MITCHELL, Tom, NELL Never Ending Learning, CMU, April 16, 2014
<p class="MsoNormal" >MITCHELL, Tom, NELL Never Ending Learning, CMU, April 16, 2014
15 min talk at SWANK Conf at Stanford. Tom is head of ML at CMU.</p>
<p>
want to learn many things and progr. improve over time</p>
<p>
<b>NELL Never Ending Lanugage learner...</b>
<b>runs forever 24 by 7</b>
<b>each day: extract more facts from the web</b> to populate the ontology
but it should also be better able to read today than yesterday.</p>
<p>
its inputs: initial ontology (categories and relations)</p>
<p>
dozens of examples of each
and occasional interactions with human trainers.</p>
<p>
<b>running since Jan 2010</b>!</p>
<p>
still learning</p>
<p>
has built up <b>70 M belief triples</b>...</p>
<p>
Maple Leafs won the Stanley Cup and so did the Red Wings whose hometown is Detroit.
Toyota acquired Hino;
Toyota created Prius and Corrola</p>
<p>
Problem 1: NELL learns mostly from unlabeled data.</p>
<p>
solution for unlabeled data: humans use multiple senses
(watching a ball bounce and hear the bounce)</p>
<p>
NELL: look at noun phrase... MOUNT EVEREST</p>
<p>
Mountain is assoc with volcanoes, winter ascents, southeast ridge...</p>
<p>
NELL learned that PEAK has a weight of 1.791 assoc with mountain</p>
<p>
it can learn from multi-views
he has many papers on this.</p>
<p>
simult learning of various aspects of an object...</p>
<p>
Problem 2: reasoning over the KB...</p>
<p>
a rule:<b> if X competes with Y and Y is a company then</b>
<b>X is a company</b>.</p>
<p>
Key idea 3:
<b>automatically extend its ontology...</b></p>
<p>
<b>new predicates to add to its ontology</b>..</p>
<p>
work of Mohamed, Hrusha, Mitchell, EMNLP 2011</p>
<p>
eg people and musical instruments...</p>
<p>
for each pair of categories c1 and C1 find pairs like</p>
<p>
sitar with George Harrison
tenor sax with Stan Getz</p>
<p>
suggests name for this relation Master or virtuoso, legend, plays</p>
<p>
NELL sample of self-added relations</p>
<p>
athlete-won-award
animal-Eats-Food
etc.</p>
<p>
Key idea 4: cumulative, staged learning</p>
<p>
learning X improves ability to learn Y.</p>
<p>
future: add a robot body to NELL (and other ideas)</p>
<p>
</p>