Speaker: Dana Wilkinson
There are a variety of domains where it is desirable to learn a
representation of an environment defined by a stream of sensori-motor
experience. This talk introduces and formalizes Subjective Mapping, a
novel approach to this problem. A learned representation is
subjective if it is constructed almost entirely from the experience
stream, minimizing the requirement of additional domain-specific
information (which is often not readily obtainable).
In many cases the observational data may be too plentiful to be
feasibly stored. In these cases, a primary feature of a learned
representation is that it be compact---summarizing information in a
way that alleviates storage demands. Consequently, the first key
insight of the subjective mapping approach is to phrase the problem as
a variation of the well-studied problem of dimensionality reduction.
The second insight is that knowing the effects of actions is critical
to the usefulness of a representation. Therefore enforcing that
actions have a consistent and succinct form in the learned
representation is also a key requirement.
This talk presents a new algorithm, Action Respecting Embedding (ARE),
which builds on a recent effective dimensionality reduction algorithm
called Maximum Variance Unfolding, in order to solve the newly
introduced subjective mapping problem. The resulting learned
representations are shown to be useful for reasoning, planning and