Microsoft has introduced a framework for generating MvM images. MvM surpasses generative-adversarial neural networks, and also allows the use of new learning algorithms in computer vision tasks.
The capabilities of generative-adversarial neural networks (GAN) in computer vision tasks are limited by two main factors. Firstly, these neural networks model distributions using statistical characteristics, such as mean and moments, rather than geometric characteristics. Secondly, traditional GANS represent the loss of the discriminator network only in the form of a one-dimensional scalar value corresponding to the Euclidean distance between the real and fake data distributions. Because of these two problems, it is impossible to directly apply metric training methods, as well as apply new loss functions and training methods.
In MvM (Manifold Matching via Metric Learning), two networks are trained against each other. The metric generator network learns to determine the best metric for the distribution generator network, and the distribution generator network learns to create negative examples for the metric generator network.
Through competitive learning, MvM creates a distribution generator network that can create a fake data distribution close to the real data distribution, and a metric generator network that can provide an effective metric to capture the internal geometric structure of the data distribution.
Unlike GAN, MvM forms a multidimensional representation of images. This allows the use of such methods, teaching without a teacher. Microsoft claims that this fact will accelerate the study of generative models and open up new potential directions that were previously considered impossible.