A few more important points about belief
networks, which are essentially Bayesian
networks and their generalizations.
So far, we have talked about using,
Bayesian networks to model simple
situations like, the sprinkler, rain and,
wetness, etc., as well as to learn facts
from text.
But, we haven't asked where these networks
come from.
So far, we have used networks which we
have imagined from intuition or judgement.
It turns out that, not only can the
conditional probabilities in the networks
can be learned from data, but the network
structure itself can be learned from data.
One of the greatest examples of learning
network structure from large volumes of
data has been in genomic medicine, or in
medicine in general.
Medical diagnosis, for example.
What treatments are best for what symptoms
has been around for many years, and has
used Bayesian networks very successfully
to build assisted systems for medical
diagnostics, especially in regions where
there aren't that many qualified doctors.
In genomic medicine, the relationship
between different genes expressing
themselves in an organism has been learned
from large volumes of experimental data
using, using techniques for learning the
structure of Bayesian networks.
Similarly, how phenotypes that is traits
that are exhibited in an organism arise
from the genes of that organism, has also
been, learned Bayesian networks which can
be inferred from large volumes of data.
When it comes to logic and uncertainty,
there is a growing realization that belief
networks are really bridging the gap
between the fundamental limits of logic
and the pitfalls of uncertainty.
The indication of this is the fact that
Judea Pearl, who invented Bayesian
networks was given the Turing award, which
is the highest award in Computer Science,
in 2012.
His initial work on Bayesian networks was
in the early 90s' and in fact, he's
written a recent book on causality, which
is still a deep subject, not completely
covered or even explored using Bayesian
networks.
Other kinds of networks that merge logic
and probability are Markov logic networks,
conditional learning fields, and many
others.
We won't have time to even touch these in
this course, but they are all forms of
belief networks that bridge the gap
between logic and uncertainty.
Coming to big data,
Well, we have seen that inference in such
networks can be done using SQL We've shown
this for Bayesian networks but the fact is
that it's all counting and map-reduce
actually works.
Because if we're just doing counting in
SQL, we can actually do inference from
large volumes of data using the big data
technologies.
As regards our hidden agenda about AI,
Deep belief networks, which we will study
a little bit in the last week of lectures,
is one direction in which connectionist
models of the brain are being explored and
extended, which sort of brings everything
back together, that all the things that
we're studying in probability, statistics,
Bayesian networks learning, eventually is
teaching us more and more about how the
brain works.
This has been a long lecture.
We've covered a lot of ground.
And it's worth recapping what we've
actually learned.
First, we began by saying that search is
not enough for general question and
answering on the web, which lead us to
reasoning.
Logic and the semantic web vision is one
way of trying to address this problem of
how computationally a web intelligence
system could actually answer a general
purpose question.
We learned, of course, that there are
fundamental limits to logic as well as
practical ones arising from uncertainty.
We went into studying how reasoning under
uncertainty could be handled using
Bayesian networks and probabilistic
graphical models in general.
Though we didn't cover the latter, we
indicated the direction in which this
field is going.
In the next few weeks, we will have a
programming assignment next week.
This will be on Bayesian inference using
SQL.
We will have a short lecture video next
week to explain the assignment, but do
start preparing by studying the SQL-based
inference that we've done in this week as
well as experimenting with a sequel engine
of your choice, I would suggest SQLITE3,
SQL Lite three, which comes bundled with
Python, and for which one can use an
in-memory data base which will pretty much
suffice, since the tables we'll be using
will be quite small.
The final week, will have the predict
lecture, where we'll put everything
together, as well as have our final
programming assignment and then the last
week, where there'll be the final exam,
and you'll be asked to complete all your
assignments by then.
So, do complete the homework and quiz for
this week and start preparing for Bayesian
inference using SQL and experimenting with
SQL Lite for the next programming
assignment.
The last week, which will have the final
exam, and the submission dates for all the
programming assignments from here onwards.
So, please prepare for next week's
programming assignment and do remember to
complete the quiz and homework for this
week, which are due only next Friday,
since we don't have a full lecture next
week.