Signal

by Devansh Tripathi·6 views·Updated Aug 4, 2023

Create your own

Chapters

Audio TranscriptAuto-generated

We live in a dynamic era filled with uncertainty and marked with unexpected events
of wars, cybercrime,
data breaches,
conflicts, protests, riots, economic recession
and much more,
Hen said.
Every citizen or any user is vulnerable to these threats present around them
and also unknown to them.
Therefore, we need a digital mechanism
that can enable
the generation of 21st century to know about any threat
that surround them
and make them aware of it.
One of the technologies which has come up recently, is location entity recognition
that extracts crucial information from online data
and
text due location associated with them,
and then informs the concerned user regarding the threat if there are any,
and then
if you boil down location, entity recognition,
the core technology behind that is named entity recognition.
Now various named entity recognition models have been in use. Clavin
not is one such model used by Signal Corporation currently,
and there are other models as well, like Spacey Flare Gate, Open LP,
which are currently in use
Now. One drawback about all these models is that
they're heavily dependent upon
a data set
which has to be big enough and also accurate enough for the model to work well.
Unfortunately,
we don't have a vast and precise data set
present with us,
because
most of the data present in the social media platform is scattered
and cluttered.
Therefore, if the data set is not strong enough, as you can see from this diagram,
the model which is being trained is also weak enough.
And hence, when it acts upon the real time data,
the results that we get will be highly inaccurate.
And that's a problem
that we are facing as of now,
when it comes to named entity recognition system for security surveillance.
Therefore, we need a model that can directly act upon real time data
and gradually improve with time.
One such model,
which has been
a
new, exciting thing
in the realm of machine learning, is reinforcement learning,
and it learns, just like a human,
just like a baby, learns to walk
it slowly starts to crawl
and then tries to stand up, and then it falls.
Then again, it crawls and then again stand ups,
and
one day it stands on its feet
and gradually,
steadily,
it learns to put its steps forward,
and I finally learned to walk.
Now This method is also employed here, where an agent
learns to
perform any action
through a feedback mechanism
it performs, the action, receives the feedback
and then
it gains a reward
and
the next action it performs is to
increase that reward.
Hence, this becomes a
process of sequential decision making
where the next decision it makes
is to
increase the reward which it received
from the previous outcome.
Thus,
this process of feedback mechanism,
action, reward, state and then action reward state
and this continuous learning mechanism between agent and involvement
is a method
that can help us
crack the dilemma of named entity recognition
and also help us tackle with the constraint of the dataset,
which with other models have been facing
therefore
named entity recognition
as a reinforcement learning problem
can be visualised as a problem where
a textual data is given to the agent
and the goal of the agent is to
maximise the named entities it can extract from the data.
Now this picture that's in front
gives a rough idea for how named
entity recognition system can work using reinforcement,
learning
the textual data will be given
and then the named entities will be named.
Now the initial action
will give us the worst outcome.
It will probably tag every word which is being present in the text.
But gradually it receives the feedback
and then it tries to improve the reward that it gets.
The more the entities it names, the more the reward it gains.
And
as
we move with time,
the actions that it leads to
will be the ones that maximises. The reward
ultimately maximises the named entities extracted from a piece of text.
Now various candidate reinforcement learning algorithms are present.
One such is called natural language policy optimisation,
which is on policy method of reinforcement learning
where a policy structure is kind of given to a reinforcement learning model,
it uses that policy
to act upon
and perform upon the
data, which is real time.
This is a mass extension of a proximal policy optimisation process.
And here what basically happens is that
a top P sampling where P being the parameter,
leaves a restricted tokens to the smallest possible set,
whose cumulative probability is greater than the parameter.
Thus,
the best set of actions is basically selected
to
perform the
next set of actions
and then gain or maximise the reward
now same way. There is other method called GRUE General Reinforcement Language
Understanding evaluation
here. What happens is that
a set of seven generating and
PTAs
happen in real time,
like this ambiguous
tokenization word splitting.
And then
there is an evaluation metric simultaneously happening upon those actions.
Now these evaluation Mac metrics
help the actions
to perform such that
the reward that it gains is the maximum one.
Thus,
it's a strict model
where evaluation takes place upon the actions
to give the best feedback possible and maximise
the reward of naming the most entities possible.
Now. Reinforcement Learning for Language model is another package
which basically uses the algorithm for
off actor critique.
Now this happens in such a way that actor is basically
a policy structure.
While critique is a valuation function,
the critique valuation function
tries to minimise the reward while the actor policy structure tries to maximise it.
This way, accumulative learning happens between the two,
hence improvising the overall reinforcement learning model.
Now any of these algorithms can be applied
to enhance the named entity information present,
and also with us are various tools which is being
in use tensorflow agent is one such tool
used for building reinforcement learning models open.
I have also developed spinning Up, which is a library
for researching and developing reinforcement learning models.
Deep Mind Lab has also performed various
results and given various results regarding reinforcement, learning
and any of these tools, along with any of the
reinforcement learning algorithm,
can be embedded into named entity recognition system,
which can help us to act with real time data.
And once we are no more constrained with the presence of a data present with us,
a strong data present with us,
we can directly employ the model on real time data present online.
And then
as we move with time,
the model can maximise
the outcome that it produces
and with this outcome is
with being presented with
Signal Corporation can hence employ a
much more better
and enhanced named entity recognition system.
And this can lead to the path for
an open source intelligence where any person any user
can use the intelligence system
of named entity recognition system
empowered by reinforcement, learning
to extract
any information concerning that security
and with noise all around present on the Internet
signal
can employ models
that can help any layman, any user or any innocent person
to be aware about their security
and also protect them
in the most intelligent way possible
and make it easy
for any user
to be aware about their security
and also
help the society
become a more secure place
like never before.
Thank you. This was presented by the
party.