Last years deep learning became a main tool in AI research and development. There are still argues on why and how it works, but most of researchers and practitioners agree on one main concept – multilayer structure of deep neural networks allows to learn complex hierarchical features from raw data and these features are better than previous designed by humans.
Most of previous deep learning research was concentrated on architectures design in somewhat way that can learn most distinguishable patterns from different kind of data like convolutional networks for images or recurrent networks for natural language. So it means that today we end up with great feature extractors, that map raw complex data to some meaningful vectors. But how we can control what we learn? Or how we can push our neural networks to study some properties we want to? Of course we can design new convolutional architectures, or attention models / memory models on top of recurrent architectures, or some tricky regularizers, but what can we do beyond that?
In this talk we are going to discuss unconventional research areas in representation learning. We will start with general strategies of pushing neural networks to learn some particular patterns, that you can use right now. They include architecture augmentation, multitask learning, multimodal learning, unsupervised learning (mostly autoencoders and generative adversarial networks), loss functions design and optimization. After we will switch to research paths, that will include differentiable programming and interpreter learning, deep function machines, algebraic-like approaches, object oriented representation learning and others. Each of the concepts will be provided with an example of real-world use and recommendations on how to implement it in practice.
Alexandr is a machine learning expert with experience in solving computer vision and time series analysis cases in Ukrainian, Russian, American, and Italian companies. He joined Sincere Tech, a USA-based company with R&D department in Ukraine two years ago as a solution architect, where he helped to develop novel approaches to signal processing and became the cofounder of the company. At Sincere Tech, he is working on biomedical signal analysis, in particular ECG, and applying machine learning for classical applications like medical diagnostics and developing novel cases as well. Meanwhile Alexandr teaches deep learning seminars in University of Verona and writes a popular blog on Medium.
Event Timeslots (1)
Track A (Upper Floor)