We studied the definition of learning maximum likelihood
estimation and consistent estimators.
We learned how to find the MLE when the parametric model is
given by a DFA. N-gram models are just one example of this where the
structure of the model is Strictly k-Local.
16 Nov 2017
Class canceled as instructor is away.
14 Nov 2017
We reviewed learning SL-k stringsets from an automata
perspective.
We reviewed learning SL-k stochastic stringsets from an automata
perspective.
We touched briefly on what it means to define learning in terms
of the MLE.
Read Chandlee on Input Strictly Local functions and why
they are of interest to phonology and morphology.
We discussed two types of classes for future directions
identification in the limit theorems: regarding all computably
enumerable stringsets from positive and negative data (Yes),
regarding the class of finite stringsets (Yes), regarding any class
of superfinite stringsets (No), and how to understand these results.
Extensions: SL tree languages, n-gram models (MLE of SL
distributions), ISL and OSL transductions (Chandlee 2014 et seq.),
SL tree transductions, and more!
12 Oct 2017
We went over a proof that SP-k stringsets are identifiable in the limit
from positive data.
We discussed informally the regular and noncounting classes of stringsets.
We defined Strictly k-Local stringsets and talked about how they
could be learned (SL handout).
10 Oct 2017
We discussed Heinz (2010)Learning Long-distance Phonotactics.
05 Oct 2017
We discussed why SP-k stringsets are identifiable in the limit
from positive data.
HW: Write a proof that SP-k stringsets are identifiable in the limit
from positive data.
03 Oct 2017
We continued discussing SP stringsets.
We sketched the idea behind on SP-k stringsets can be learned.
Read Heinz (2010)Learning Long-distance Phonotactics by Oct 10, 2017.
28 Sep 2017
We finished the discussion from last time.
We defined the Strictly Piecewise stringsets
(see here).
HW: Study the proof for learning the "bar" class of stringsets
and come to class prepared to ask about it.
26 Sep 2017
We talked about a concrete procedure which provides the n-th
string in the standard enumeration of all
strings.
We explained how it can be modified to learn a
non-trivial class of stringsets. Here is the
the handout.
21 Sep 2017
We discussed HW exercises and reviewed one definition of learning.
Alëna explained why most subsets of sigma star do not have a
solution to the membership problem in this handout.