WorksheetsMeta-Learning, Few-Shot & Zero-Shot Learning & Graph
Total questions: 20
Worksheet time: 10mins
Meta-learning means
Learning faster models
Learning how to learn
Learning without data
Learning deeper networks
Few-shot learning refers to
Many labeled samples
Learning from very few examples
Learning without labels
Learning once
Zero-shot learning means
No training
No testing
Recognizing unseen classes
No model
Meta-learning operates across
Samples
Tasks
Classes
Features
A task in meta-learning contains
One sample
Support set and query set
Only test data
Only training data
Query set is used for
Learning representations
Updating meta-parameters
Evaluating task performance
Data augmentation
Prototypical Networks are
Gradient-based
Metric-based
Rule-based
Probabilistic
Prototype represents
Random sample
Mean of class embeddings
Maximum value
Weight vector
Distance metric commonly used is
Manhattan
Cosine
Euclidean
Hamming
Meta-training differs from normal training because
No loss
Trains across tasks
Uses test data
No labels
Zero-shot learning relies on
Attribute information
Random guessing
Large batch sizes
CNN depth
Few-shot learning is common in
NLP & Vision
Databases
Networking
Operating systems
One-shot learning means
One task
One example per class
One epoch
One model
Meta-learning improves
Memory usage
Generalization to new tasks
Training loss only
Batch size
Graph data consists of
Rows and columns
Nodes and edges
Pixels
Tokens
Graphs are useful for
Images only
Sequential data
Relational data
Tabular data
Directed graph means
No edges
Bidirectional edges
Edges have direction
No nodes
Undirected graphs have
One-way edges
Two-way edges
No attributes
No nodes
Graph Machine Learning learns from
Independent samples
Connected data
Images
Text
Examples of graph data include
Molecules
Road networks
Knowledge graphs
All of the above
