Empirical Approach: How to Get Fast, Interpretable Deep Learning

摘要

We are witnessing an explosion of data (streams) being generated and growing exponentially. Nowadays we carry in our pockets Gigabytes of data in the form of USB flash memory sticks, smartphones, smartwatches etc. Extracting useful information and knowledge from these big data streams is of immense importance for the society, economy and science. Deep Learning quickly become a synonymous of a powerful method to enable items and processes with elements of AI in the sense that it makes possible human like performance in recognizing images and speech. However, the currently used methods for deep learning which are based on neural networks (recurrent, belief, etc.) is opaque (not transparent), requires huge amount of training data and computing power (hours of training using GPUs), is offline and its online versions based on reinforcement learning has no proven convergence, does not guarantee same result for the same input (lacks repeatability). The speaker recently introduced a new concept of empirical approach to machine learning and fuzzy sets and systems, had proven convergence for a class of such models and used the link between neural networks and fuzzy systems (neuro-fuzzy systems are known to have a duality from the radial basis functions (RBF) networks and fuzzy rule based models and having the key property of universal approximation proven for both). In this talk he will present in a systematic way the basics of the newly introduced Empirical Approach to Machine Learning, Fuzzy Sets and Systems and its applications to problems like anomaly detection, clustering, classification, prediction and control. The major advantages of this new paradigm are the liberation from the restrictive and often unrealistic assumptions and requirements concerning the nature of the data (random, deterministic, fuzzy), the need to formulate and assume a priori the type of distribution models, membership functions, the independence of the individual data observations, their large (theoretically infinite) number, etc. From a pragmatic point of view, this direct approach from data (streams) to complex, layered model representation is automated fully and leads to very efficient model structures. In addition, the proposed new concept learns in a way similar to the way people learn - it can start from a single example. The reason why the proposed new approach makes this possible is because it is prototype based and non-parametric.