Lecture 8 – Connecting Brains and Machines

Overview – How can we connect Brains with Machines?

Part 1 – Artificial Intelligence; Part 2 – Neuroprosthetics

Artificial Intelligence

What is Artificial Intelligence?

According to Google:

Artificial intelligence is the theory and development of computer systems able to perform

tasks normally requiring human intelligence, such as visual perception, speech recognition,

decision-making and language translation.

PSYC4982 2016 Definition:

Artificial intelligence is technology that mimics biological intelligence.

A Brief History of AI

Going back as far as we can remember – in the

Greek myths there are concepts of AI in Talos – a

bronze statue that comes alive and terrorized

people. Galatea was a sculpture carved from ivory,

and the sculptor was so fascinated by Galatea that

he prayed to Aphrodite to bring Galatea to life.

And Aphrodite obliged. Once you have a concept

of AI, you can have a concept of an artificially

intelligent being.

Aristotle was fundamental in creating the

foundations of logic. In particular, he identified

logical arguments known as syllogism, and the

way this works is that you have a couple of

premises, and based on some logical rules you are

able to form a conclusion.

This is the foundation of a lot of approaches in AI.

Skip the Dark Ages and go straight to the

Renaissance. René Descartes had the concept of

dualism. Nowadays we reject dualism and think

that the mind is instantiated in the brain. But an

AI approach relies on there being some truth to

dualism: if intelligence depends on the brain, you

can’t have an artificially intelligent computer. If

we want to mimic the brain we have to do it in

something other than the body, and as far as AI is

concerned, the mind is separate from the brain.

Whitehead and Russell wrote Principia

Mathematica, with the aim of creating

foundations of mathematical theory

and pure logic. The idea was that you

have a series of axioms from which

you could derive all possible truths on

the basis of those axioms. As an

indication of the complexity of the

language used to construct this system,

the equations may not seem

particularly meaningful to the layman

– but even with this complexity we still

have yet to arrive at the basic facts of

logic (or in this case, 1 + 1 = 2).

Their dream was crushed in 1931 by a logician

called Gödel. This isn’t quite correct, but it gives an

indication of Gödel’s contribution. He showed that

if you had any formal logical system, you’re unable

to prove all of the truths. He showed that there were

true statements within the logical system that

couldn’t be shown to be true based on the axioms of

the system – there were provable truths, and

unprovable truths. You can’t create all of the truths

through the logical system. In the same token, you

could say that there are falsehoods that cannot be proven to be false. So he’s trying to indicate

the failure of logical systems – we need a new approach (to be discussed in this lecture).

3 laws of robotics. If we were going to have

intelligent beings, we should have some

constraints on them, and make sure that they

don’t overtake the world and kill us. These

laws were meant to prevent the robot uprising.

But he also did write a series of books in which

you could circumvent these laws. Maybe you

can’t prevent the robot uprising though these

kinds of logical or law-based approaches.

The first computer was invented following the

end of WWII. It was as big as a room, and could

perform many complex scientific operations, but

it was very unstable and parts had to be replaced

by the hour. We’ve really come a long way since

then. The computers in our phones are a lot more

powerful than the very first computer.

In 1950, the Turing test was developed by Alan

Turing (who helped to end WWII), and it was

proposed to test machine intelligence. A

machine could be said to be intelligent if it could

convince a human that the machine was in fact a

human. So you’d have conversations over the

computer, and if you couldn’t tell if the “person”

on the other end of the line was a human or a

computer, then the machine passes the Turing

test. To this day, no machine has passed this test.

AI and Gaming – Machine Victories

Just had a broad overview of the AI history. Now we take a more specific pathway in games.

There have been several advances in the last few years, and it also provides a nice way for us

to test and develop new methods of artificial intelligence.

The computer was programmed to be able to play noughts

and crosses to the expert level. In 1952 this was a big deal,

but nowadays anyone with a level of programming expertise

will be able to solve this problem.

There wasn’t a lot of progress until the 90’s because of the

complexity of the gains that were involved, and also the

requirements of computing power and the subsequent costs.

In 1992 a computer could play Backgammon at the human

expert level.

Draughts was conquered in 1994.

In 1997, IBM’s Deep Blue machine beat the world

champion in chess – a really big deal at the time. You can

see in the TV the way that Deep Blue performed – it searched

through all of the possible game states into the future given a

particular move – e.g. if you move a Knight to a particular

position, what consequence would it have in the next turn,

and in corresponding turns. This problem was solved through

a brute force

method.

Jeopardy was conquered in 2011. This didn’t

really need to wait until 2011 because you

could just gather

all the facts and

arrange them in

such a way that

the computer

can access them.

Most recent developments have used deep

neural networks (Google DeepMind). In

2014 they published a study showing that the

computer can perform at an expert level at an

ATARI video game – a video game system

of the 1980’s – with classics like pacman and

asteroids etc. Their work was published in

Nature. They combined aspects of

neuroscience and psychology – they

combined reinforcement learning (from

instrumental conditioning Skinner’s

approach), and they used deep neural

networks, inspired by the way that the

human brain is constructed.

The results are shown below. Y-axis are

the games that the computer played. X-

axis is how well the machines performs

relative to human

levels. 100% = human

level of performance,

and anything above

that means that the

machine is doing

better than the human.

The computer was

able to play space

invaders a lot better

than a human can.

The turret could take

out the mothership on

most occasions.

The most recent case was in 2016, where the

deep learning approach was able to defeat the

world champion Lee Sedol in the game of Go.

The paper was published in Nature.

The game of Go involves a board of 19x19

positions. 2 players, one with black stones

and the other white. Each player takes a

turn to place a stone at an intersection on

the board. The goal of the game is to

surround the opposition stones with your

stones. When you do that, you can remove

their stone from the board and that counts

as a point for you. So you’re trying to gain

territory by surrounding your opposition’s

stones. The main aim is to gain as much

territory as possible.

So what is it about Go that makes this

result so interesting? Chess was conquered

in 1997. Why did it take until 2016 to

master Go?

Both games came from roughly the

same area, Indochina, with Go being

older than chess. The board in Go is

bigger than chess, so there is an

increase in the number of positions.

Remember that even a small increase in

the board size drastically increases the

complexity of the game. There are more

possible moves per turn in Go, and the

number of board configurations in Go is

downright ridiculous.

The number of board configurations in Go is astronomically bigger than in chess – the ratio

of the difference is greater than the number of atoms in the known universe. That’s insane.

This is a huge problem for both humans and machines to solve.