Lecture 2

Question 1

Deep Learning problem

Answer 1

Large training set
Expensive to make

Question 2

Few shot learning

Answer 2

N-way k-shot
N classes, k examples per class

Question 3

Few shot learning (train/val/test)

Answer 3

Make sets on class level (Few shots for new task)

Question 4

Few shot learning (Total)

Answer 4

Learn a learner that can learn in few shots, not an algorithm that can classify well

Question 5

Pre-training, Fine-tuning

Answer 5

Copy weights from previous task (expensively learned)
Fine tune with fewer shots

Question 6

Deep meta-learning

Answer 6

Learn good hyperparameter -> learn a good learner

Question 7

Hyperparameters of NN

Answer 7

Initialization parameters
Optimization Algorithm (SGD)
Loss function (MSE, cross-entropy)
Architecture (num and kind of layers)

Question 8

Matching network

Answer 8

Compute embeddings of support and query using G(x) and F(x)
Compare embeddings (using cosine similarity)
(not sure why meta learning yet)

Question 9

Prototyical network

Answer 9

Matching network but with class centroids
(uses euclid distance instead of cosine sim)

Question 10

Model Agnostic Meta Learning

Answer 10

Ff kijken hoe dit werkt

Question 11

3 approaches

Answer 11

Metric-based
Optimization-based
Model-based