We will discuss what few-shot learning is, how to do it right, what kind of problems it can solve and how we use it in our projects.
Deep learning now is achieving great results for new tasks, but a the same time modern architecture becomes more and more data and resource hungry. You should collect a huge dataset for achieving an acceptable performance of your model. Few-shot learning technique aims to learn a model to recognize unseen classes during training with limited labeled examples.
In this talk, we will deeply look at the few-shot learning approaches. Investigate different architectures and datasets. Figure out when the few-shot learning works and when does not. In the end, you should be able to understand all landscape of few-shot learning field and could continue your journey in this area.
Website: https://fwdays.com/en/event/data-science-fwdays-2019/review/practical-few-shot-learning
6. Life is good in Deepland!
1. “No need to code anymore”
2. For any given problem, just:
a. label some training data
b. define an objective function
c. train neural network
d. …
e. …
f. ..
g. sell your startup for millions!
Why?
The Revolution will not be Supervised
8. 1. Data labeling is pretty expensive
2. New classes have arrived
Why?
9. 1. Data labeling is pretty expensive
2. New classes have arrived
3. The Revolution will not be Supervised
Why?
The Revolution will not be Supervised
16. Overview
1. Initialization based methods: “learning to fine-tune” & “learning an optimizer”
1. Distance metric learning based methods: “learning to compare”
1. Hallucination based methods: “learning to augment” & GANs
A Closer Look at Few-shot Classification
28. Large Scale Object Detection & Instance
Segmentation
Initially
collected
dataset
n classes
n+1 class
n+m class
Annoy & Faiss
29. 1. A Closer Look at Few-shot Classification
2. Prototypical Networks for Few-shot Learning
3. Matching Networks for One Shot Learning
4. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
5. On First-Order Meta-Learning Algorithms
6. Learning to Compare: Relation Network for Few-Shot Learning
7. Exploring the Limits of Weakly Supervised Pretraining
Few-Shot-Learning