1. AI: Neural network
We have brain, so we can think
But AI don’t have brain.
Do you know how AI can think?
The neural network is a mathematical model of our
neurons in the brain. And it is a fundamental
concept on which AI is based!
The neural network have changed during the three
AI booms.
<The first AI boom「Perceptron」>
The perceptron is a neural network that mimics
human vision and brain function.
<The second AI boom 「Multilayer perceptron」>
The multilayer perceptron has three nodes. Other
than that, not much different from the perceptron.
<The therd AI boom 「Deep learning」>
The deep learning is very famous right now. It
allows neural networks to capture features
themselves. So, It can learn more active.
<Perceptron>
The perceptron, a neural network invented in 1957, attracted attention for its ability
to mimic human vision
and brain function, leading to the first AI boom.
However, the boom came to an end
when its weaknesses were pointed out,
such as its inability to learn
problems that are not linearly separable.
<Multilayer perceptron>
The second boom occurred in 1986 with the development of the "error back
propagation" method. The Error Back Propagation Method uses a "multi-layer
perceptron", a neural network that is an advanced version of the perceptron, allowing
for more complex learning.
However, at that time,
before the advent of the Internet,
the boom came to an end again
because there was not much data
available for machine learning,
and the learning accuracy
of multilayer neural networks did not improve easily.
<Deep learning>
A technology called "auto encoder," developed in 2006, made it possible for neural
networks to capture features themselves. The learning method using autoencoders
and multilayer neural networks is called "deep learning," and has been the
breakthrough of the third AI boom.
2. AI: CNN
Hook
What do you think this
picture is of?
Can a computer recognize it
as you do?
Key details
・What is CNN?
CNN is "Convolutional Neural Network,”. It
is a neural network with several deep
layers and is a network that generates
value mainly in the field of image
recognition. CNNs are mainly valuable in
the image recognition task called "general
object recognition. In this field, it is
utilized as an algorithm with excellent
performance.
More detail
1,Convolution: By restricting the connections between
neurons well and using a technique called weight sharing,
a process like image convolution is expressed within the
framework of a neural network.
2,Pooling: Reduces the size of the feature map by
extracting the features that satisfy the conditions
from the specified region.
3,Affine: Instead of treating the input as a two-dimensional
array, we will treat everything equally as a single list. All
values get their own vote in terms of whether the current
image is ○ or x, but we are not completely democratic in
this process.
In understanding how CNN works, there are three
main layers: Convolution, Pooling, and Affine.
3. Artificial Intelligence: Reinforcement learning
Hook
A learning method that
automatically derives the optimal
solution
Key details
Reinforcement learning repeats trial and error and learns the
appropriate control method.
Application Examples of Reinforcement learning are Advanced
Control, Autonomous Driving, Robotics, Scheduling and
Calibration.
Reinforcement learning being used in a real system for example,
Shogi AI, Igo AI and Cleaning robot. There are 5 steps.
・First step is about an environment. This step defines the
environment. Second step is about a reward. This step is
necessary to repeat some times for reward. Third step is making
agent. This step is necessary to choose some methods for
example, choosing neural networks or lookup tables, choosing
an appropriate learning algorithm. Fourth step is agent training
and validation. This step prove the performance of the learned
policy. Final step is policy deployment. All of them are
Reinforcement learning steps.
More detail
・Advanced Control: Controlling nonlinear systems is a difficult problem, often
addressed by linearizing the system at different operating points. Reinforcement
learning can be applied directly to nonlinear systems.
・Autonomous Driving: Given the success of deep neural networks in imaging
applications, driving decision-making based on camera input is an area where
reinforcement learning is well suited.
・Robotics: Reinforcement learning is useful for applications like robotic
grasping, such as teaching robotic arms how to use different objects in pick-and-
place applications. There are many other applications of robotics, such as human-
robot and robot-robot collaboration.
・Scheduling: Scheduling problems can be found in many situations, such as
controlling traffic lights or coordinating factory floor resources for a given
purpose. Reinforcement learning can be used as an other to evolutionary
methods for solving these combinatorial optimization problems.
・Calibration: Applications that involve manual calibration of parameters, such as
electronic control unit (ECU) calibration, are good candidates for reinforcement
learning.
・Neural networks: It is a mathematical model called an artificial neuron that
expresses nerve cells in the human brain and their connections.
・Lookup tables: Refers to data structures such as arrays and associative arrays
created to improve efficiency by changing complex calculation processing with
simple array reference processing.
4. Artificial intelligence: sentence generation
Hook
AI may win the Nobel Prize
in Literature in the near
future !?
Key details
GPT-3 is the most notable current technology in sentence
generation AI.
GPT-3 (Generative Pretrained Transformer) is one of the pre-
trained natural language processing.
Natural language processing is a technology that uses AI to
analyze large amounts of text data.
GPT-3 learns a massive corpus of text with more than 175
billion parameters.
Thanks to that, GPT-3 can automatically generates AI-written
text that has the potential to be practically indistinguishable
from human-written sentences, paragraphs, articles, short
stories, dialogue, lyrics, and more.
In the U.S. and other countries, there are a number of services
that utilize GPT-3.
More details
● What is GPT-3?
GPT-3, a new language model from the whizzes over at OpenAI, generates AI-written text that has the
potential to be practically indistinguishable from human-written sentences, paragraphs, articles, short
stories, dialogue, lyrics, and more.
OpenAI trained GPT-3 on a massive corpus of text with more than 175 billion (yes, with a "b")
parameters, making it the largest language model—ever.
Lost? In non- technical terms, GPT-3 was shown how millions of people write and taught how to pick
up on writing patterns based on user entry. You just feed it some input, and the model will generate
intelligent text following the submitted pattern and structure.
● How does GPT-3 work?
GPT-3 is a language model, which is a statistical program that predicts the probable sequence of
words. Trained on a massive dataset (from sources like Common Crawl, Wikipedia, and more), GPT-3
has seen millions of conversations and can calculate which word (or even character) should come next
in relation to the words around it.
When you type in an initial set of words, like "go to the store to buy...", GPT-3 will start predicting
what would naturally come next based on its training. Probably something like:
Eggs,Milk,Bread,Fruit,Vegetables,Snacks,Drinks,Etc.
● Downsides to GPT-3
GPT-3 may seem like the perfect AI-communications solution, but it's not without its imperfections.
There are a few downsides to this powerful machine learning technology:
1, Lack of true intelligence: GPT-3 is a deep learning model that uses machine learning algorithms,
but it's still not "intelligence." This AI is only using existing text to predict future results—it's not
necessarily coming up with anything truly original as it lacks true understanding and meaning (unlike
something like Artificial General Intelligence (AGI)).
2, Privacy risk: It's unclear whether GPT-3 retains any portion of the training data, making it a
potential privacy issue.
3, Bias: GPT-3 can be fooled into creating incorrect, racist, sexist, and biased content that's devoid of
common sense and real-world sensibility. The model’s output is dependent on its input: garbage in,
garbage out.
5. Artificial Intelligence: Meta AI
Hook
There is a small director, who is
not humans, in games!
Key details
Meta AI is the type of Artificial Intelligence. It is used
in games to adjust the difficulty of games. In detail, it
gets some information about the context of games, the
level of games, and so on. After that, it adjust difficulty
based on these information. So, they work like a
game director.
There are two types; Classical Meta AI and Modern
Meta AI.
Crassical Meta AI has been existing since the
1980. It judge the skill of player, and then adjust
difficulty passively.
Modern Meta AI has been existing since the
2007. It adjust difficulty more actively by using more
particular way. For example, by the spawn position,
the number, and the kind of enemy, or more
sophisticated way.
More details
There are two example for each Meta AI.
First, the use of Classical Meta AI in Xevious. Xevious is a 2-D shooting
game. It was released in 1982. In this game, the game director use a
humorous variable and easy Meta AI. The name of the variable is
“DIFFICULTY”. It is used for judging the skill of player. If a player take
more mistakes, the value of “DIFFICULTY” will be lower. And Meta AI
adjust the strongness of enemy based in the value of “DIFFICULTY”.
Thanks for these components, both good player and not good player
can enjoy this game.
Second, the use of Modern Meta AI in Left 4 Dead. Left 4 Dead is a 3-
D First Person Shooting (FPS) game. It was released in 2008. The goal is
to leave the city that there are a lot of zombies. The players cooperate
to achieve this goal. In this game, the very complex way is used. At
first, the AI estimates the “emotional strongness” of each survivor as a
value. For example, when they are injured by zombie, the value
become higher. The AI adjust the amount of zombie based in the value
of “emotional strongness”. If the value does not reach the peak, the
amounts of zombies and strongness of zombies will be higher. If the
value reach the peak, the AI will give the relax time to the players.
Thanks for these complex methods, this game will not be simplified,
and the player more likely to continue this game.
6. Edge AI
Hook
Doesn’t self-driving car really run over human?
Key details
Self-driving car can detect human and stop immediately using ”Edge
AI”.
The machine which equips “Edge AI” can decide immediately.
“Edge AI” can think immediately because it is mounted on devices
such as IoT devices and sensors.
“Edge AI” processes inferences in the device based on data which is
collected at the device and decide immediately.
In contrast to Edge AI, there are Cloud AI.
The machine which equips “Cloud AI” cannot decide immediately.
It takes time for “Cloud AI” to recognize image or information.
Hence, If self-driving car uses “Cloud AI” cannot detect human and
stop immediately.
More details
The advantage and disadvantage of “Edge AI” are shown below. First, advantage of
“Edge AI” are low communication costs and security strength. “Edge AI” only sends the
data which is necessary for learning to the cloud after processing on the device.
Because the volume of data is smaller than that of cloud AI, which sends all data in the
device, ”Edge AI” can reduce communication costs. Then, “Edge AI”, which perform
inference and learning in different locations(device and cloud), can process critical data
which has the risk of information leakage while keeping it inside the device. “Edge AI”
has the advantage, which enhance the security of the network environment by
processing without going through the internet. Then, disadvantage of “Edge AI” are low
processing capacity and high hurdles for introduction and operation. Considering the
size of the device and its power consumption, there is a limit to the resources that can
be loaded in the device. Because the CPU and GPU used in ”Edge AI” have lower
specifications than those used in the cloud, it is difficult to process large data. In
addition, because inference and learning are performed in different locations, “Edge
AI” cannot execute complex and advanced processing. Then, ”Edge AI” has high
operational hurdles because making system design and maintenance operations are
easy to get more complex.
Give you some example of “Edge AI” except self-driving car. First, it is monitoring
manufacturing facility. The sensor which load ”Edge AI” enable equipment in the
factory to monitor in real time. It can detect minor changes in operating conditions and
prevent equipment from breaking and stopping. Second, it is behavior of customer
analysis. The camera which loads “Edge AI” and is installed in stores and facilities
analyze data of customer behavior. By analyzing the path which a person has walked, it
can serve a changing layout of product and reviewing of product shelves. Finally, it is
safety management of construction site. The gateway which loads ”Edge AI” provides
operation information of construction equipment and special vehicle s for
construction. By monitoring construction works in real time, it serve a safety
management of laborer.
[1] https://conexio-iot.jp/blog/40#6178bc8b23748342cb2e1a84-1635302878375