WorksheetsFinal Test
Total questions: 50
Worksheet time: 1hrs 18mins
In the context of Unsupervised Learning, what is the primary purpose of "Dimensionality Reduction"?
To increase the number of features to capture more details.
To label the dataset manually for future training.
To reduce the number of features while retaining essential information.
To classify data into distinct groups based on distance.
Which of the following scenarios indicates that a model is suffering from "High Variance" (Overfitting)?
The model has high error on the training set and high error on the test set.
The model has low error on the training set and low error on the test set.
The model has extremely low error on the training set but high error on the test set.
The model is too simple to capture the underlying trend of the data.
What is the mathematical objective of the "Cost Function" during the training phase?
To minimize the difference between predicted outputs and actual targets.
To maximize the number of features used.
To calculated the accuracy percentage directly.
To increase the complexity of the model to fit all points.
In a Confusion Matrix, what does a "False Positive" represent?
The model correctly predicted the positive class.
The model incorrectly predicted the negative class when it was actually positive.
The model incorrectly predicted the positive class when it was actually negative.
The model correctly predicted the negative class.
In Reinforcement Learning, what is a "Policy"?
The signal indicating how good an action was.
The environment in which the agent operates.
The strategy or rule that defines the agent's behavior at a given state.
The final goal the agent is trying to achieve.
If you remove the "Test Set" and evaluate your model only on the "Training Set", what is the most likely outcome?
The model will underfit.
The evaluation will overestimate the model's true performance.
The model will perform better on real-world data.
The training time will increase.
What defines the "State" in a Reinforcement Learning scenario?
The action taken by the agent.
The configuration of the environment.
The long-term goal of the agent.
The reward received after an action.
Which algorithm is considered a "Lazy Learner" because it does not build a model during the training phase but memorizes the training dataset instead?
Linear Regression.
Decision Tree.
K-Nearest Neighbors.
Artificial Neural Network.
What is the risk of using "Mean Imputation" to handle missing data?
It reduces the size of the dataset significantly.
It can be sensitive to outliers and distort the data distribution.
It cannot be used on numerical data.
It makes the model too complex.
Which statement correctly highlights a similarity between biological neural networks and artificial neural networks?
Both use digital signals to communicate between nodes.
Both consist of interconnected units (neurons) that process information.
Both are entirely software-based systems.
Both have perfect memory and unlimited processing speed.
Which of the following lists the main components of an artificial neuron?
Input signals, weights, summation function, activation function, output signal
Axons, dendrites, synapses, neurotransmitters
Memory, storage, CPU, output
Sensors, actuators, controller
What is the main purpose of the bias in an artificial neuron?
To adjust the output independently of the input, allowing the activation function to shift
To store previous outputs for later use
To increase the number of inputs a neuron can handle
To replace the activation function when it is not needed
Which of the following correctly matches an activation function with one of its characteristics?
Sigmoid – outputs values between -1 and 1
ReLU – allows for faster training and avoids vanishing gradient for positive inputs
Tanh – always outputs positive values only
Linear – introduces non-linearity to the network
Which activation function is most commonly used in the output layer for binary classification problems?
ReLU
Sigmoid
Tanh
Softmax
Which activation function is commonly used in the output layer for multi-class classification problems?
ReLU
Sigmoid
Tanh
Softmax
How do hidden layers in a neural network affect the decision boundary of a model?
They make the decision boundary always linear, regardless of the problem.
They allow the network to create complex, non-linear decision boundaries.
They prevent the network from learning any useful patterns.
They only affect the output scale, not the boundary shape.
Why are Convolutional Neural Networks (CNNs) often preferred over common Artificial Neural Networks (ANNs) for tasks like image recognition?
CNNs require no training, while ANNs do
CNNs automatically learn spatial features and reducing the number of parameters
CNNs can only work with small datasets, whereas ANNs cannot
CNNs always produce perfect accuracy on image tasks.
How does a Convolutional Neural Network (CNN) learn to detect the correct features in images?
The features are manually programmed by the designer before training
Filters are randomly initialized and adjusted during training using backpropagation to detect patterns that reduce the loss
CNNs use a fixed set of filters that never change
Features are learned only after training is complete without any weight updates
What does precision measure in a classification model?
The proportion of correctly predicted positive instances out of all predicted positive instances
The proportion of correctly predicted positive instances out of all actual positive instances
The overall proportion of correctly predicted instances
The difference between predicted and actual values
Which image processing technique involves adjusting an image to make it more suitable for human vision, such as by increasing contrast or brightness?
Image Restoration
Image Compression
Image Enhancement
Image Segmentation
Which layer in a CNN is responsible for downsampling the feature map, thereby reducing computational cost and preventing overfitting?
Convolutional Layer
Pooling Layer
Fully Connected Layer
ReLU Layer
What happens during the 'Flattening' step in a CNN architecture?
The image is converted to grayscale.
The 2D feature matrix is unrolled into a long 1D vector.
The resolution of the image is increased.
Negative values are removed from the matrix.
Which key innovation allows ResNet (Residual Network) to train extremely deep networks (e.g., 150+ layers) without suffering from the vanishing gradient problem?
Inception Modules
Skip Connections (Shortcut paths)
Depth-wise Separable Convolutions
Region Proposal Networks
MobileNet is designed for efficiency on mobile devices. What specific type of convolution does it use to reduce computation?
Standard Convolution
Depth-wise Separable Convolution
Atrous (Dilated) Convolution
Transposed Convolution
In the context of Object Detection, what distinguishes 'One-Stage Detectors' (like YOLO) from 'Two-Stage Detectors' (like Faster R-CNN)?
One-stage detectors are slower but more accurate.
One-stage detectors combine localization and classification into a single pass, making them faster.
One-stage detectors use Region Proposal Networks (RPN).
One-stage detectors cannot detect multiple objects.
What is the purpose of Intersection over Union (IoU) in object detection?
To convert the image to grayscale.
To measure the accuracy of a predicted bounding box against the ground truth.
To combine multiple feature maps into one.
To detect edges in an image.
What is the function of Non-Max Suppression (NMS) in the YOLO object detection pipeline?
To generate the initial region proposals.
To increase the brightness of dark objects.
To eliminate redundant bounding boxes and keep only the winner.
To separate the foreground from the background.
What is the main difference between Semantic Segmentation and Instance Segmentation?
Semantic segmentation draws bounding boxes; Instance does not.
Instance segmentation distinguishes between individual objects of the same class while Semantic does not.
Semantic segmentation is used for videos; Instance is for images.
There is no difference; they are the same.
What is the purpose of "Data Augmentation" in training an AI model for defect detection?
To artificially increase the diversity of the training dataset (e.g., by rotating or flipping images).
To reduce the size of the images to save storage.
To manually label every defect in the dataset.
To remove noise from the images.
What is "Edge Computing" in the context of AI Machine Vision?
Processing data exclusively in a centralized cloud server.
Processing data locally on the device (camera/sensor) or nearby server.
Using only the edges of an image for analysis.
Computing the cost of the entire system.
What is "Transfer Learning"?
Moving a robot from one factory to another.
Using a pre-trained model as a starting point for a new, similar task.
Transferring data from a hard drive to the cloud.
Teaching a human operator how to use the AI system.
In an AI-based visual inspection system, what is the output of the "Inference" stage?
A trained model file.
A prediction or decision based on new input data.
A labeled dataset.
A new training algorithm.
You are using Transfer Learning to train a defect detector. You decide to "freeze" the early layers (backbone) of the pre-trained model. Why?
To allow the model to learn new shapes from scratch.
Because early layers detect generic features (edges, curves) that are transferable, while later layers need retraining for specific defect types.
To increase the training time and complexity.
Because the pre-trained model was trained on the same dataset.
In the context of Deep Learning, how does the number of parameters generally affect a model's capabilities?
Fewer parameters make the model more creative.
More parameters correlate with higher intelligence.
Parameter count has no effect on performance.
More parameters simply increase the file storage size without improving intelligence.
Once text is tokenized, how does the model represent the meaning of words mathematically?
As alphabetical lists features.
As vectors in a high-dimensional space.
As simple binary code (0s and 1s).
As percentages.
When a Transformer predicts the next word, it generates raw scores for every word in its vocabulary. What are these raw scores called?
Vectors
Embeddings
Logits
Weights
Why are "Hallucinations" a common issue in Generative AI?
The model becomes too hot physically.
The model relies on insufficient data, noisy data, or lacks constraints.
The model refuses to answer based on ethical guidelines.
The embedding dimensions are too small.
In Transformer Model Why is "Positional Embedding" added to the input data?
To tell the model which language is being spoken.
To preserve information about the order of words so the model understands context.
To encrypt the user's data.
To increase the speed of processing.
What is the purpose of the "Attention Mechanism" in Transformer architecture?
To make the model pay attention to the user's voice.
To assign weights to words relative to other words to understand relationships.
To alert the user when an error occurs.
To focus solely on the last word of a sentence.
"Zero Shot" and "Few Shots" are techniques associated with which method?
Fine-Tuning.
Prompt Engineering.
RAG.
Tokenization.
In the context of AI Agents, what is the specific function of "Planning"?
To store long-term memories in a vector database.
To decompose complex goals into smaller sub-goals and perform self-reflection.
To connect to external APIs like Google Search.
To generate the final text response to the user.
Which of the following best describes the "Pre-trained" aspect of a GPT model?
The model learns in real-time as you speak to it.
The model was trained on a large dataset before being deployed.
The model requires the user to re-train it manually before use.
The model is hard-coded with fixed dictionary definitions.
How do AI Agents overcome the "Goldfish memory" limitation of standard LLMs?
By using a dual-memory system: Context and Vector Database.
By writing everything down in a physical notebook.
By strictly limiting the conversation length.
By increasing the learning rate and epoch
Which of the following is NOT a step in the "Action Loop" of an AI Agent?
Plan
Execute
Ignore Errors
Observe & Debug
How does the Gemini Sidebar differ from a standard chat in terms of context awareness?
It cannot see any data.
It "sees" your variables, dataframes, and runtime state.
It only knows about Python libraries, not your code.
It requires you to copy-paste everything manually.
What is the primary ethical issue described as the "Black Box" problem in predictive maintenance?
The AI system is too expensive to maintain.
The AI model is stolen by a competitor using reverse engineering.
The deep learning model cannot explain the rationale behind its decision.
The AI system relies on data that has been poisoned by a malicious insider.
A factory uses computer vision to ensure workers wear helmets (Safety). However, management begins using the same camera to track how long workers spend in the restroom. What ethical risk does this represent?
Function Creep
Algorithm Aversion
Model Drift
Data Poisoning
Unlike a chatbot error, what is the specific high-stakes consequence of an AI failure in an industrial manufacturing environment?
Social media backlash
Physical harm to workers
Minor data loss
Slower internet speeds
Which privacy technique involves adding a mathematical "noise" pattern to a dataset to prove ownership if it is stolen?
Differential Privacy
Watermarking
Data Hashing
Encryption
In the context of cybersecurity, what is a "Dataset Poisoning" attack?
Deleting the entire database.
Stealing the model using reverse engineering.
Corrupting the training data to teach the AI incorrect rule.
Sending spam emails to the system administrator.
