Paper 6 DeepMetaLearning Flashcards
1
Q
Model based methods
A
SNAIL
Meta nets
2
Q
Metric based methods
A
GNN
Relational nets
Matching nets
3
Q
Optimization based
A
MAML
Reptile
LSTM
(Everything else)
4
Q
N-way, k-shot Learning
A
N classes, k examples per class (for each tasks)
Train, val, test split is done on task level
The model trains on Tasks and can more quickly learn new Task
5
Q
Few shot learning
A
N-way, k-shot learning with a small k
6
Q
Matching network
A
Metric based
Compare pair wise cosine similarity of query to all support samples
7
Q
Prototypical network
A
Metric based
Extends matching network by using class centroids instead of samples for support classes
8
Q
MAML
A
Optimization based
Learn inital parameters of network to learn new tasks from few data
9
Q
A
