Introduction to Machine Learning

Video Script

Hello everyone, welcome to the online lecture series of artificial intelligence and the
data analytics e-learning module. This is the introduction to machine learning and your
course instructor Dr Bo Yang at the Department of Computing of the Hong Kong
Polytechnic University. Right, here is our first question: Why do we study machine
learning? Bill Gates said that a breakthrough in machine learning would be worth 10
Microsoft's. The former President of Stanford said that machine learning is the hot new
thing. Well, it seems that machine learning is something new and very valuable. Believe
it or not, machine learning is something that has been changing the world. Okay in
today's lecture we are going to briefly introduce machine learning in the following four
aspects. First what is machine learning? Right here we would better first look at what
is learning. From Merriam Webster, we can see that learning is defined as the activity
or process of gaining knowledge or skill by studying, practicing, being taught or
experiencing something. Actually, for human beings, from the time when we grew up
from kids to teenagers to adult, it is actually the learning process. Okay then what is
machine learning. From the computer scientist Tom Michelle, a computer programmer,
is said to learn from experience E with respect to some class of tasks T and performance
measure P. If its performance at tasks in T, as measured by P, improves with experience
E. Well, it seems very convoluted and very abstract to understand but you don't need to
understand. You don't need to memorize this definition at all. To be simple, machine
learning is something or is a piece of program that can learn from experience and then
to solve some tasks. That's the very simple way to understand what is machine learning.
Okay now let's look at the brick, the history of machine learning. Actually, machine
learning is not something new and it is an old thing. It was invented in 50 years ago
actually. In the 1950s, the first perception algorithm was invented and implemented as
a circuit, not computer yet, and then after years of progress, different types of programs
were invented. In the 1980s, some fundamental ideas like back propagation were
invented, which actually those ideas are the fundamentals of the knowledge of machine
learning techniques. Well then in the 1980s, actually machine learning and AI faced a
winter time. It's not always a good time, because at that time the computer games and
computer programs were not able to solve difficult problems as human expected,
making not so many people believed it anymore. But after 10 or 20 years, in the recent
10 to 20 years, especially back to the 5 to 6 years ago, in defeated the human GO
championship. This was broadcasted, and as I remember broadcasted globally, so
nobody believed that this program, this piece of program can defeat and can solve so
much complex task but it was successful. So now we are in the era of the big wave of
machine learning. Machine learning was not new and was invented, developed for
many years, almost half of century. Why, because it's not simple. Machine learning
involves many different disciplines, for example the mathematics including probability,
statistics, optimization and also some Biology, Psychology and some neuroscience.
Well, the good news is you don't have to be an expert in all these areas and then start to
learn machine learning. It's not necessary. So, you have the ability to learn machine
learning right now. Okay now let's talk about some types of, basic types of machine
learning. So generally, there are three types of machine learning - Supervised learning,
unsupervised learning and reinforcement learning. The first supervised learning, this
model's algorithms are generally trained with targets or human labels to solve some
classification tasks or regression tasks. For unsupervised learning, basically those
models are usually trained without targets. They just train them with raw images or raw
data without any annotations. Then these models can solve some tasks, for example like
clustering or dimension reduction. The last category of machine learning models is
reinforcement learning. Usually, the reinforcement learning models are trained with
mistakes, for example like the algorithms are designed to play computer games, so there
is no correct or wrong answer. There are mistakes, if you win something and then you
give rewards and then your model will be better and better, so that is the strategy of
training with mistakes. Okay here is a very simple example for supervised learning,
particularly for image classification task. So, we can see that here we have many images.
Well here i just show four images but in practice we have many images together with
labels. These labels or targets are allocated by human beings. Okay like for example we
have this picture, we know it's cat and we give it a label, and then we give the label Bird
to this image. Actually, we take many images and then we allocate these images with a
categorical label. So, the supervised learning is like we define some models and then
we feed these images together with these labels to train the model. Okay, to train, to
teach the model, to learn some important features, for example car, there are some
wheels, like this a cat they have some eyes or ears, something we hope that this model
can learn these features. So, after we train this model, it's like we have successfully
taught someone, and then we give the model a new image. We call the test image and
then there's more. Take this test image into the algorithm and then the algorithm can
predict the category of this image. For example, in this case again, the model we
compare, we think about that: Okay the appearance or the geometry is similar to this
image. Now this image, test image belongs to car. It seems that this model can learn
from this experience and the data and then they can predict. Similarly, there's another
example for supervised learning particularly for regression tasks. So here we again we
have training labels. There are several guys like with different genders and the body
height and the age and so on, so we have the label of their body weight. And then we
train the model after which optimizes this model, and then we give a new guy, the
information of new guy - Male, 1.7 meters and 42 years old and then this model is able
to predict. Okay this guy is likely around 60 something kilograms, something like that.
It's like they learned this knowledge from the data and the labels, the model is able to
predict something. Here's the difference between this regression task with the previous
slide, is that here we use continuous label like the numbers - they are continuous, but
the previous one is categorical label - it's discrete. So here is another example for
unsupervised learning. So again, we have a lot of images but this time there is no label
associated with any image. We only have lots of images and probably now you know
that it is very easy to get many millions of images from online. So, what is unsupervised
learning? Basically, we define some models, we learn features from these images and
then we can cluster the similar images into one group. For example, the first one is this
one - the two dogs. They look similar from the eyes, noses or ears something like that.
When the other guy is panda, these two pandas they look similar. The model - the
unsupervised learning models or the algorithm, they are able to automatically group the
similar images into the same cluster. They can also reduce the dimension of the data
point. For example, we know this cat image in the background is not informative. The
key features are the noses or eyes and the other features. We can learn a more
discriminative vector or a representation for this image. In that case, instead of storing
this high-resolution image in the computer, we can just store a very compact
representation and then transfer it online. Still the information is not lost, the key
information is still there. All right, of course there are many other tasks like density
estimation and so on. The answer was learning. They can all solve these tasks. But by
the way, we don't need any labels here. There is no label at all. We don't know these
two images belonging to the same category. The model will learn to put them together.
Right, so here is another example of, or actually an illustration of reinforcement
learning. Reinforcement learning is like this - so we have an actor or agent or a just a
piece of a program, a piece of algorithm. This actor is able to sense the state of the
environment. For example, when we play computer games and we look the status, for
example we can visually take images of the environment. Given this state information,
this actor will take some actions by itself, like for example you go left or go right or go
up or go down to shoot some enemies or something like that. Right that the actor will
make some actions to change the environment and then the environment will give some
rewards back. For example, you shoot many stars and then you will get higher scores
back. In that way the state you can get an updated state simultaneously so the actor will
try to get more reward by doing some more sensible actions. So, in this way the actor
is trained with reward. Another is the mistake - of course reward is that if it is positive,
we call it reward; if it's negative, that is a penalty or mistake. And then according to the
reward, the actor will try to adjust its own strategies, its own actions so that the reward
can be maximized. Okay in this way, the way the actor evolves and learns by itself from
the environment and interacts with the environment, this kind of algorithms we call
them a reinforcement learning algorithm. So now we have introduced three types of
algorithms - Supervised learning for classification and the regression tasks,
unsupervised learning for clustering or some other tasks, and reinforcement learning
for some like computer games. Now let's look at some real applications. Okay so here
we can see that given a video, this is a video okay as clip of a video, so we designed
some supervised algorithms. That algorithm is able to precisely detect all the objects
like including the cars, pedestrians and all the other guys here, like so many crowds
here, the people here. All of them can be detected and even tracked. So, in the
autonomous driving applications, this is a very important, the essential part, otherwise
the car will hit people or will go to the wrong directions. But once the car with this
machine learning algorithms embedded, those cars can make right decisions, make
quick decisions and to avoid any traffic or any dangers. So here is another example for
unsupervised learning. We can see that on our left-hand side, this is a video clip of a
horse. Right so we change the appearance of this horse to a zebra, so we can see that
the texture or the appearance of the horse is almost very perfectly changed such that the
right-hand side, this kind of a video it looks very realistic but in fact it is a fake video.
In fact, if you have a zebra video, you can convert it back to a horse video as well. So,
the point here is that we have the horse video but actually we don't know how it looks
like if this guy is a zebra. We don't know, we have no idea. But the algorithm can
generate such a realistic and very visually appealing of any videos, so without any label,
the algorithm can generate very realistic videos that is amazing. Okay so here is another
example of reinforcement learning. Actually, this guy in the video was my previous
colleagues, and so at the left-hand side we have the camera mounted on this robot. So,
this color video and this steps video is the observation or the states of the environment,
here the environment is this room. We have the sequence of images. There are boxes or
some other things. But then in the robot we have a reinforcement learning algorithm
embedded, so this robot itself can sense in the environment and make decisions to
navigate itself to avoid the obstacles, for example here you can see that, even the
environment changes, it is able to make the right decisions to avoid any collusions. You
can see that it's not just a repeat. It's not programmed, it's not hard programmed. This
agent, this robot has to be able to understand the environment by itself. That is because
of the smart reinforcement algorithm embedded. So, this is the three demos for the three
types of machine learning algorithms. Hope this could be inspiring for you guys and of
course there are many amazing applications and demos online if you search or if you
follow some newspapers. This now machine learning can do many things, can solve
many real-world problems. Okay so here is the summary of our brief introduction to
machine learning. Alright, the first three points. Machine learning aims to solve tasks
from experience and improve its performance. We can see that we have different types
of algorithms, like to detect cars, pedestrians, to navigate or to generate some new
videos from experience. Experience is collected from the data, for example we have
added in many data and then this algorithm can learn from our annotations. For the
reinforcement learning, we can teach the robot to go around. If the robot hits the wall
and then we will get to give it a penalty and then the robot memorizes, figure out that
okay in the next time i will not hit the wall, I will go right hand to go to the correct way,
otherwise i will crash, so that is the solved task from the experience, and then improve
the performance. Performance means okay you can correct or protect something. You
can generate some more visual opinions or they can survive longer and in the different
complex environment, like for navigation and so on, so that is the machine learning,
the very general concept of machine learning. It's not something very difficult. The
second point is machine learning generally has three types of models - Supervised
learning, unsupervised learning and reinforcement learning. Well of course there are
many different environments of this algorithm or learning problems, but anyway these
three are the general categories. Last one is machine learning has a wide range of real-
world applications, including computer vision, natural language processing, robotics,
etc. Of course, there are many applications, you can even include the chemistry or
biological study or some many areas. Nowadays machine learning is widely used in
different areas. Okay right so these are the three points to summarize our today's lecture.
The closing mark is why do we study machine learning? Yes, it is changing the world
and i hope you catch this and from now on you have this opportunity to study machine
learning and you will have the opportunity to change the world. All right, thank you for
attending this course, and I hope you have a good journey to study machine learning.
Thank you and bye.