Humans
have been taking a beating from computers lately. The 4-1 defeat of
Go grandmaster Lee Se-Dol by Google's AlphaGo artificial intelligence
(AI) is only the latest in a string of pursuits in which technology has
triumphed over humanity.
Self-driving cars are already less accident-prone than human drivers, the TV quiz show Jeopardy! is a lost cause, and in chess humans have fallen so woefully behind computers that a recent international tournament was won by a mobile phone.
There
is a real sense that this month's human vs AI Go match marks a
turning point. Go has long been held up as requiring levels of human intuition
and pattern recognition that should be beyond the powers of number-crunching
computers.
AlphaGo's
win over one of the world's best players has reignited fears over the pervasive
application of deep learning and AI in our future – fears famously expressed by
Elon Musk as "our greatest existential threat".
We
should consider AI a threat for two reasons, but there are approaches we can
take to minimise that threat.
The
first problem is that AI is often trained using a combination of logic and
heuristics, and reinforcement learning.
The
logic and heuristics part has reasonably predictable results: we program the
rules of the game or problem into the computer, as well as some human-expert
guidelines, and then use the computer's number-crunching power to think further
ahead than humans can.
This
is how the early chess programs worked. While they played ugly chess, it was
sufficient to win.
The
problem of reinforcement learning
Reinforcement
learning, on the other hand, is more opaque.
We
have the computer perform the task – playing Go, for example – repetitively. It
tweaks its strategy each time and learns the best moves from the outcomes of
its play. In order not to have to play humans
exhaustively, this is done by playing the computer against itself. AlphaGo has
played millions of games of Go – far more than any human ever has.
The
problem is the AI will explore the entire space of possible moves and
strategies in a way humans never would, and we have no insight into the methods
it will derive from that exploration.
In
the second game between Lee Se-Dol and AlphaGo, the AI made a move so
surprising – "not a human move" in the words of a commentator – that
Lee Se-Dol had to leave the room for 15 minutes to recover his composure.
This
is a characteristic of machine learning. The machine is not constrained by
human experience or expectations.
Until
we see an AI do the utterly unexpected, we don't even realise that we had a
limited view of the possibilities. AIs move effortlessly beyond the limits of
human imagination.
In
real-world applications, the scope for AI surprises is much wider. A
stock-trading AI, for example, will re-invent every single method known to us
for maximising return on investment. It will find several that are not yet
known to us.
Unfortunately,
many methods for maximising stock returns – bid support, co-ordinated trading,
and so on – are regarded as illegal and unethical price manipulation.
How
do you prevent an AI from using such methods when you don't actually know what
its methods are? Especially when the method it's using, while unethical, may be
undiscovered by human traders – literally, unknown to humankind?
It's
farcical to think that we will be able to predict or manage the worst-case
behaviour of AIs when we can't actually imagine their probable behaviour.
The
problem of ethics
This
leads us to the second problem. Even quite simple AIs will need to behave ethically
and morally, if only to keep their operators out of jail.
Unfortunately,
ethics and morality are not reducible to heuristics or rules.
Source:
phys.org
No comments:
Post a Comment