CVPR 2022 | meta learning performance in image regression task

Paper title ：

What Matters for Meta-Learning Vision Regression Tasks?

Thesis link ：

https://arxiv.org/abs/2203.04905

 

Code link ：

https://github.com/boschresearch/what-matters-for-meta-learning

 

Meta learning is widely used in small sample classification and functional regression because it can quickly adapt to tasks never seen in training . However , It has not been well explored in the regression task of high-dimensional input such as images . This paper makes two main contributions , It helps to understand this almost unexplored field . First , We designed two kinds of cross category visual regression tasks with unprecedented complexity in the past meta learning field , That is, object recognition locking and pose estimation .

So , this paper (i) The performance of common meta learning techniques on these tasks is evaluated in detail , and (ii) This paper quantitatively analyzes the effects of various deep learning techniques commonly used in recent meta learning algorithms to enhance generalization ability , Including data enhancement , Domain randomization , Task enhancement and meta regularization . in addition , this paper (iii) It provides some insights and practical suggestions for training meta learning algorithms in visual regression tasks . secondly , We suggest that in conditional neural processes (CNPs) Add functional contrast learning to the task expression learning space in (FCL), And train in an end-to-end way .

Experimental results show that , Due to improper selection of loss function and too small meta training set , The results of previous work are misleading . say concretely , We found that without fine tuning ,CNPs Better than on most tasks MAML. Besides , We have observed that the absence of task enhancements for design can lead to serious under fitting .