environment
- To determine whether the model generalizes to actual sensor noise, environmental variation, and device conditions
Answer: D) To determine whether the model generalizes to actual sensor noise, environmental variation, and device conditions
Explanation:
Real-world sensor inputs can differ from laboratory training data because of noise, hardware variation, temperature, placement, or environmental changes. Testing under representative conditions helps reveal issues that might otherwise appear only after deployment.
41. What is the role of a hardware accelerator in TinyML?
- To execute supported machine learning operations more efficiently than a general-purpose processor in suitable workloads
- To replace the need for all model inputs
- To guarantee that every model fits into memory
- To eliminate the need for firmware development
Answer: A) To execute supported machine learning operations more efficiently than a general-purpose processor in suitable workloads
Explanation:
Hardware accelerators can speed up supported neural network operations or reduce energy consumption. Their benefits depend on operator compatibility, model structure, data movement, and the overhead of using the accelerator.
42. What does model benchmarking measure in a TinyML project?
- Only the number of files in the firmware folder
- Performance characteristics such as inference time, memory usage, energy consumption, and predictive quality
- The number of programming languages installed on a computer
- Only the size of the training dataset's filename
Answer: B) Performance characteristics such as inference time, memory usage, energy consumption, and predictive quality
Explanation:
Benchmarking measures whether a model meets the practical constraints of its target device. It should include relevant resource measurements and predictive evaluation, rather than relying only on model size or desktop inference speed.
43. Which technique can help reduce the computational cost of a TinyML neural network?
- Adding unnecessary layers to every model
- Increasing every activation tensor without limits
- Using a compact architecture with fewer parameters and suitable operations
- Duplicating all model weights without changing the architecture
Answer: C) Using a compact architecture with fewer parameters and suitable operations
Explanation:
Compact architectures are designed to reduce parameter counts and computational operations. Choosing suitable layers and input dimensions can help reduce memory and latency, while accuracy should be measured on representative data.
44. What is the purpose of a confidence threshold in a TinyML classification system?
- To increase the processor's clock speed automatically
- To remove all low-confidence inputs from the sensor
- To guarantee that every prediction is correct
- To determine whether a prediction score is high enough for the system to accept a classification or trigger a defined response
Answer: D) To determine whether a prediction score is high enough for the system to accept a classification or trigger a defined response
Explanation:
A confidence threshold can help a device decide whether to accept a prediction, ignore an uncertain result, or request further processing. The threshold should be selected using validation data and the consequences of false positives and false negatives.
45. Which statement about TinyML and cloud-based machine learning is correct?
- TinyML emphasizes local inference on constrained devices, while cloud-based machine learning can use remote computing resources
- TinyML always requires a continuous high-speed internet connection
- Cloud-based machine learning cannot train neural networks
- Both approaches have identical memory and power requirements
Answer: A) TinyML emphasizes local inference on constrained devices, while cloud-based machine learning can use remote computing resources
Explanation:
TinyML focuses on machine learning under strict device-level constraints. Cloud systems can provide greater computing capacity for training or inference, while hybrid architectures can combine local predictions with cloud-based analysis when connectivity and application requirements permit.
46. Why might a TinyML developer use fixed-point arithmetic?
- To make all calculations independent of numerical precision
- To perform supported numerical computations using representations that can be more suitable for constrained processors
- To remove the need for mathematical operations
- To ensure that every model uses unlimited memory
Answer: B) To perform supported numerical computations using representations that can be more suitable for constrained processors
Explanation:
Fixed-point arithmetic represents numerical values using scaled integers rather than general floating-point values. It can be useful on embedded hardware with limited or less efficient floating-point support, but numerical range and precision must be managed carefully.
47. What is an important security consideration for a deployed TinyML device?
- Security is unnecessary when inference occurs locally
- Every firmware image should be publicly writable
- Firmware integrity, secure updates, and protection against unauthorized access should be considered
- All model outputs should automatically be sent to an unsecured server
Answer: C) Firmware integrity, secure updates, and protection against unauthorized access should be considered
Explanation:
Local inference does not eliminate security risks. Embedded deployments should consider secure boot where supported, signed firmware updates, access controls, protection of sensitive data, and safe handling of model inputs and outputs.
48. A developer has trained a sound-classification model, but the model file exceeds the flash memory available on a microcontroller. Which approach is most appropriate to try?
- Increase the number of model layers and keep all parameters unchanged
- Store the entire model in RAM without checking memory limits
- Ignore the flash memory constraint and deploy the model as it is
- Apply suitable quantization, evaluate pruning or distillation, and consider a smaller architecture before measuring the converted model size
Answer: D) Apply suitable quantization, evaluate pruning or distillation, and consider a smaller architecture before measuring the converted model size
Explanation:
Quantization can reduce parameter storage, while pruning, distillation, or a compact architecture may further reduce model size. The developer should measure the converted model and firmware footprint, verify operator compatibility, and confirm that accuracy remains acceptable.
49. A battery-powered TinyML device detects machine vibration anomalies. Its predictions are accurate, but battery life is much shorter than expected. What should the developer investigate first?
- Measure the full system's power consumption, including sensor sampling, processor activity, inference frequency, wireless communication, and sleep behavior
- Increase the number of inference operations without measuring power
- Disable all model evaluation permanently
- Increase the size of the model to improve battery life
Answer: A) Measure the full system's power consumption, including sensor sampling, processor activity, inference frequency, wireless communication, and sleep behavior
Explanation:
Battery consumption depends on the complete system, not just the model's inference cost. The developer should profile active and idle power, sensor duty cycles, communication, and inference frequency, then optimize the components responsible for the highest energy use.
50. An industrial monitoring system must detect abnormal vibration on a microcontroller with limited RAM and flash memory. It must operate without continuous internet access and issue alerts with low latency. Which design is most appropriate?
- Send every raw vibration sample to a remote cloud model before generating an alert
- Deploy a compact, validated model with suitable quantization, optimize preprocessing and tensor memory, and perform inference locally on the microcontroller
- Use the largest available neural network without measuring resource usage
- Store all sensor readings indefinitely in RAM and postpone every prediction until the device reconnects to the internet
Answer: B) Deploy a compact, validated model with suitable quantization, optimize preprocessing and tensor memory, and perform inference locally on the microcontroller
Explanation:
This design supports low-latency inference without continuous connectivity. Quantization and compact model architecture can reduce memory requirements, while optimized preprocessing and tensor allocation help the model fit available RAM. The final system should be tested on the actual hardware for accuracy, latency, energy consumption, and reliable alert behavior under realistic operating conditions.