NEW
Font size
WorksheetsPixel Pursuit
Total questions: 12
Worksheet time: 6mins
In a Vision Transformer (ViT), what is the purpose of the learnable [class] token that is prepended to the sequence of image patch embeddings?
To provide a global representation of the entire image, which is then fed to the final classification head.
To mark the beginning of the sequence, similar to a start-of-sentence token in NLP.
To store the positional embedding information for all the patches.
To reduce the dimensionality of the patch embeddings before they enter the transformer encoder.
You are designing a generative model for faces. A key issue is that changing a single attribute in the latent space (e.g., adding glasses) also unintentionally changes other attributes (e.g., hair color). This problem is known as:
Mode Collapse
Poor Feature Disentanglement
The Vanishing Gradient Problem
Catastrophic Forgetting
What is a primary advantage of using Depthwise Separable Convolutions, as seen in architectures like MobileNet, compared to standard convolutions?
They significantly reduce the number of computations and parameters while maintaining similar performance.
They increase the model's accuracy by a large margin on all tasks.
They are immune to the vanishing gradient problem.
They eliminate the need for activation functions.
A U-Net architecture is highly effective for biomedical image segmentation. What is the main function of the "skip connections" between the contracting path (encoder) and the expansive path (decoder)?
To add regularization to prevent overfitting.
To speed up the forward pass of the network.
To help the decoder recover fine-grained spatial information that was lost during down-sampling.
To reduce the number of channels in the feature maps.
The Fréchet Inception Distance (FID) is a common metric for evaluating the quality of images generated by GANs. What does it primarily measure?
Pixel-wise Mean Squared Error between real and generated images.
Structural similarity between real and generated images.
Similarity of feature distributions (mean and covariance) between real and generated images, as captured by a pre-trained Inception network.
The peak signal-to-noise ratio of the generated images.
In object detection, what is the main purpose of Non-Maximum Suppression (NMS)?
To generate initial bounding box proposals.
To eliminate redundant, overlapping bounding boxes for the same object, keeping only the one with the highest confidence score.
To normalize the coordinates of the bounding boxes.
To increase the number of detected objects.
Which of the following statements about ensemble methods is most accurate?
Boosting methods primarily aim to reduce the variance of a model.
Bagging methods, like Random Forest, build models sequentially, where each new model corrects the errors of the previous one.
Stacking involves training a meta-model to make final predictions based on the outputs of several base-level models.
Ensemble methods are only effective when all the base models are of the same type (e.g., all decision trees).
In Reinforcement Learning, what is the key difference between a model-based and a model-free approach?
Model-based methods learn a policy directly, while model-free methods first learn the value function.
Model-based methods attempt to learn a model of the environment (transition probabilities and rewards), while model-free methods do not.
Model-based methods are always more sample-efficient than model-free methods.
Model-free methods can only be used in environments with a discrete action space.
What is the "catastrophic forgetting" problem in the context of continual or lifelong learning?
When a model completely fails to learn a new task.
When a model's performance on a previously learned task degrades significantly after it is trained on a new task.
When a model overfits to a sequence of tasks.
When the model's loss function diverges during training.
You have a graph with 10⁵ nodes and 2×10⁵ edges. You need to find all strongly connected components (SCCs) efficiently. Which algorithm is most appropriate?
BFS on each node
Kosaraju’s or Tarjan’s algorithm (O(V+E))
Dijkstra’s algorithm from each node
Floyd-Warshall algorithm
You train a classifier on a highly imbalanced dataset. Accuracy is 95%, but the model never predicts the minority class. Which metric is more appropriate than accuracy?
R² score
Precision, recall, F1-score
Mean Squared Error
ROC-AUC only
You are implementing a real-time recommendation system where new user ratings arrive continuously. You need to maintain the top-k items for each user efficiently. Which data structure and approach is most suitable?
A simple list sorted after each rating; recompute top-k every time.
A min-heap of size k per user, updating in O(log k) per new rating.
A hash map storing all items without ordering; retrieve top-k by scanning the entire map.
A queue storing the last k ratings; the newest items are always the top-k.
