Speaker: Marek Rychlik, Mathematics
Title: From Perceptron to Deductron - a tour of Neural Nets and Optical Character Recognition
Abstract: Over the past decade neural networks proved to be a winning strategy for many applied math problems. I will outline some classical nets, and the less known Recurrent Neural Nets (RNN). I will explain the connectionist approach which replaces "programming" with "training", using my new neural architecture, the Deductron.Optical Character Recognition (OCR) is a traditional playground for the connectionist approach, along with speech recognition and other complex signal analysis. OCR is about recovering text from images. The problem is essentially solved for English print. The recent focus is on cursive scripts (e.g. Arabic and Pashto) and handwriting, which present a formidable challenge. The most spectacular progress was achieved in the past decade with RNNs. The National Endowment for Humanities (NEH) funded a project (which I am leading) to perform OCR on 1 million pages of Pashto (the language of Afghanistan, similar to Arabic), and also traditional Chinese, which utilizes over 60,000 characters. I will discuss this project, and several research problems suitable for graduate research.