Self-assessment quiz
Presentation & objectives
The following quizzes are here to help you check that you understood the articles you had to study. At the end of a quiz, you will be given explanations on your answers. If some of them are wrong, you will have the possibility to click on the question you failed to try again.
These quizzes are provided for self-assessment and will not be graded or stored.
Don’t hesitate to reach out on the Discord server for any precision/explanation!
Quizzes
What is artificial intelligence?
# What is Artificial Intelligence (AI)?
- [x] A family of techniques aimed at extracting knowledge from data to solve automation tasks
> ✅ AI involves methods for analyzing data to create systems that can automate complex tasks.
- [ ] A specific type of algorithm designed for image processing
> ❌ AI is a broad field encompassing more than image processing algorithms.
- [x] A concept that encompasses machine learning (ML) and deep learning (DL)
> ✅ ML and DL are subsets of AI, focusing on learning from data and solving specific tasks.
- [ ] A single mathematical model used for decision-making
> ❌ AI includes many models and techniques, not just one.
- [ ] A deterministic system that does not rely on probabilities
> ❌ AI often includes probabilistic systems, especially in machine learning and generative models.
# What distinguishes Machine Learning (ML) from traditional algorithms?
- [x] ML learns from data instead of being explicitly programmed
> ✅ Machine learning systems learn patterns from data rather than relying solely on pre-defined rules.
- [ ] ML always guarantees the correct solution to a problem
> ❌ Machine learning does not always guarantee correctness, as it depends on the data and model used.
- [x] ML involves a training phase to estimate parameters
> ✅ Machine learning models are trained using datasets to optimize parameters for better predictions.
- [ ] ML requires deterministic algorithms
> ❌ ML often involves probabilistic models and algorithms.
- [ ] ML only works with structured data
> ❌ ML can handle both structured and unstructured data, such as images, audio, and text.
# What is supervised learning?
- [x] A type of ML where the model learns from labeled data
> ✅ Supervised learning involves training models on datasets where both inputs and corresponding outputs are known.
- [ ] A method to group similar data points without labels
> ❌ This describes unsupervised learning, not supervised learning.
- [x] A task where the model predicts outputs based on examples
> ✅ Supervised learning models are trained to predict outputs using labeled input-output pairs.
- [ ] A process where the model receives rewards for actions
> ❌ This describes reinforcement learning, not supervised learning.
- [ ] A method exclusive to decision trees
> ❌ Supervised learning is not restricted to decision trees; it includes various models like linear regression and neural networks.
# What is a feature in machine learning?
- [x] An individual measurable property of the data
> ✅ Features are the attributes or variables used as input to machine learning models.
- [ ] A type of deep learning algorithm
> ❌ Features are not algorithms; they are data attributes.
- [x] A representation of data used for training a model
> ✅ Features are the representation of input data that models use for learning.
- [ ] A matrix of output predictions
> ❌ Features are input properties, not outputs.
- [ ] A hyperparameter in neural networks
> ❌ Hyperparameters control the model's training process, while features represent input data.
# Why is dimensionality reduction important in machine learning?
- [x] It simplifies the data by reducing the number of features
> ✅ Dimensionality reduction removes irrelevant or redundant features, making the data easier to process.
- [ ] It guarantees better model performance
> ❌ Dimensionality reduction improves efficiency but does not always guarantee better performance.
- [x] It helps in visualizing high-dimensional data
> ✅ Reducing dimensions enables better visualization of data in 2D or 3D space.
- [ ] It eliminates the need for feature selection
> ❌ Dimensionality reduction complements feature selection but does not replace it entirely.
- [ ] It automatically trains the model
> ❌ Dimensionality reduction is a preprocessing step, not a model training process.
Applications of machine learning
# What are examples of object detection?
- [x] Detecting and localizing objects in an image with bounding boxes
> ✅ Object detection identifies and localizes instances of objects within an image, often using bounding boxes and probabilities.
- [ ] Translating text in an image
> ❌ Text translation involves natural language processing, not object detection.
- [x] Using YOLO or DETR models to identify objects in an image
> ✅ YOLO and DETR are popular models for object detection tasks.
- [ ] Segmenting a tumor in medical imaging
> ❌ This is an example of semantic segmentation, not object detection.
- [ ] Predicting the age of an individual from an image
> ❌ This is a classification task, not object detection.
# What are common applications of large language models (LLMs)?
- [x] Translation, summarization, and question answering
> ✅ LLMs excel in natural language understanding tasks such as translation, text summarization, and answering questions.
- [ ] Analyzing audio for sound events
> ❌ Audio analysis falls under signal processing, not LLM tasks.
- [x] Conversational AI and role-playing
> ✅ LLMs are widely used in chatbots and systems designed for interactive conversations.
- [ ] Classifying images based on their content
> ❌ Image classification is a task in computer vision, not NLP.
- [x] Software coding assistance (e.g., GitHub Copilot)
> ✅ LLMs like GitHub Copilot are used to assist in software coding by generating relevant code snippets.
# Which application involves time series forecasting?
- [x] Predicting future energy demands
> ✅ Time series forecasting predicts future values, such as energy demands, based on historical data.
- [ ] Segmenting brain signals in EEG recordings
> ❌ Segmentation involves dividing data into distinct parts, not forecasting future trends.
- [x] Weather prediction
> ✅ AI is commonly used to forecast weather based on time series data.
- [ ] Recognizing bird species from their sounds
> ❌ Recognizing bird species involves audio classification, not time series forecasting.
- [ ] Generating music from an input melody
> ❌ Music generation does not rely on time series forecasting techniques.
# Which tasks are addressed by audio-based AI?
- [x] Sound event detection
> ✅ Sound event detection identifies the start and end times of specific sounds in audio recordings.
- [x] Marine mammal identification
> ✅ AI can identify marine mammal species from underwater acoustic recordings.
- [ ] Translating spoken words between languages
> ❌ Translation falls under NLP tasks, not audio analysis.
- [x] Musical source separation
> ✅ AI can separate the instrumental and vocal tracks in a song.
- [ ] Classifying tumors in medical imaging
> ❌ This task is part of computer vision, not audio processing.
Standard algorithms for machine learning
# What are the main differences between supervised and unsupervised learning?
- [x] Supervised learning uses labeled data, while unsupervised learning uses unlabeled data
> ✅ Supervised learning requires labeled datasets for training, whereas unsupervised learning identifies patterns without labels.
- [ ] Supervised learning is faster than unsupervised learning
> ❌ The speed depends on the algorithm and the data, not necessarily on whether it's supervised or unsupervised.
- [x] Unsupervised learning groups data into clusters
> ✅ Unsupervised learning often involves clustering data into meaningful groups.
- [ ] Supervised learning does not require a loss function
> ❌ Supervised learning optimizes a loss function to minimize prediction errors.
- [ ] Unsupervised learning uses a training set for prediction
> ❌ Unsupervised learning identifies patterns and structures without making predictions.
# What is the purpose of the k-nearest neighbors (k-NN) algorithm?
- [x] To predict labels based on the closest training samples
> ✅ k-NN predicts the label of a query point by examining its k nearest neighbors in the training data.
- [ ] To partition data into clusters
> ❌ Partitioning data into clusters is a task for algorithms like k-means, not k-NN.
- [x] To classify new data points using majority voting among neighbors
> ✅ k-NN uses the majority class among the nearest neighbors to classify a query point.
- [ ] To train a model by optimizing parameters
> ❌ k-NN does not involve training; it relies on storing and comparing with the training dataset.
- [ ] To estimate centroids in feature space
> ❌ Estimating centroids is a feature of k-means, not k-NN.
# Which distance metrics are commonly used in k-NN and k-means?
- [x] Euclidean distance
> ✅ Euclidean distance is the most commonly used metric for both k-NN and k-means.
- [x] Manhattan distance
> ✅ Manhattan distance is another valid metric, particularly for grid-like data structures.
- [ ] Cosine similarity
> ❌ Cosine similarity measures angles, not distances, and is less common in k-NN and k-means.
- [ ] Jaccard similarity
> ❌ Jaccard similarity is used for set-based comparisons, not for k-NN or k-means.
- [ ] Hamming distance
> ❌ Hamming distance is used for categorical data, not typical for k-NN or k-means clustering.
# What is the complexity of predicting a query in k-NN?
- [x] O(Nd), where N is the number of training samples and d is the feature dimension
> ✅ Prediction involves calculating distances from the query to all training points, making it O(Nd).
- [ ] O(log N)
> ❌ Logarithmic complexity is not applicable to k-NN, which requires scanning all points.
- [x] Computationally expensive for large datasets
> ✅ Since prediction requires comparing with all training points, k-NN is computationally expensive for large datasets.
- [ ] Constant time for all datasets
> ❌ Prediction time grows with the size of the training dataset and the dimensionality of features.
- [ ] Independent of the feature dimension
> ❌ Feature dimension directly impacts the computational cost of distance calculations.
A broad introduction to deep learning
# What is the primary focus of deep learning?
- [x] Learning directly from raw data with minimal preprocessing
> ✅ Deep learning methods learn features directly from raw data, reducing the need for manual feature extraction.
- [ ] Using only pre-extracted features for training models
> ❌ Deep learning performs end-to-end learning, including feature extraction.
- [x] Training models with automatic differentiation and backpropagation
> ✅ Deep learning relies on automatic differentiation and backpropagation to optimize model parameters.
- [ ] Working exclusively with labeled data
> ❌ Deep learning can work with both labeled and unlabeled data, especially in self-supervised learning.
- [ ] A manual process of optimizing each layer
> ❌ Deep learning optimizes all layers simultaneously during training.
# What is a fully connected layer in a neural network?
- [x] A layer where each node is connected to all nodes in the previous layer
> ✅ Fully connected layers compute a weighted sum of all inputs, followed by an activation function.
- [ ] A layer that operates only on image data
> ❌ Fully connected layers are general and not limited to image data.
- [ ] A layer with shared weights
> ❌ Shared weights are characteristic of convolutional layers, not fully connected layers.
- [x] Used in multi-layer perceptrons (MLPs)
> ✅ Fully connected layers form the basis of MLPs, often used for structured data tasks.
- [ ] The final output layer of every neural network
> ❌ Not all networks end with fully connected layers; some tasks use specialized layers like softmax.
# What are the main functions of convolutional layers?
- [x] To capture spatial patterns in data like images
> ✅ Convolutional layers extract spatial features using small filters, making them suitable for image and audio data.
- [ ] To connect all neurons in the previous layer
> ❌ This describes fully connected layers, not convolutional layers.
- [x] To share weights across different regions of the input
> ✅ Convolutional layers use shared weights (kernels) to efficiently process data with regular structure.
- [ ] To process unstructured text data
> ❌ Convolutional layers are typically used for structured data like images, not unstructured text.
- [ ] To reduce the size of the input
> ❌ Reducing input size is typically the role of pooling layers, not convolutional layers.
# What is automatic differentiation in deep learning?
- [x] Computing gradients automatically for all layers
> ✅ Automatic differentiation calculates gradients efficiently, enabling the training of deep networks using backpropagation.
- [ ] Optimizing hyperparameters
> ❌ Hyperparameter optimization is separate from automatic differentiation, which focuses on computing gradients.
- [x] A key step in backpropagation
> ✅ Automatic differentiation is essential for backpropagation, which updates weights based on gradient descent.
- [ ] A method for creating new neural architectures
> ❌ Automatic differentiation is not for designing architectures; it facilitates efficient training.
- [ ] Replacing all manual computations in neural networks
> ❌ While it automates gradient computations, other manual decisions (e.g., architecture design) remain essential.