Research On The Semi-supervised Few-shot Classification Based On Siamese Network And GMM

In recently years,machine learning and deep learning technology has made a remarkable breakthrough.Deep learning technology based on convolutional neural network has achieved great performance in field of image classifcation.At the same time,these methods also revealed some problems: the need for large amount of labeled data,long training peroid,poor interpretability.In a real-world scenario,large-scale datasets do not always exist.A complex convolutional neural network can hardly be trained from a small amount of samples.Meanwhile,humans show great abilities to learn from very little information.The gap of learning ability between present machine learning models and humans has stimulate great interest in few-shot learning.How to learn a model from a few samples that generalizes well has become a hot topic in the field of machine learning.Current research in few-shot learning often adopts prior knowledge to make up for the insufficient samples.For example,training the model with other relevant data sets and then fine-tuning it with given data,or using semi-supervised learning to extract information from unlabeled data to benefit the training process.The thesis proposed a semi-supervised learning algorithm combining Siamese network and gaussian mixture model.Siamase network proves to perform pretty well on a small amount training samples.Two data point are fed into Siamese Network to be judged whether they are similar.This feature allows us to combine the data points to produce more training data.The gaussian mixture model is a widely used unsupervised classification model.The features in the image need to be extracted manually.Manually extracted features often require domain specific knowledge and are with great limitations.This problem can be effectively addressed with Siamese Network,which does not rely on specific domain specific knowledge to extract features,but only on convolutional neural Network.Combining the advantages of the two,Siamese Network is used to extract distinguishing features,and then gaussian mixture model is used to classify objects.The model can be trained with a small amount of labeled data sets and a large amount of unlabeled data sets.And the results turn to be good enough.Experiments on mini Image Net and omniglot show that the architecture based on Siamese Network and gaussian mixture model can be well applied to semi-supervised few-shot image classification tasks.