Zero Sample Learning Method And Its Application Research

With the improvement on internet and computing ability,self-learning methods made breakthrough in many fields due to the effective information extraction from big data.In many datasets,certain type of samples may lack of ability for gathering,and this will make a bad effect on algorithm.To overcome this issue,computer scientists developed one-shot learning method,and it evolved into zero-shot learning category.Zero-shot learning is one of the most significant research areas for computer science,especially for transferable learning,with broad application prospects and potentials.Zero-shot learning algorithm recognizes or classify the unknown test samples through potential features extracted from training samples.Note that any test samples are forbidden during the whole training phrase.This algorithm may relieve the problem caused by edge data,and in medical case,certain class may face the problem of lacking samples as the difficulty for information collecting or mutation issues.In the experiment,totally three datasets are used.Animal with Attributes dataset and ImageNet-2 dataset for zero-shot learning,and MIASDB medical dataset for testing the algorithm applying on medical issue.In medical area,we use Word2Vector feature to replace semantic features as the dataset does not offer potential features for zero-shot learning.Firstly,modifying Semantic Auto-Encoder by adding kernel attribute and linear discriminant term,improves the result on the test data.Secondly,simplifying the sparse coding zero-shot learning method by replacing the dictionary to semantic features,makes the projection from semantic space to visual space with sparsity.,and the method achieves effective performance in zero-shot datasets,however,it performs instability on medical data as the inaccuracy semantic features.Finally,applying genetic algorithm framework into zero-shot learning method and testify the practicability.Zero-shot learning builds the connection between features of seen and unseen data by potential features.However,the open medical data lack the well-defined artificial labeling features which slows down the progress of zero-shot learning for medical application.The experimental results show that replacing semantic features by Word2Vector features performs instability,the results are highly correlated with the potential distribution of extracted words.