SEKE 2012 Proceedings - Knowledge Systems Institute

A never-ending language learner called NELL was
described in [3]. NELL u tilized semi-supervised learning
methods and a collection of knowledge extraction methods
to learn noun phrases from specified semantic categories
and with specified semantic relations. NELL has four
component learners: a pattern learner, a se mi-structured
extractor, a morphological classifier, and a r ule learner.
NELL also accommodates human interaction to approve or
reject inference rules learned by the rule learner component.
The work in [1] reported an agent system called ALICE
that conducted lifelong learning to build a set o f concepts,
facts and generalizations with regard to a particular topic
directly from a large volume of Web text. Equipped with a
domain-specific corpus of t exts, some background
knowledge, and a control strategy, ALICE learns to update
and refine a theory of the domain.
The results in [10] defined continual learning to be a
continual process where learning occurs over time, and time
is monotonic. A continual learner possesses the following
properties: the agent is autonomous; learning is embodied in
problem solving, is incremental, and occurs at multiple time
steps; and there is no fixed training set. Knowledge an agent
acquires now can be built upon and modified later. A
continual learning agent system called CHILD was
described in [10].
YAGO2 is a lar ge and extendable knowledge base
capable of unifying facts automatically extracted f rom
Wikipedia Web documents to concepts in WordNet and
GeoNames [7]. YAGO2 exhibits its continuous learning
capability by allowing new facts to be incrementally added
to an existing knowledge base. Knowledge gleaned by
YAGO2 is of high quality in terms of coverage and
accuracy.
The results in [4] deal with clustering with inconsistent
advice. Advice such as must-link and cannot-link can be
incorporated into clustering algorithms so as to produce
more sensible groups of related entities. Advice can become
inconsistent due to different reasons. Clustering in the
presence of inconsistent advice amounts to finding
minimum normalized cuts [4].
Bias shifting can be a useful technique in the area of
transfer learning [9]. However, there are a n umber of
important differences between the aforementioned work and
the focuses of research in this paper. (1) i 2 Learning
emphasizes on the stimulus for perpetual learning, i.e., the
learning episodes of the agent are tri ggered by
inconsistencies it encounters during its problem-solving
episodes. This is not necessarily the focus in related work.
(2) i 2 Learning has a problem-solving slant, i.e., learning to
incrementally improve performance for solving problems at
hand, whereas the related work in [1,3,7,10] is p rimarily
geared toward the general task of knowledge-acquisition, or
building an ontology or a dom ain theory. (3) The learning
episodes also differ: i 2 Learning adopts discrete learning
episodes (as triggered by conflicting phenomena), whereas
the learning episodes in related work of [1,3,7,10] is largely
continuous, not necessarily triggered by any events. (4)
Inconsistencies are utilized as essential heuristics for the
perpetual i 2 Learning, whereas inconsistent advice in [4] is
only used as constraint for clustering. (5) Mos t of the
related work (with the exception of CHILD) is web-centric
in the sense that learning is carried out with regard to web
texts. i 2 Learning, on the other hand, accommodates a broad
range of heuristics in its learning process.
3. i 2 Learning: A Framework for Perpetual
Learning Agents
A perpetual learning agent is one that engages in a
continuous and alternating sequence of problem-solving
episodes and learning episodes. In such an alternating
sequence, learning takes place in response to a whole host
of stimuli, including inconsistencies encountered in the
agent’s the problem solving episodes. Learning episodes
result in the agent’s knowledge being refined or augmented,
which in turn improves its p erformance at tasks
incrementally. We use learning burst and applying burst to
refer to reoccurring learning episodes and knowledge
application (problem solving) episodes.
The proposed i 2 Learning framework focuses on a
particular scenario for the aforementioned perpetual
learning agents: one in which the learning episodes of the
agent are tri ggered by inconsistencies it encounters during
its problem-solving episodes and the perpetual learning
process is embodied in the continuous knowledge
refinement and revision so as to overcome encountered
inconsistencies.
A perpetual learning agent has the following
components: (1) a kn owledge base (KB) for persistent
knowledge and beliefs, domain or ontological constraints,
assumptions and defaults; (2) a meta knowledge base (mKB)
for the agent’s meta-knowledge (knowledge on how to
apply domain knowledge in KB during problem solving); (3)
a working memory (WM) that holds problem specific facts,
and facts deduced with activated beliefs from KB; (4) a
reasoning mechanism to facilitate problem solving process;
(5) a component called CAL (Coordinator for Applying burst
and Learning burst) that recognizes inconsistency and
initiates learning bursts; (6) a bias space containing
candidate biases for the learning process; and (7) a learning
module i 2 Learning that carries out inconsistency-induced
learning to refine or a ugment KB, mKB, or WM (or any
combination of the three) so as to ov ercome encountered
inconsistencies. Figure 1 captures the structure of perpetual
learning agents.
When a co nflicting situation arises in WM during its
problem solving process, the agent’s CAL detects it,
suspends the current problem solving session, initiates the
next learning burst via passing the specific inconsistent
circumstance to t he learning module, and waits for the
result from the learning module. The learning module
i 2 Learning in turn carries out the learning process by
recognizing the type of inconsistency in the conflicting
circumstance, and selecting the appropriate in consistency-
250