Interpolating between Optimal Transport and MMD using Sinkhorn Divergences

Reading notes on Interpolating between Optimal Transport and MMD using Sinkhorn Divergences

The purpose of this paper is to show that the Sinkhorn divergences are convex, smooth, positive definite loss functions that metrize the convergence in law.

Countless methods in machine learning and image processing reley on comparisons betwen probability distributions. But simple dissimilarities such as the Total Variation norm or the Kullback-Leibler relative entropy do not take into account the distance d on the feature space χ\chi. As a result, they do not metrize the convergence in law and are unstable with respect to deformations of the distributions’ support. Optimal Transport distances (sometimes refeered as Earth Mover’s Distance) and Maximum Mean Discrepancies are continuous with respect to convergence in law and metrize its topology when feature space χ\chi is compact.

Read More

DeepEMD A Few-Shot Image Classification method

A reading note on DeepEMD: Few-Shot Image Classification with Differentiable Earth Mover’s Distance and Structured Classifiers.

Illustration of Earth mover distance


Deep Neural Networks achieve high performance under the large labelled datasets. For some circumstances, no enough labelled images are provided. One of most well-studied machine learning algorithms is few-shot image classification. Only with small labelled data, few-shot algorithms can categorize new images.

Read More



Read More

What are the difficulties in avoiding direct discrimination on the basis of a protected characteristic (e.g. gender) when creating AI systems?

This is my formative writing task in Ethics and Regulation of Artificial Intelligence(LAWM161) at University of Surrey.

Professor: Mikolaj Barczentewicz

Over the past few years, lots of services or products with artificial intelligence have come to people’s life. After several waves of research in Artificial Intelligence, some AI technologies have been emerging like Machine Learning, Deep learning and Neural Networks. Although great progress has been made, research on ethics and laws still lags behind the development of technology. Besides, it turns out that prejudice and discrimination between people will also appear in AI systems easily due to the wide spread of AI systems. There are many reasons for discrimination. In the training process of the AI model, some inappropriate training data may be used, which may contain discrimination information. At the same time, the designer of the AI system will also entrain some personal emotions when designing. Or the idea that the decision made by the Ai system is discriminatory. Even the wrong use of the AI system by users will make them feel treated differently. In this article, I will focus on the direct discrimination when creating an AI system, and discuss some difficulties to decrease direct discrimination.

Read More


1. 什么是Guard


  1. 用户的注册信息存入数据库(登记)
  2. 从数据库中读取数据和用户输入的对比(认证)


Read More