Keynote speech I: Big data, non-big data, and algorithms for recognizing the real world data

Abstract

In this talk, we focus on recognition of static images, motion images from a video, and speech waves spoken by simultaneously multiple speakers. The necessary size of data for learning depends on algorithms for recognizing patterns. Real world data of static images, motion images, and speech waves includes many kinds of problems to be solved for their recognition. The most important one is the separation of segmentation and recognition in both time and space domains as well as overcoming their non-linear variations of these patterns. The segmentation problem is strongly coupled with the recognition problem. Without segmentation, recognition is impossible and vice versa. We need to create a sophisticated algorithm for decoupling of the two. If the recognition algorithm itself can also solve both the problems of segmentation and overcoming problem of non-linear variations of these patterns in the inside process of recognition, big data is not required for learning. On the other hand, deep learning is requiring big data of segmented samples for storing them in the form of connection weights among nodes of multi-layer. Deep learning is basically based on the segmentation of patterns in both learning and recognition stages. We propose two algorithms of matching. The one is called two-dimensional continuous dynamic programming (2DCDP) for spatial segmentation-free recognition of static images. An expanded version of 2DCDP called incremental two-dimensional continuous dynamic programming (I2DCDP) can carry out time segmentation-free and speaker-independent recognition of a single speech wave spoken by multiple speakers without speech separation. The other one is called time-space continuous dynamic programming (TSCDP) for both time segmentation-free and location-free recognition of complex human/object motions from a video even in the moving background. The two algorithms can solve automatically the decoupling problem of segmentation and recognition. They can also solve the problem for overcoming non-linear variations of static images, motion images and speech waves by through the inside process of recognition algorithms. Therefore, a quite small size of data of static images, motion images and speech waves, respectively, is enough for recognizing actual these real data of wide range. We will show many experimental results for confirming our argument.