BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//Talks.cam//talks.cam.ac.uk//
X-WR-CALNAME:Talks.cam
BEGIN:VEVENT
SUMMARY:Visualizing and Understanding Recurrent Networks - Andrej Karpathy
 \, PhD student\, Stanford University
DTSTART:20150806T110000Z
DTEND:20150806T120000Z
UID:TALK60270@talks.cam.ac.uk
CONTACT:42888
DESCRIPTION:Recurrent Neural Networks (RNNs)\, and specifically a variant 
 with Long Short-Term Memory (LSTM)\, are enjoying renewed interest as a re
 sult of successful applications in a wide range of machine learning proble
 ms that involve sequential data. However\, while LSTMs provide exceptional
  results in practice\, the source of their performance and their limitatio
 ns remain rather poorly understood. Using character-level language models 
 as an interpretable testbed\, we aim to bridge this gap by providing a com
 prehensive analysis of their representations\, predictions and error types
 . In particular\, our experiments reveal the existence of interpretable ce
 lls that keep track of long-range dependencies such as line lengths\, quot
 es and brackets. Moreover\, an extensive analysis with finite horizon n-gr
 am models suggest that these dependencies are actively discovered and util
 ized by the networks. Finally\, we provide detailed error analysis that su
 ggests areas for further study.
LOCATION:Cambridge University Engineering Department\, LT1
END:VEVENT
END:VCALENDAR