UNIVERSITY OF HERTFORDSHIRE
             COMPUTER SCIENCE RESEARCH COLLOQUIUM

                            presents


         "Hypothesis testing, Linear threshold functions, and
          Information Bottleneck Phase Transitions - What can
                  they tell us about Deep Learning?"

                        Prof. Natali Tishby
            (The Hebrew University, Jerusalem, Israel)

                     8 July 2015 (Wednesday)
                          1 pm - 2pm

                  Hatfield, College Lane Campus
                       Seminar Room B154 

                  Everyone is Welcome to Attend

                  Refreshments will be available





Abstract:

The Information Bottleneck method provides a way of describing
representations of relevant information in one variable with respect to
another one. It can be considered as a cascade of representation changes
through phase transitions, or cluster splits, that are most efficient for
encoding the relevant information. Interestingly, the information gained
by a cluster split can be effectively captured by a linear threshold
function, or a formal neuron. On the other hand, with limited data,
one can gain much by quantizing (coarse graining) the representation
before applying the classifier. We discuss the informational gains and
losses of these elements and relate them to the success and possible
design principles of Deep Neural Networks.

---------------------------------------------------
Hertfordshire Computer Science Research Colloquium
http://cs-colloq.stca.herts.ac.uk