Deep Learning-Based Few-Shot Learning Method

Few-shot learning technology refers to the technique of quickly learning and generalizing to a new task in a situation where labeled samples are limited.This technology is an active exploration of artificial intelligence towards true intelligence,and has been empowered in applications such as scarce labeled data and limited scenarios.In recent years,few-shot learning methods have developed rapidly,but how to better learn the task-relevant knowledge and how to alleviate the over-fitting problem caused by the inconsistent data distribution between training and test phase are still the biggest challenges of few-shot learning.With regard to this problem,this paper proposes a distribution estimation method and an iterative domain adaptation learning strategy,which solve three specific tasks of classification,segmentation and detection,on the basis of two limited scenarios.Compared with the prior work,the main contributions of this paper are as follows:First,in the few-shot classification task under novel class with scarce labeled data scenario,this paper first points out the disadvantages of commonly used point-based estimation approaches,which are inherently noisevulnerable and easy-to-be-biased.In order to alleviate these risks,we propose a distribution-based estimation algorithm instead.Specifically,this algorithm uses the variational inference strategy in Bayesian learning to refine the distribution estimation pipeline,and transforms the original few-shot classification problem into two simpler subproblems: precisely estimating the posterior distribution and leveraging the posterior distribution to complete the inference.In the process of solving the subproblems,this paper proposes a distribution aggregation rule and an inference rule.A categoryspecific distribution that describes the uncertainty of each sample is thus obtained,instead of the original deterministic category representation.This solution greatly improves the insensitivity of the model to few supervision.Second,in the few-shot segmentation task under the same scenario,this paper implants the former distribution-based estimation framework into a baseline algorithm of this task,and further verifies the portability and scalability of this framework with minimal overhead.Finally,in the few-shot detection task under new domain with no labeled data scenario,this paper first proposes a domain adaptive detection algorithm.Through the adversarial training of the domain discriminator and the feature extractor,the domain discriminator is able to distinguish the domain attribute,while the feature extractor can extract features which confuse the domain discriminator.On this basis,this paper further proposes a multi-stage iterative learning strategy,and at the same time interacts between the domain discriminator and the detector via an attention mechanism.As a result,the detector which is suitable for the source domain,is gradually transferred to the target domain.This paper conducts sufficient experiments on public datasets of fewshot classification,segmentation,and detection tasks.These experiments consist of performance comparison with contemporary state-of-the-art algorithms,properties analysis of the proposed algorithms,and visualization results.Extensive experimental results on these benchmarks well demonstrate the effectiveness of our few-shot learning framework over previous methods,in terms of superior classification accuracy and robustness.