Machine learning and AI in action | TECH(talk)

Machine learning and AI in action | TECH(talk)


hi and welcome back to Tech Talk I’m Ken
Mingis executive editor Computerworld I’m here today with Juliet Beauchamp and
InfoWorld’s Serdar Yegulalp about artificial intelligence and
machine learning we’ve got Sardar back for more of a an interesting discussion
a quick reminder if you’re watching us live we’re streaming on our YouTube
channel please subscribe if you like what you see
and we’re also streaming in the computer world LinkedIn page please follow the
page if you want to keep track of what we’re up to for future episodes Sardar
welcome back thanks again hi thanks for being here
we actually have some questions from from the last show that some people had
submitted so I thought I would throw those your way and see if we can get
some answers for the viewers sure thing okay question number one from soo J
which frameworks or tech languages can be used to build web-based machine
learning models well the easy point of entry here is the Python language since
that’s the one that’s currently dominating the field of working with
machine learning I should point out that Python itself is not what is actually
being used to construct machine learning models it’s more of a wrapper for the
libraries that are written in C or C++ that do all the actual heavy lifting
okay but it makes the job far easier so getting started with that and getting
started with the libraries that are associated with Python for machining
like tensorflow those are always a great start I should
point out that the go language is actually starting to also show a lot of
progress in this space although not quite to the same extent again also the
giulia language they’re both coming up in this space with many of the same kind
of conveniences but python has been successful in big part because it is so
convenient to work with and so that’s always been a very good low barrier of
entry for getting started with this kind of work yeah I was gonna in fact just to
sort of follow up on that what is it about these languages that makes them
better or more ideal for for coating around ml well one of the things about
working with something like machine learning is that a lot of the heavy
lifting again is done in a separate library which is written in a much more
powerful language than Python itself is but working directly with those
libraries in the languages that they’re written and is something of a headache
and Python takes away a lot of that headache by abstracting away a lot of
the behaviors that you would normally have to deal with like memory management
or the handling of objects so you can you can deliver very high-level
instructions to the library for how to handle a machine learning workflow and
that way a lot of the work that you would
normally have to do yourself is pushed out of the way you can concentrate on
things like keeping your data clean or you know making sure that the whole
pipeline is functioning correctly got it okay cool that’s where that’s really
useful information a second question from mode is how one how can one get
started with artificial intelligence and machine learning what coding languages
should they learn you’ve sort of already talked by then ago and also the julia
language which is which is starting to become more popular who’s specifically
developed for math and statistics and and scientific computing and is designed
sort of to be a best of all worlds type of thing and that it’s supposed to be
the easiest uses Python and also as fast as C++ in many respects but it’s still
it’s still very young still only a 1 point something release at this point
but it’s getting a lot of attention for how it addresses these problems so I
would I would pay attention to where or Julia is going within the space in the
future with Python is never a bad place to start with it and on another level
that the how to go about doing it generally get started with some toy
projects one of the beautiful things about Python is that you can set up
something on a computer on your own and get started right away with models that
you can download you know from the from the internet publicly and you can that
way you can get used to the workflow that’s involved where you acquire a data
set you make it clean you you bring it into conformance you train a model
around it and then you deploy it even if only on a toy web serve and you serve
predictions from it the last part in particular can be the
hardest that the deployment the whole pipeline assembling the pipeline for the
deployment can be trickiest and because there isn’t even really one you know
doctrinaire way to do it there are ways to do it but there isn’t it’s it’s not
like there’s a de facto standard that it’s been established there’s simply a
set of best practices that you have to try to hew to as closely as you can so
being having some local prac this for doing this as always is always
a good thing if you want to try doing this stuff in the cloud where you have a
you know access to much bigger datasets or you know much better training
hardware data bricks for instance has ML models and ml training systems that you
can use in the cloud in an experimental way where you can you can work with them
a little bit and then throw them away when you’re done or you can you know
purchase a full plan upgrade from the free one and then do real work in
earnest there so play with it that’s my best suggestion that’s always a good way
to start just a quick reminder if you’re tuning in we’re talking to Sir darf
event for world about artificial intelligence and machine learning
answering some earlier questions and then he’s gonna get into some more some
more details about what’s coming up what’s what’s the third question we got
to yes our third question is from Shah and they asked what data analysis skills
are needed for using / running a AI so beyond knowing Python I guessed right I
mean obviously you need to have some skill with programming with wrangling
you know a piece of software into into the pipeline and and to think logically
about problems but the big thing that I would start with is if you don’t have
any knowledge just statistics then start there you want to pick up a course in
statistics and understand things like p-values and correlations you know what
these things mean what they don’t mean because if you don’t you’re gonna be
foundering and you’ll you want to do that also to separate your math skills
from your programming skills because the two don’t always have a lot of overlap
programming is often a lot more about logical thinking and about designing a
project about you know being able to break a problem down lots of little
steps and math is totally separate from that so if you–if you separate those
two and you focus on on statistics as its own learning experience that’s going
to pay off enormously further on as I remember statistics from college so much
for my AI ml coding career that’s already over with good alright that’s
that’s really good advice so it’s really important to have yeah so what we wanted
to talk a little bit about you know moving forward now before we see if
there’s any new questions about how organizations are really starting to use
AI and ml we had someone on last week who dealing with robotics and I was
surprised to find out when Keith Shaw was here that you know machine learning
is is is becoming more of a component of some of the robotics of it yeah
developed yeah but even just going beyond robotics there’s so many other
practical use cases that are on are not as easy to physically see as a robot so
what are something that’s always the problem isn’t it yeah no it’s true I
mean it’s so easy to look at a row button and say obviously they’re using
artificial intelligence or and or machine learning but it’s a little more
difficult to look at how people are running like data analysis using AI so
what are some examples that you’ve seen there prizes are using that that can be
tough you know it’s always it’s always easy to point to something like a
self-driving car but so glamorous yeah yeah and so many of the real benefits of
this sort of thing are not glamorous and in-your-face they’re they’re a little
pardon the term nerdier they’re a lot more out of the way what it boils down
to is that the biggest benefits come from being able to take data and
recognize patterns in it that give you insight that you didn’t have before it’s
not necessarily about predicting what’s going to happen it’s about getting
insert insight into what is happening right now so this could be for instance
fraud detection this is actually one of the most common uses of this kind of
pattern detection I see as you know if somebody suddenly makes a purchase from
a geographic place that they have never been to before that’s a possible flag
for fraud and you have an AI ml system that will take bits of data like that
and correlate them and then and then generate a report and say there’s a high
likelihood that this person’s account has been broken into as being used for
fraudulent activity and there’s other things in the same vein for instance
customer segmentation figuring out who uh what bucket a customer is likely to
fall into if they buy certain kinds of things or if they engage with certain
subsections of your site and because these things are not very obvious
because they tend to be much more business oriented they’re a lot harder
for somebody to look at it and wrap their heads around but that’s where the
real value is because they drive they drive revenue that’s a really
interesting point that you made that there
applications of a IML that aren’t for predicting for dealing with what’s going
on right now I like that point and yeah mm-hmm there was a report that I was
actually sent by Forrester you know the title of it was shatter the seven myths
of machine learning and one of the things they said in there specifically
that you know correlates with us is machine learning really isn’t about
predicting the future it’s it’s at its best when the future looks like the past
as they put it in other words you’re not you’re not trying to necessarily predict
what’s going to happen you’re trying to get insight into what happened or what’s
going on right now right if if those things are going to continue more or
less unabated then there will be some predictive value there but it’s best to
not assume that they will have absolute predictive value otherwise you fall into
the trap of assuming that everything that you’ve been doing up until now is
what you what you can continue to do or what you know you should do things
differently because the Machine says so I’m just wondering you know is this one
of the reasons that it seems like there’s some a lot of a IML use that
seems to be creeping into security because you know yeah trying to it it’s
always visible there it’s always been used for there in that respect where you
will you will have some kind of pattern detection and only recently they started
to advertise it more explicitly as a nml because those are buzzwords and they
sell and at its core a lot of these things are not necessarily about the
really sophisticated forms of AR anymore like deep learning they can they can
just be a simple a simple Bayesian algorithm but yeah that’s a big part of
why as security software has started to make more noise about using using these
things because it’s a perfect use case when you’ll have you know hundreds of
thousands of user behaviors and you want to be able to just pull the one needle
out of a haystack that looks like somebody trying to break into the system
you know how do you do that you can’t do that by pulling through it manual you
have to you have to have some way to sift and so this gives you a way to set
right yeah flagging something that would otherwise take an incomprehensible
amount of time for Justin yeah this is the other pattern that I see a lot is
where you have the Machine basically being trained to serve as your copilot
or to ride shotgun with you not to take over what you’re doing but to augment it
you know we were talking about self-driving cars
before and I thought the real the real win here is not gonna be that we’ll have
a car that drives itself but that you’ll have a co-pilot that will never fall
asleep on you that will always be able to tell you Oh somebody’s about to cross
the cut in front of you or you’re backing up there’s somebody to your left
that looks like a pedestrian watch out that’s really useful and it’s useful in
the most immediate way it’s not something that’s gonna take another five
or ten years to dope out it’s already it’s already useful right now we can
already do that and I’m like that’s a really good way I think of looking at it
and to sort of distill it down to people who maybe are interested in are maybe
getting into Python and they’re interested in becoming AI AM and ml
developers but this when you first hear I am AI AM L it seems so daunting oh
yeah but when you apply it to something like that it actually it makes a lot of
sense and it makes it a lot easier to understand just to either the lay person
or the beginner right yeah I mean if you want to build something specific you
know one of the piece of advice I give to people who want to get into software
development generally is find some problem that’s close to you and work on
solutions to that problem that involve programming and it’s the same thing with
ml or I find something that’s relatively close to you it’s not doesn’t even have
to be very big and look for a way to apply machine learning to that again it
doesn’t even have to be in a huge way just in a way that makes it
incrementally easier and then as you begin to understand how that works you
can you can ride in the scope of it a little bit you know the people who work
at the Tesla what-have-you you know they have the benefit of throwing you know
hundreds of people and billions of dollars at a problem and and having
these huge ambitions the individual a developer does not have that Vantage
they have to focus on the tiny things but the tiny things may prove to be much
more immediately useful than a skunkworks project that burns through
billions of bucks and doesn’t produce anything got it we should probably take
a quick pause and see if we have any questions or comments I have a question
for Sardar but before I jackpot about use cases so here’s one kind of
philosophical maybe um how can a I progress our daily life ooh that’s
opening didn’t yeah how’s it gonna make life better for us earlier well I was
I’ve always gone with the idea that like I said it will serve as a it was a serve
as a co-pilot and that it will always it will be
something that will that will flag the things that will sometimes slip past us
and that’s not that’s not always the easiest thing to embed into our lives I
mean right now we had things like smart devices in the home and I’ve always
found them to be not very useful because they seem they don’t really seem to be
introducing new things they’re not really very new they’re just taking
something that already existed and automating and not always automating way
that’s terribly useful if I want to turn the lights off in a room I can I can
turn the lights off in a room perfectly well myself so I would I would really
want to see things that are that are not just extensions of the obvious
automating what we already do I would like to see things that are that are
more like you know genuine problems that we have that have not yet been touched
in any way by automation that’s much harder then people are asking about the
use of AI for education auto mechanics and manufacturing if you have any
insight into that that’s really broad I actually follow a blogger a fella by I
fell fell from Canada who does a lot of blogging about education and about the
about the impact of technology in education and whenever he speaks about
AI generally he always seems again to speak of it in an auxilary role that
it’s not something where you’re going to be using it to replace a teacher grading
papers but rather where you would be using it to get better insight and to
say you know how students are performing over the course of several years in a
given course or you know it it’s it’s not a way to it’s not a way to to get
rid of the teacher but a way to make his job easier or to give him insights that
he would never be able to have before that would allow him to do new things
with his job that it was simply not possible because the time that would
have been taken from to get that insight would have been so consuming and for
things like auto auto mechanics auto repair you know one of the one of the
things that I can think of office out my head is things like Diagnostics and you
know the other day I had my car in the shop because I had a punctured tire I
ran over a I ran over a nail and I said well that’s that you know that’s that’s
a very easy obvious thing but what if what if you have something
it’s not so obvious you know it’s the same way with medicine if a patient
comes in and they have a collection of symptoms that seem like they could be a
fungal infection but may in fact be something else and you’ve got this AI
assistant that has gathered millions of medical records from across decades and
can provide you with more in a wastes ways to narrow things down and ways you
haven’t thought of before that’s beautiful yeah it’s interesting it seems
like so much that AI and n ml is is related to scale the ability to take in
a huge amount of information and data that no human being could possibly go
through distill it down and then find the you know what the anachronism there
and then figure out what’s going on and then maybe you know yeah absolutely I
actually do have a question for you sir when it comes to this especially in the
case of medicine where you have this hyper personal data about people that is
also often protected at least here in the United States over by HIPAA how do
you have enterprises going about collecting this data fairly and legally
in a way that still maintains people’s privacy that’s a really tough one the a
lot of the work that I’ve seen as to where this could go revolves around the
idea of taking taking the data and operating on it not as anonymously as
possible like for instance one of the more the more out there suggestions I’ve
seen that it’s actually not quite as out there anymore over the last year or so
is what’s called homomorphic encryption where you take a set of data you encrypt
it and then instead of decrypting it to perform operations on it you perform
operations on it while it is still encrypted so you never have to actually
see the original data in order to work with it I guess what I’m trying to say
is there there plenty of technical solutions both that are developing and
that already exists to the problem but the real problem is always going to be a
social one you need to have strong laws like HIPAA to protect personal privacy
and you need also to have a sensibility among the people who develop these
solutions that personal privacy has to be paramount that I think is one of the
real shortcomings that we don’t necessarily have a culture of
consciousness about personal privacy and the development of software as a whole
and that definitely extends to the way that we would start using a I’m in
though don’t you think that’s sort of related
to the fact too that people don’t realize how much of privacy is is it’s
easy to to give it away you know what without intending it and you know it’s
interesting that sort of touches on one of the things we talked about the last
time you were here which is sort of the ethics around coding and AI and machine
learning and making sure that that you know the people who were writing the
code for this stuff have in mind you know ethical situations yeah not
against everything right that’s true yeah I mean are we seeing anything more
around that right now I’m certainly seeing a little bit more open discussion
about it like the other day when I went to one of one of Google’s products for
machine learning in the cloud there were a number of like call out paragraphs in
the documentation that were specifically aimed at the user saying you know make
sure that the day that you were that they did the data that you were
supplying is is as free of bias as possible and we’ll have as positive an
impact as possible on the people you’re trying to serve with it I said you know
a year ago I wouldn’t have seen anything like this and now they’re at least
making an attempt to bring bring the subject up in the actual documentation
for you know the the the software that’s being used to build these solutions
that’s a nice step forward it’s not the absolute you know it’s not the only step
but it’s it’s nice that there’s more casual consciousness about it and that
it’s filtering into all of these different aspects of it not just the
usage of it the construction of it sure I mean the AI has to be built by humans
and humans are have an inherent bias so it’s good that to be conscious of that
bias when you’re building something that is going to be widely applicable yeah
and that’s why you need multiple parties to check each other’s biases twice good
point do we have more questions Michelle we do
let’s do a few about careers so somebody is considering making a career into AI
in ml is that a good decision that’s like saying should I have a
career in medicine it’s very open-ended it’s it’s not a bad
idea but you have to keep in mind what what is gonna be involved you know first
as I mentioned earlier you’re going to need it I have a solid grounding in
statistics and then in math and in software development and then second
you’re going to have to figure out you know an actual project that you want to
aim at I’m not really an emirate of the idea of
just becoming a general purpose AI or ml developer in the same way that somebody
is not necessarily just a generic software developer people have the
specialization they will specialize in you know full-stack or those specialized
in databases so they will specialize in you know software that’s designed to
protect endpoints or what have you you know they will they will find a specific
problem to work on and you can start by being a generalist you can start by
getting an you know general acquaints with it but it’s always best to to have
to narrow down as you go and find something specific that you can apply
yourself to and to build solutions for if somebody was asking about if you
needed to be a programmer or if nine programmers can do it but it sounds like
you should have programming skills we’re getting to a point where it’s it’s
actually becoming easier to use some of these solutions without necessarily
having full-blown programming chops you know if you’re if you’re using Python
and you don’t really know very much you can stick pieces together and you can
achieve some fairly basic results but you’re always going to be constrained by
the limits of the pieces that you are given to work with sure and if you’re
not capable of developing your own pieces on the back end that match or
exceed the pieces that you’re given then you’re always going to be at the mercy
of whatever tool kit you’re using point of using the toolkits is if you want to
solve an immediate problem that’s great you know that’s what they’re for but for
your own progression eventually you have to learn how to transcend just just the
Lego bricks you want to learn how to basically you know start distilling your
own bricks much better metaphor really works as a metaphor but you get the idea
yeah do you know of any online courses that people could look into there’s
there’s tons of them and the bad news is that it can be very difficult to wade
through and figure out which ones are actually decent Khan Academy new to me
usually have have pretty good stuff I have I confess that I have not plowed
through the vast majority than those tours simply the two that I tend to have
the best reputation associated with them it’s not a bad idea to start start
trying them out and then abandon them if you feel that things are just going too
slowly or not being explained clearly enough okay that’s fair great
before we let you go any other thoughts about where things are going this year
since what sort of it’s January is still 20 we’re just into the new year so yeah
and and I get the impression that that there is definitely a sobering of the
way that AI is is being approached it’s no longer being seen as a magic solution
everything is no one will be seeing as this black box out of which miracles can
come that it’s a very specific tool for a very specific set of problems and that
means that it doesn’t mean that people are gonna stop trying to do moonshots
you know I don’t think that that the air has completely gone out of you know the
self-driving car idea I just think that people are starting to regard the basic
idea as being less less important early than fixing a great many other things
that aren’t necessarily visible problems you know the problem the problems of
climate change the problems of mass migration the problems of you know the
political economy of human rights says someone once put it those are all things
that involve tons of data and those are all things that I think would be very
amenable to being examined with AI and ml and that could be used to generate
real solutions with instead of just figuring out how to do the same things
that much more you know efficiently or automatically yeah that’s awesome yeah I
agree yeah great well thank you so much Sardar
thank you yeah they’re always enlightening and we have to say back
we’ll keep this going definitely thanks once again for having
me I always really enjoy it thank you yeah thank you and thank you all so much
for watching this episode of tech talk if you liked this video and you’re
watching on YouTube be sure to give it a thumbs up and subscribe to our Channel
and if you’re over on the computer world LinkedIn page also give it a thumbs up
and follow the LinkedIn the computer world LinkedIn page that’s all the words
got it thank you so much again and we’ll see you next time

One thought on “Machine learning and AI in action | TECH(talk)

Leave a Reply

Your email address will not be published. Required fields are marked *