Deep Learning is really good when dealing with images where conventional machine learning methodologies fell short. Still when training a deep neural network we need a lot of labeled examples unlike a human which can learn an object from even a single image. Collecting labeled images is not only cumbersome but also expensive. Training a classifier with few examples will simply overfit on the training dataset and will not generalise well.
All in all in this talk I will cover some of the approaches that can be used to train a Neural network based image classifier when given few examples from different classes. Audience will get to learn the concept of few shot learning, current research trends, common approaches to tackle this problem.
2. Hello!
I am Vaibhav
You can contact me on linkedin
https://www.linkedin.com/in/techytux/
● Payment Fraud @ Klarna
● Previously worked in Moderation Team @
OLX Group
● Software Engineer turned Machine
Learning Engineer
13. Classify a reference image into one of N candidate
classes, given K representatives of each class.
For example, 2-shot 4-way classification:
Concept : K Shot N Way
14. Humans can learn new concepts fast
without much training data
But for Deep Learning Networks we need
hundreds or even thousands of examples
What did we
learn
15. Motivation for few shot learning
● Machine learning fails to generalize when little supervised information is available.
● Getting labeled data can be expensive.
● Adding domain information / features for each task is cumbersome and does not scale well.
● Gradient based methods like Adagrad, ADAM, Adadelta, Nesterov momentum weren’t designed to
perform well under the constraint of a set number of examples
● Random initialization of weights hurts the networks ability to converge with few updates
16. What are commonly used datasets for few
shot learning
Popular
Datasets 2
17. Created by Lake in 2015
105 x 105 Grayscale Images
1623 chars from 50 alphabets each with
20 handwritten examples
Omniglot
20. Approaches to do low shot learning
Data Based
● Siamese networks (Koch et al. 2015)
● Memory Augmented Networks (Santoro et
al. 2016)
● Matching networks (Vinyals et al. 2016)
● Meta Optimization (Ravi & Larochelle 2017)
Metric or Parameter Based
● Tap into external data sources
● Generate new labels using semi supervised
learning. Propagate labels using a distance
metric
● Generate new examples using GAN’s. For
example: Create new images of birds from
different angles if there are examples
present in the training set
21. Related Work
● Fei Fei li -- Bayesian framework -- previously learned classes can be leveraged to help forecast future ones.
Fei Fei et al, 2003; Fei Fei et al, 2006
● Lake -- HBPL -- Concept learning through probabilistic programming
○ Penstrokes as domain information
Lake et al. 2015
34. Matching Networks
● Matching nets is built on top of
○ metric learning on Deep Neural features and
○ memory networks, basically external memory with attention mechanism
used to access the memory
● Proposes that training and test conditions must match
35. Matching Networks
Vinyals et al. 2016
● Network is shown sample set S and asked to give
probability that x(hat) is an instance of class in set S.
38. Optimization as a model for few shot learning
Ravi & Larochelle 2017
39. ● Instead of using gradient based optimization they propose
○ LSTM based meta learner to learn exact optimization algorithm used to
train another neural network classifier
● General initialization of network so that it converges faster
Meta Optimization