Enter the URL of the YouTube video to download subtitles in many different formats and languages.
JASON MAYES: So to\nkick off this chapter
it's time to lay\nsome foundations
around all the\nbuzzwords you may have
You may have heard of a\nnumber of terms related
to this field of study, things\nlike artificial intelligence
machine learning or\neven deep learning.
So let's take a moment\nto formally define them
Now, artificial intelligence,\nor AI for short
is essentially the science\nof making things smart
or more formally speaking,\nhuman intelligence
However, this is\na very broad term
and right now, you're\nable to create systems
Well, narrow AI\nis simply a system
that can do one, or\nmaybe a few things
as well or better than a\nhuman expert for that task.
A great example would\nbe text classification.
As a web developer, you\nmay have at some point
been asked to make\na contact form where
That message is\nsent to a company
and someone decides what subteam\nit should be forwarded to.
Well, for advances\nin technology
you can now train a system to\nautomatically route the message
to the correct team based\non the content, if there's
Or, how about the\nmedical industry
that can use an\nAI system to scan
through grainy medical\nimages of brain
scans to accurately identify\ntumors in such images
with an accuracy that can\nexceed a professional?
This can make analysis\nfaster, more accurate
and cost-efficient\nthan ever before
which is great for both\ndoctor and patient.
Now, there's a lot\nof systems right now
that work hand in hand with\ntheir human counterparts
to create a workflow\nthat is more
So next up, you have machine\nlearning, or ML for short.
Machine learning is\nan approach to achieve
AI that was spoken about\non the previous slide.
Essentially, this is\nthe implementation
of the actual program\nthat runs and can
learn from prior\nexperience to find patterns
It can then use this\nknowledge to classify
previously unseen examples of\nthe same type in the future.
Now, what makes these machine\nlearning programs powerful
is that they can be reused\nand trained with new data
without changing\nthe code itself.
So if I create a\nmachine learning
system that recognizes cats,\nit can use the same code
without modification\nto then recognize dogs
just by feeding it different\ntraining images for it
And this is very powerful,\nand a big difference
to how you used to\nprogram in the past.
With traditional\nprogramming, you
may have had a bunch of\nconditionals or lookups
to check if a word was\nassociated with spam.
If it was, you would\nblock the email.
However, the spammer\ncan get savvy of this
modify the word just slightly,\nand the system has been broken.
Thus, a tug of war between\nspammer and programmer
develops, which is not\na good use of your time.
Fast forward to\ntoday, and you can now
use machine learning\nto solve this problem.
Instead, thousands of\nusers mark emails as spam
and the machine learning\nwill automatically figure out
what words and features are\nmost likely to have contributed
You can retrain the model\nevery day with fresh content
and now, no human\nneeds to be involved
to maintain manual\nlists, freeing up time
And there are many common use\ncases out there for machine
learning that go\nbeyond texts, too.
For example, object\nrecognition, to know
Here, you can see how many\nand where in the image
One thing to consider\nis that object detection
is subtly different\nfrom image recognition.
When someone refers to the\nterm image recognition
the machine learning\nsystem will tell you
that something\nexists in an image
but it won't tell you where\nor how many, and sometimes
misuse these terms,\nso watch out for that.
Or what about linear regression?
This sounds a little\nscary, but all it means
is that you are predicting\na numerical value
for some other\nnumerical input value
just like in your high school\nmathematics where you might use
a line of best fit through\nsome points on a graph
to make a correlation\nbetween something
on the x-axis and something\nelse on the y-axis.
For example, what's the price\nof a house if the square footage
With enough data,\nyou can predict this
with machine learning, if\nyou plot all your known data
points on the graph like\nthe ones shown here.
You can use linear regression\nto essentially find
the line of best fit\nautomatically for you
which you can then use to make\npredictions for any house size.
Or how about natural\nlanguage processing
to understand human\nlanguage itself?
With this, you can detect\nif a comment is spam
before it's even sent to\nthe server to be stored.
Or maybe you could\nunderstand the sentiment
For example, is a\nsocial media post
about a topic positive,\nnegative, or neutral?
Today, this area of research\nis so advanced, you can even
get machines to\nsummarize text for us
or answer complex questions\nfrom a passage of text
like you see on this slide,\nwhere it can scroll us
to the answer on\nany web page or even
You've even got audio-based\nML for speech recognition
turning human voice\ninto readable text.
I'm sure many of\nyou have smartphones
with digital assistants\nor tried the web speech
APIs of JavaScript\nin the browser
and this is all\npowered by machine
You have the ability\nto generate audio
too, like the demo\nlinks on this slide that
turns your voice into a musical\ninstrument of your choice
for example, which brings\nus on to the generative
or creative use cases of\nmachine learning, one of which
you can see here, which is\ncreated by Nvidia's research.
The key thing to\nnote is that none
of the faces in this\nanimation are real.
They've been dreamt up by\nthe machine learning model
just like if I asked you\nto imagine a purple cat
you probably could do so, even\nthough you've never seen one.
Here, the machine\nending has learned
the essence of what a\nhuman face is composed of
and then it's asked to generate\nnew ones that it thinks
These are just a few examples\nof what machine learning is
There are many more use\ncases out there, too
and this will continue\nto grow in the future
as industry and\nresearch continue
And by using machine\nlearning in your solution
you can reduce the amount\nof time spent programming.
Imagine you wanted to recognize\na marker pen in an image.
Sure, you could write\nsome custom code
to do this, like\nI did here, where
I try to define logical rules\nthat would allow you to find it
in a given image, maybe based\non the color and edge detection
And maybe after\nweeks of coding, you
would have something that\nworks in certain situations
but would fail in others,\nlike poor lighting
or if the pen branding\nor color changes.
Alternatively, you can\nuse machine learning
to recognize such objects\nsimply by feeding it
example images of the\nobject in question
and get a more reliable solution\nin a fraction of the time.
Secondly, it will allow you\nto customize your product
Taking our marker pen\nexample, let's say
you wanted to use this for\npencils instead of marker pens.
You could simply reuse your\nexisting machine learning
solution and feed\nit images of pencils
instead, delivering a\nsolution to your next customer
Finally, it can help you solve\nseemingly unsolvable problems
if you are using\ntraditional programming.
As a human, I can\nrecognize a face
but I have no idea how\nI actually do that.
And it's really hard\nfor me as a programmer
to translate the\nsimple tasks that I
do as a human to code-based\nlogic, yet this sort of task
is relatively simple for\na machine learning system
With machine\nlearning, you change
the way you think about\nhow you solve problems.
Traditional programmers are\ntrained to think logically
and mathematically, but\nfor machine learning
the focus shifts to\nmaking observations
about the data fed into the\nsystem, which uses statistics
behind the scenes\nthat can update
its understanding of the\nworld for the task at hand
rather than using\npure, hard-coded logic.
Deep learning is,\nessentially, one technique
you can use to implement\nthe machine learning
programs spoken about\non the previous slides.
You can think of this as one\nof many possible algorithms you
can choose from to make the\nmachine learning program
There are, of course, many\nother techniques, too.
Here, a concept known\nas deep neural networks
is shown, which essentially\nare code structures that
are arranged in many layers but\nloosely mimic how scientists
believe the human\nbrain to work, learning
patterns of patterns the\nfurther down the layers you go.
Well, imagine that the\nearly stages, a single part
of a network, can recognize\nsomething simple, like lines
Go one level deeper,\nand that line data
from the previous\nlayer may be combined
to allow you to recognize\nshapes, which ultimately
are just a collection of lines.
And one level deeper\nstill, those shapes
might combine to allow\nyou to recognize objects.
For example, a face might be\nrepresented by several shape
features that always\nappear in certain positions
Generally, the\ndeeper the network
the more advanced\npatterns it can recognize
but this comes at a cost\nof processing power.
So in summary here, you\ncan see how these three
The deep learning\nis the algorithm
you can use to drive the\nmachine learning program
and this machine\nlearning program
gives us an illusion of\nartificial intelligence
Now, these core concepts\ngo back to the 1950s.
They're not actually very\nnew, but it's only now
that you have the resources\nat cheap enough cost
such as the RAM, the CPU,\nor the graphics card
known as the GPU, to make\nthese ideas feasible
which is one reason for its\nrecent growth in this domain.
And you're living in\na truly exciting time.
It's not often to be\npart of a new industry
and you are at the\nstart of a new wave
right here for how\nyou can create smarter
In fact, machine learning could\ninfluence every industry out
And over the years,\nhuman society
has actually gone through\nmany different ages.
You might be familiar with\nthe Industrial Revolution
for example, and\ncurrently, you are
living in the scientific\nor digital revolution.
However, we are fast\napproaching what
I believe to be the machine\nlearning revolution
Notice how for each revolution,\nit's shorter than the last
but makes more human\nprogress in terms
of innovation than the\nprior one in terms of area
In fact, this next\nage could lead
to more progress and\ninnovation than all of those
Let's move on to how such\nsystems can be trained