Transcripts
1. 1. Introduction: Hello everyone. I'm pretty sure you
have seen and read the war and AI artificial
intelligence repeatedly. And most likely,
you already a user, all the various AI enabled
tools in your life. Now, do you wonder how all
this works on a high level? And would you be
interested to know more, how is your job and your life will transform the way they die. In the next five years. Does AI have any
limitations on boundaries? And finally, where are we headed with this rapidly
evolving technique? If you have any of
those questions, then please sign
up to the course. I'll give you a brief history
of AI, how the gut here. I'll give you various
examples from different industries and how
they are currently using AI. And I'll try to demystify
some of the myths around AI, what you can and what you cannot expect from the I within
the next five years. And finally, where
are we headed as we collect more data and develop more advanced
AI algorithms. By the end of this course, you'll be able to use
these insights to plan ahead for your life
and your own career. My name is said, I studied both natural and
artificial intelligence in Cambridge University
and then I took my tea, also co-founder
of cognitively t, which is an AI company
dedicated to improving health. Thanks for watching and I'll hopefully see
you in the class.
2. 2. Project: Great. So as for the course project, I want you to make a video
of yourself describing a tool or an
application that you're currently using and
is powered by AI. And I want you to describe in your video how you think AI is currently using that product. And maybe try to mention tree top features or highlights that are
particularly powered by AI. And if you can kind of
try to predict or try to understand how these features
would or would have not existed without a UI that
that would be added value. Then finally, tried to see how this application or
tool that you are describing, well change in the
next five years as day I is going
to advance more. So based on what you will
hear in the next lessons, this is going to help you
to kind of get an idea of where we're headed with
new advancements in AI. So full, that will help you
to predict what's going to happen in the next
five years. Great. Then finally, when you are done, please, please
post your projects into the project gallery.
3. 3. AI TimeLine & History : Alright, so today is our first session
and I'm going to walk you through a
quick history of the eye from where roughly started and where we
are today and kind of awareness possibly
going in the future. It's going to be
a quick timeline. I think it's fair to say that
Alan Turing, back in 1951, he was possibly among the, one of the early individuals
who had dreams about AI. And he had imaginations and I'm going to read
out the code from him. He said that at some stage, therefore these
machines, we should expect them to take control. So that was kind of
pace imagination back, back then in 1951. Of course, if things didn't
happen that quickly, it was much more difficult apparently than the human
generations out initially. So in failures after that. Back in 1957. There were these initial
variances of neural networks. In simple terms,
neural networks are these big networks
of connections. And you give them input, you multiply them a set of weights and you
get an altitude. This is one layer of
those neural networks. This one layer neural networks. They were called
perceptron and they could solve only linear problems. So they got very
popular from 1957 to 1960 to Frank Rosenblatt. He was among the visuals main individuals behind
introducing this perceptrons. And after 1960 to the excitement around these
networks started to go down. Particularly because
they could only solve linear problems and they
were heavy to train. So there weren't enough or
strong processing units to be able to train
these networks. And there weren't big
enough data sets. So those were two main
limitations around, or three main limitations
around preceptors. Then a few years later,
another hallmark perhaps, was the publication of this, we'll call it analog VLSI, implementation of
neural systems. We could perhaps mark
that as another hallmark. But Prof. more excitingly in 1997 was this big IBM
machine called Deep Blue. That one, Kasparov in chess. So that was a historic
move in 1997. And As we will get into this later in
one of the sessions, I'll just highlighted here. This big, big machine. It was the main
advantage it had, or humans was having a
very huge big memory. So it could kind of based on whatever move
that Kasparov hat, it would have made a
very big decision tree and kind of predicting what
would be the next best move. If Kasparov comes with
this one or that one. So that would kinda predict
all the possible ways until the end of the game in
a very big decision tree. And that's kind of how the
algorithm, we'll call it. We could call it a
smart algorithm works. The main advantage therefore, was having a very
big memory and being able to kinda build
that decision tree. Now, moving forward in 2000s, something exciting came round. It's just called graphic
processing units. And it wasn't advancement
in the hardware. And the graphic processing units allow for parallel
processing of an operation. Massive parallel
processing, one operation. This was actually something that could very much help
with neural networks. Now we'll get to that in, in 20122010 around there. Okay, so in 2 thousand, therefore we had this
introduction of GPUs. As I said earlier, one of the limitations of the early neural networks,
the perceptrons. In addition to hardware, was lack of data, big data, big labeled data. So in 2009 and around 20092010, universities,
institutes, et cetera, they started coming up with
these big labeled datasets. Perhaps the most popular
ones by now is the ImageNet. The world of visual processing, visual object recognition. This ImageNet competition
that had existed at MIT. It's a very big dataset, millions of images
that are labeled in terms of what objects in them. I mean, this, this
year is we have many more labeled datasets. So that was kind of
prompts, another hallmark. Now, in 2010, we have those GPUs and big
labeled datasets. So what happens next? In 2012, there is this
ImageNet competition and a new neural network which pupil now refer to as deep learning algorithms
was introduced by Alex Krizhevsky in sales team in that competition and it won all the other
algorithms by distance. So to put it simply, it was kind of a
multilayer perceptron. Now because both GPUs
and big labeled datasets allowed or compensated for the limitations of
those perceptrons. So it kind of became
possible price, of course, they also had some innovations
and novelties too. Make the algorithm work. Fewer parameters wasn't
densely connected. Right there in 2014. Our grew up on a few others. We showed there are strong
similarities between how humans process visual images and deep neural networks do. The more similar the
models of object provision become to humans to actually perform
better in real world. In 2016. It's not a big hallmark. And that is when AlphaGo, some of you might have heard of google DeepMind in London. They built this AI algorithm, which was based on deep learning and reinforcement learning. And it could play
the game of Go, which is much more difficult if you'd like
compress the chest. Because in chess, as I said, you could predict or built this decision tree budding
in the game of Go. That is not possible. Like effectively, you can have a decision to have
older possible moves. Because it's could say it is infinite number
of possible moves. Very different in terms of how the game of golf
worse compared to chess. What you need here is
kind of an intuition, rather than having a big memory. How they did it in simple terms. The deep neural
network was trained by the the games that
champions of the Go play. And by looking at this is
the scene of the game. Now, this is the next best
move that this champion did. You train the algorithm
with such inflammation? And in 2016, the, the algorithm developed when
the champion of the game. And one exciting thing which
perhaps is it related to what the codes I said
at the beginning from Alan Turing was that in one of the versions
of the algorithm, you have to AI models that
compete against each other. And one of them gets better and better by playing with
these other AI model. Today, we have self-driving cars and we are moving to now. Many applications around us do benefit from some
sort of AI algorithm. Either it is face
recognition, self-driving, speech recognition
and translation, Google search and
so on and so forth. Smart advertisements,
personalized recommendations. And we will try to mention
some of those and give you more examples
throughout the course. This stage that we are in is also referred
to as industry 4, which is the combination of IoT, Internet of Things,
artificial intelligence. Iot is responsible
for collecting data and making the appearances and things connected
to each other. And then Ai processor and gives you additional insights, right? And moving towards the future. Stay with me. In the next few courses, you will see what is likely to come within the next
five to ten years. I don't think we can accurately
predict much beyond that, but we'll see what, how possibly AI will
replace some of the jobs. What would be the
potential impact on your career and on our life?
4. 4. Machine learning and Good-Old-Fashion AI: Today we're going to learn about machine-learning and
good old-fashioned AI, also referred to as goal five. So what is machine learning? So machine learning definition
is essentially a fill of a study that gives computers the ability to learn without
being explicitly programmed. So essentially we are
teaching machines to learn to do things without
explicitly programming them. And in that history I gave, in the previous session, we saw some of the games that machine-learning techniques
and AI techniques have been able to conquer, such as chairs back in nineties. And then recently
the game of Go. So this is kind of a
timeline of AI and also some of the
high-level definition of what is artificial
intelligence, what is machine learning? What is deep learning? I'm sure you've heard all these different
terminologies and if you wonder, which refers to what
then does exactly that. So the term artificial
intelligence, that's the broader umbrella. It's any technique, any
machine-learning technique, or any algorithm that enables machines to
mimic human behavior. It could be envisioned, it could mean, speaking, could be in any other domain. This is generally referred to
as artificial intelligence. So whatever application
that falls under, this is called AI application. Then more specifically,
they are a subcategory of algorithms that are referred
to as machine learning. Terminology started
at around 1980s. So that's when machines
learn effectively. And they're also this session we will talk about the different branches
of machine learning. So a subcategory of that
is called deep learning, which is based on
neural networks. So of course these
are overlapping. So if we want to
separate the two, we have this machine-learning
techniques that are not based on neural networks, such as simple classifiers,
regression, etc. And we do have Raul
network-based machine learning. And when the depth of
that neural network is more than a few layers than they are called
deep learning. Today, I'll walk you through the broader categories
of machine learning. And then the next session, when we talk about modern AI, I'll give you a little bit of information on high-level
understanding of what is deep learning and some
applications. Right? So what are the different types of learnings for machines? And actually it's not just for
machines, even for humans, we have supervised learning, unsupervised learning, and
reinforcement learning. Supervised learning
is effectively when we tell the agent here
either the algorithm or the human being or even
training animals. Then you give them a stimulus, an object, an image, and you tell them what it is. So it's supervised learning. And you repeat that
process over and over again until you make sure that
they actually understood. This is supervised learning. Unsupervised learning
is effectively you get just give the
images or stimulus. You don't give the
labels or annotations. So they may not exactly
know what it is, but based on the features of those images or stimuli
you've given them, they'll be able to cluster them and separate
the two like say, okay, these are apples
and oranges, right? Maybe I don't know the label, the name of that fruit. But I know these look similar to each other and these others
look similar to each other. These two are different. Clustering algorithms fall under the
unsupervised learning. Then we also have this concept
of reinforcement learning, which I'll talk about
more in the next session. But broadly speaking, in
reinforcement learning is the ability to learn
by exploration. You put an agent
into an environment. And by exploring
the environment and finding out the boundaries
of that environment, they learn to do things. And human kids are a very good
example for this scenario. Babies, they don't
know how to walk, they don't know how
to respond or react. Or essentially, no, very
little when they are born. And by exploring
the environment and being naturally exposed
to different things. Receiving a reward or a penalty. By that exploration. There are things that they
charge and they might be hard, so they'll avoid it next time. That's kind of a
natural penalty. Or they might eat something and they find
it very delicious, so they will
continue to do that. So this is kind of
reinforcement learning. Bike that exploration,
they come to some rewards or penalties
and then based on that, they decide to repeat that
action or not to repeat it. That's kind of how
the learning happens in the world of
reinforcement learning. Alright, so now I'm going to walk you through
two examples here. Two key subcategories
of learning, supervised learning and
unsupervised learning. So supervised learning,
as I mentioned, we give the model. It could be any model. It could be a good, old-fashioned AI model,
such as a classifier, or it could be a neural network. So this is regardless
of the model, this type of learning is
called supervised learning. We show a input images or input data here,
series of apples. Then we tell the model the
labels are annotations. We tell them a little cases. Look these pictures, these are apples and these
are the labels. So we give a few
repetitions to make sure the model has understood
and learned the input data. Then we in the test phase. So we have trained the model. The model has learned
the concept of apples. So we show an apple. And then what do we
expect them all to say? Well, we expect them all
to say it's an apple. That's called
supervised learning. Now, a very good example
of supervised learning, which I'm sure you are
familiar with is Face ID. This is smartphone. And these are some
of the sensors in the front side of
the, of the iPhone. And these are the instruments for you to input the
data essentially, right? So it will read your face data
and let me play the video. So here, this is the
training phase, right? You will see the iPhone essentially will
look at your face from different perspectives. And it will understand
your geometry of her face. And it will build a model of
how your face looks like. So this is essentially giving
the data and labeling it, so it will train the model. Now, what happens next
is that you can see in this video through that sensors
in front of the camera. It's, we'll look at the face
is building that model now. And then after that, in the test phase, which is after you've
set up your face ID, then you look up
your mobile phone, you will see whether it is
going to unlock or not. If it has recognized your face
than it will on like here, the idea is that it should
be able to unlock your face. Regardless of the changes in your everyday fashion or
changes you might have, even your growing periods. So this is, this is
kind of the idea. It builds a face
model that adopts to changes into your face. This is called invariant
face recognition. So regardless of the
changes into your face, as long as the changes
our identity preserving, then the Face ID
should be able to recognize your face, right? So that was an example
of supervised learning that now what is
unsupervised learning? So this is a set of fruits. And we don't have the
labels in this example, so we're not telling the model that these are bananas
or these are apples. We just give the
images to the model. And the model will
be able to separate them based on how
similar they are. The apples visually
looks similar, and then the bananas pitch, so they are into three
different clusters. So clustering algorithms are an example of
unsupervised learning. So you're not, you haven't
given them the label. You have not given
them all of the label based on the features that
is important to the model. In this case, the
visual similarity. The clusters will be formed. Example of that is
your smart oh boo. So for example, here, if you go for example,
to you or to all boom, you will see that based on the photos you've
taken previously, iPhone has categorized for you. People are clustered,
pupil or different faces. And even if you haven't
labeled them, right, in this example, the
individual has labels, but you don't necessarily
have to label them. So what happens is
that the face ID, the face is detected, and similar faces are
categorized into one cluster. And then you can either choose to add a label to that or not. It's up to you. If you're interested to know more behind the
scenes, what happens? This is what happens
in a given photo. The, the algorithm phase detects the faces and the upper body
and then matches the two. This is, this face
belongs to this body. And then it goes, this part goes to her face model and this one goes
through a body model. So the part that I wanted to emphasize here is this
clustering parts. The faces that look
similar to each other are clustered
into one category, then you can choose to
label it or not, right? So that's, that was
another real-world example of unsupervised learning.
5. 5. Modern AI: This session kind of
talk through modern AI, particularly on topics about deep learning and
reinforcement learning. So we briefly discussed
what is traditionally I. And here are a few
more examples. Expert systems, the
michele neural networks, which we mentioned in
the history of AI. Fuzzy logic is a good example when some of these languages, A-I, A-I specific
languages, probably. These were some of the
older advancements in AI. We then had some advancements in machine learning
and computer vision. We mentioned the IBM Deep Blue. The big benefits of the Deep Blue was
having a big memory, being able to build a search tree of what are
the possible actions. So I'll do the game
and therefore be able to give us cores and predict
what would be the best next, next move for the computer. So moving to modern AI, we can mark this primarily with two advancement
in hardware and GPUs. And then big labeled datasets
such as the ImageNet. So deep neural networks, they are neural networks that
have a depth more than one. That's kinda technically
referred to them. The deeper the network. Typically you have
more parameters, you do larger training datasets. And there'll be
more non-linearity that the algorithm goes through. B can also mentioned the Bayesian statistics and Bayesian algorithms as one
of the modern tools and AI. My focus will be
mostly on giving you a better understanding of what is a deep
neural network. What are the operations
involve? Kinda high level. And then a good example of that also would be
the game of Go. The game of Go is one of the
most complex strategy games. Perhaps treat thousand year-old, which one of these board game? And in comparison
to chest, chest. After the first two moves, around like 5400,
possible next moves. So it's easy to build
all the possible moves and assign a score
to each of them. But in Go, there are close to 130 thousand
possible moves and the search is based, it's pretty last at. The number will be greater than the atoms in the universe. So it's not possible to predict all the
possible future moves. And that's where they kind of
intuition comes, comes in. And so in this case,
the, the AlphaGo, which was the
algorithm created by DeepMind that wound the
champion of the game. It was based on deep
reinforcement learning and kind of learn by
observing and learning the policy and
assigning scores into possible next
future future moves without building all
the moves, right? So, yeah, what is a deep
convolutional neural networks? So this is kinda the first layer of deep neural nets
were found books. This is the input image. And there'll be
operations like the stem. You can see here they
are called convolution. So these are filters that are
convolved over the image. And then you get an output. So different filters
applied throughout the image and you build
different feature maps. And these feature maps go
through a nonlinear operation. Most popular one is the
operation that you see here, rectified linear function. And then they go through
a local pooling, meaning that in that, in this blue window, everything is pulled together, close is mapped into one value over the next layer or supplier. And then they go through
a local normalization. So this is the first layer, and then each layer has operations very
similar to this one. And you can have. Several layers are originally the the AlexNet 20128 layers. I'll show a figure
of that as well. But before that, this is kind of how that could look like. So this is the input image and these are the filters
applied each letter. Then they can out with a one layer becomes the
input to the next layer. And you have a series of
layers that are considered feature learning or
feature extractors. They map the input image or
input data into a feature. And then you haven't for you. Typically fully connected layers that are doing the task
of classification. So they get the feature
map and then they map it into your final labels. These finer layers or come
fully connected because each node here is connected to all the
nodes in the next one. So that, that's why they're
called fully-connected. Alright, so I thought this
might be interesting for you. This is the original drawing
of the 2 thousand paper. The similar kind of
deep neural network, a model that want the
ImageNet competition. And that's how this new wave of deep neural networks started. This was trained with supervised learning using
1.2 million images. The output here
is 1000 category. So it can predict or classify thousand different
image categories. And in terms of the
number of parameters, there were 60 million
parameters in total, and 650 thousand
neurons are nodes. Right? And as I said, are,
well seven layers. Or you could consider
with the input, there could be a flyers. So that's the 2000s
Krizhevsky network. Now, we talked briefly about
reinforcement learning. I thought it would
be a good idea to give you a high-level
example here. This is something you have
possibly seen many times. What happens here? The, the agent here, the dog, and the environment here, the girl that is
throwing this thick. This is a typical easier to understand case of
reinforcement when the agent follows and goes
to pick up the stick, and after it does
that, it's rewarded. So it's, the agent
is observing what, what's happening in
the environment. And then based on the observation of getting
rewards or getting penalize, it will repeat the
actions that lead to more rewards
and less penalty. That's called
reinforcement learning. In the context of algorithms and the concrete
isn't that different? So here is a Atari game that he's using deep
reinforcement learning. The concept is very similar to what I just described in terms
of reinforcement learning. I'll run the video so you can
see generally what it is. So this is the game. So this isn't a target
and you have most likely played when you were younger. So these are the results of the Atari game after just
ten minutes of training. So you see that the game itself. So how does the training work? It's very simple. The only thing that the
algorithm knows is the input, which is what you
can also see on the screen and the score. So the algorithm essentially
just moves around these small plates and it either gets randomly
rewarded or penalize. After every move. It recalculates it's understanding
of their environments and learns to repeat the actions that
lead to that reward. Gaining a score or
missing your score. Avoid it will avoid the
activities or actions that caused a penalty
missing this score. So yeah, now, that was
after ten minutes. Now after 120
minutes of training, you can see it's very
pro tem play very well. And that's an example of
reinforcement learning. So there is no other
training going into this. The curious no like pre-specified
algorithm or anything else is just a policy. And the agent here is
learning how to do this by just doing
it and getting feedback that's improving itself over time by being in the environment and
trying things out. Cope. So we talked about
reinforcement learning, we talked about deep learning. And in this example, this is a deep
reinforcement learning. And it's the only part
that is different from a typical reinforcement
value here is that in terms of
predicting that score, that the reward, that penalty. The algorithm here is using a different element
for it to do. To do that, predicts
what is the score of each and for
the water removed that the algorithm
is going to make. And that is using a
deep neural network. Otherwise, it's the same as any other reinforcement
learning algorithm. Cool.
6. 6. AI in Healthcare: In health care and applications
of AI in healthcare. So this is a topic that
I'm most passionate about. Of course. I will start with this grade
book, Deep Medicine. From Eric. There is a coat and the book
says, What's wrong game. Today's healthcare system
is those missing care. And the way it
relates to that is by utilizing artificial
intelligence that might actually look
counter-intuitive, but actually, if bringing AI into our killer Nicole and
medical practice, good measures in place, there's a good chance that
the clinicians can spend more time to be more humanly involved in the process
of carrying the patients. And that's how, that's what
the theme of the book, and that's what potential
AI can bring in health care in the near future. So in terms of the timeline and you know roughly where VR, in terms of AI being
adopted in health care. You are aware of this for
industrial revolutions. And I think in the
first session I briefly also refer to this. It's fair to say that medicine or healthcare
is roughly here. We are in yet fully into the fourth industrial
revolution in healthcare. It's partly because
of regulations and some of the other
frictions exist in the health care. And digital transformations. Take you a little bit
more time in this sector. And it is definitely more sensitive because human
lives are involved. Now, what I want to go
through these few minutes is for different scenarios in which AI can be
utilized in healthcare. You might have already seen
some of these applications. One is in improving infrastructure and
access to health care. Which leads to reducing costs of health care,
improves quality. It makes health care more
accessible and more affordable. It's kinda makes it
scalable like many people, regardless of their location, language, etc, can
access healthcare. One such good example, I would say our chatbots in medical care and
medical advice. Babylon GP, the GP, a tanh has such a
chatbot, already. Integrate it so we can
consolidate the chat, but that gives you
some initial diagnosis if you like, or, you know, gives you an idea of when
you ask us to conditions, it gives you an idea of like what could be potentially wrong. And then it gets kinda connect
you with a health care professional, doctor, GP, etc. Now, that's kind of an infrastructure that
brings health care accessible to a wide variety of people regardless
of their location. So that's, I think
a good example, but there's still a
lot to be done on improving access to health care through these digital
transformations. So another scenario, which is reasonably obvious and
you might've heard the news, some of the
applications is use of AI and machine learning
in diagnostics. Or eight for diagnosis will
give you two examples here. One is cognitive,
cognitive ability. We have developed this tool, which is an AI powered tool for detecting
cognitive impairments. And it uses ai. Explainable AI. Another example are, there are
a variety of applications. This is one example
that uses images, your ultrasound images, to
detect signs of breast cancer. There are other AI driven algorithms that
work, for example, on images of brain
as well to detect tumor and a few other
such applications. So diagnostic, I would say, is one of the areas that
AI is more adopted in. And he's kinda more intuitively,
if you know what I mean, diagnostics eight for diagnosis are getting a momentum
at the moment. So third example is using AI and machine learning for
prevention and monitoring. This is also getting reasonably popular, particularly
with variables. So we have variables, we have, if we're collecting more data from an individual over time. Apple HealthKit is a
good example of that. It's collecting your exercise, sleep, other activities. Also. It's tracking your heart rate. It can do ECG monitoring. So putting all this
data together, it can give you insights about your life is still your
cardiovascular health. And if you apply those insights, you can potentially detect signs of cardiac potential
cardiovascular problems. You take those early so we can prevent or improve
your life is still. Another example for fellas is the optima and it's
a wellness app. Objectively measures your
everyday kind of performance, again, against your
lifestyle measures, and you can use it to
improve your life is still that's broadly on
prevention, on monitoring. And we'd AI, that's actually one of the trends
for healthcare, is moving towards
prevention rather than late disease detection, which is more costly. Now, finally, the fourth dimension
or scenario I wanted to bring
here is treatment. In comparison to the
other three I mentioned, treatments is I would say
behind. There are examples. I would pick this
tree. Drug discovery. He's an obvious choice. Using AI and machine learning. Pharma companies can
accelerate the process of discovering new drugs based
on already approved drugs. So kind of narrowing down
the list of drugs that are likely to work on a new disorder and running go clinical trial on a limited or narrow down
list of potential drugs. And instead of
running many clinical trials on the wider list, which is much more expensive
and takes more time. Robotic surgery is another area that is capable of helping with. Digital therapeutics is also a term is used
recently, more often. And it can refer to some
of the digital treatments. You can you can be given
at prescriptions, etc. It's pretty early stages still, but there are some applications
of it already existing. Some of those, for example, or games that could help you improve some of
your mental status. And some of those games, healthy to improving ADHD. If you Google, you'll find some of those have
already FDA approval. So this is a field
that is of course evolving and it has very limited applications
at the moment. But there is a good chance
that it gets momentum within the next five to ten years. Yeah, if you will like to know
more about other stories, other potential scenarios
that might happen. With AI in healthcare. This is a good book
I would recommend. It's an e-book. Hope you will enjoy it. I'll see you in the next lesson.
7. 7. AI, Society and Jobs: In this session, I'm
going to talk about AI and its impact on the
society and the future of jobs. So let's start with this. We know AI, internet of things to sell technologies
are enabled. And if we look at them from this perspective in health care, which is the example we went
through in the last session. We see that the same services
we're getting today, they can be provided
more effectively via EI. They can become more
widely accessible, more affordable, can be
given with higher-quality. Now, this is true in education, finances, health care, etc. And I'm gonna give you two
examples from education. And then in retail. In education. Imagine
the schools of tomorrow. Not tomorrow, maybe
today we have seen tasted a little
bit of this already, but maybe not as part of our
formal educational systems. So imagine we have
these new technologies. You could be educated
anytime your comfort and your educational
platform will adapt to the needs you have,
the student needs. The curriculum would be
customized towards the career. You're after annual
ritual mentors. The exams will be customized
based on what you want to achieve and the whole platform
can be gave you fight. In retail. Maybe this is something we
are more familiar, Beth. I've seen recommendation engines like things like an
Amazon and Google, where you get personalized
recommendations based on your interests and your history. You might have already
seen chatbots that provide kind of customer
support 24 hour. The example of chatbots, I also mentioned in the previous session
how they can help in health care provide a high
level recommendations. Right? So moving into, these are few examples of how AI can be used in different
days, different services, and provide those three
features we mentioned, making them more accessible
to a wider population, making them more affordable, and also giving them
a higher-quality. So does I provide
opportunities for developing countries
or is it only useful for developed countries? We know maybe US and China
are currently leading AI. But the AI has lots of uncharted territories
and there's a lot of room for growth and that many countries
can benefit from. And I'll give you example of
a concept called leap-frog, which is how developing economies can actually
use technology and discuss AI to jump a few steps forward and maybe
catch up with the game. Now, imagine this
black icon here. This is a developed economy. Previously they had
to go every step, build infrastructure
step-by-step until it gets up to here. Where you have a
developed economy with all these services
in health care, finance as education, etc. Now, imagine if you are
a developing economy. Now, you can use technology. Here. This is a jump
jumping equipment. You can use it to jump and skip a few steps and jump over here and effectively
day I doing that. But take the example
of education. When you provide this platform, remote communication and
online support and automation. There are lots of services across different
different industries, including healthcare, education, finances that can be
provided over it, the same infrastructure
you have built only once. So instead of going
into a country, building lots of hospitals, lots of schools than
maintaining them. Here you're building
one infrastructure for these mobile platforms
and mobile communications. And through that way, that platform you are
providing all these services. Leap frog. Now the question I
get asked many times is how ai can affect
jobs in the future? Are we going to go
with chocolates? So, well, this is, this is a good visualization of how automation and the future, near future would affect each of these
different industries. The reference is always C, D. And you see on the right, this is a probability of
automation by sector. So on top, food preparation
assistance cleaner is helpers, et cetera. Those jobs have the highest probability
of being automated. So there'll be less humanly
kind of involved if you like. Because there'll be automated. Maybe they don't
need much of it. Creative work, or they lack
some of the aspects of unique human capabilities
such as taking care, etc, which I'll talk to you about possibly in the
next, next, next session. So these are kind of the
jobs that are more likely to be displaced or replaced
by AI or machines. So whereas as you go down
the list here, there are, these are the jobs such as teaching professions,
health professions. So these are the jobs
that engage, both need, they need more creativity
and also that aspect of human to human
communication, social care. Those aspects are strong in it. These are the aspects that
are strong in humans, but less sewing machines. Therefore, the chances
of automation or less. But the overall, this Fourth
Industrial Revolution is not that different from
previous revolutions. And what happened previously
is likely to happen. And this is making this
prediction that yes, they'll be jobs displaced. But more likely we will create more new jobs than the jobs
that will be displaced.
8. 8. Summary and Conclusions: Alright, so this is
our last session. We are going to have a quick summary of what
we learned together. Initially, we talked about
this timeline of AI, how AI was evolved, and how we got here. And importantly, we discussed ai as an enabler, transforming
different industries. And this is like the
impact we can have. It can make things more widely accessible,
affordable, higher-quality. We discussed this
in the context of health care as an
example of postal, gave examples of
education, retail, and you can generalize
that other industries. We also talked about
limitations of AI and some of the unique human
characters. In particular. I hope you take this v to u as this two-dimensional
diagram where compassion and creativity
and a strategy. These are unique
human characters are characters in
which humans are strong compared to things
that AI is just try and grab, like optimization and things
that can be automated. We also discussed
how I can change the current jobs and
the future jobs. In particular, we
discussed that AI is not going to destroy all the jobs and we
become old job list. Instead. Similar to other industrial
revolutions in the past. There'll be some jobs
that they'll be displaced and new jobs will emerge. In particular jobs that are easier to automate,
they'll be displaced. And the jobs that are neither
one of those human aspects, more, they'll emerge,
they will remain. So we'll have new jobs created in those particular
two-dimensions. I mentioned. With that, what I
wanted to emphasize the importance of life long
learning and education, particularly on those
core humans skills. And so we need to become
better at strategic thinking, creativity and the softer skills human to human
interaction connections, compassionate, being
compassionate individuals. So these are the skills
that we can work on. Discriminates and
distinguish ourselves from AI, if you like. Right? So thank you all for being
with me until this far. You hope we have
been able to shed some light on some of
the myths around AI. And hopefully that
will contribute to some insights for you
to see how we are currently using AI and
how AI is possibly going to change how we live
and work in the future. And I hope you can benefit
from that by planning ahead. Have a fantastic day and hope to see you again
in future courses.
9. Bonus: Future of Healthcare with AI (Precision Medicine): Here we want to walk
you through the impact that artificial intelligence
can have on health care. To do that, let's
just start with locating where health care crumbly East with regard to
adopting new technologies. As you see here in this figure, healthcare is yet
behind in adopting AI. There are various
reasons for that, which we'll talk about
in a different session. But for now, this
presents itself both as an opportunity because there
is a lot that can be done, but also a challenge. Now a question you can ask is, if we adopt a UI, where does that get us? So it will get us
from the realm of intuitive medicine and I'm pretty cold medicine
to precision medicine. So in intuitive medicine, that's when the clinician
uses their intuition. So it's not based on data
or in protocol evidence. Based on intuition. They might suggest diagnosis
or a treatment plan. In M protocol medicine, the clinician uses
the limited data available to them and perhaps a clinical interview with the patient looking at
some images, et cetera. And based on that,
they come up with a possible diagnosis and a
probable treatment plan. And maybe in six months, they'll see the patient again to see if the truth and
that hasn't worked. So it's kind of experimenting
with the patient to see what happens then if
the treatment with park. For ours in precision medicine, based on the data, a wealth of data that can be collected from a patient
over time and variables, why those signs, etc. The clinician, with the help
of that data and then I, making sense of that data
can make precise diagnosis, precise personalized diagnosis, and then treatment will follow. So let me give you
an analogy here. When you buy a car, the manufacturer of
the car will guarantee that the car will be
functional for 34 years. So they do that based
on the sets and t they have about the quality of the car
they have manufactured. In in health care settings. When you go to a hospital, can you expect them to guarantee a diagnosis or treatment plan
they give you? No. That's because there is
so much uncertainty. The realm of
intuitive and impro, cold medicine that the
outcome cannot be guaranteed. But moving to precision
medicine, the help of AI, what we can expect
is that we move from this uncertainty to the
realm of certainty. And under such scenarios
when you can make it precise diagnosis and
a precise treatment plan, then even health care
services can be guaranteed.
Seyed Khaligh-Razavi, AI & Entrepreneurship (AM Cambridge Uni)
