Naked Science Forum
Non Life Sciences => Technology => Topic started by: evan_au on 30/03/2016 07:54:06
-
In this week's podcast on Artificial Intelligence, the terms "Algorithm (http://en.wikipedia.org/wiki/Algorithm)" and "Artificial Intelligence (http://en.wikipedia.org/wiki/Artificial_intelligence)" were sometimes used almost interchangeably.
What is the difference between an "Algorithm" and "Artificial Intelligence"?
Is it possible to implement an "Algorithm" using "Artificial Intelligence"?
Is it possible to implement "Artificial Intelligence" using an "Algorithm"?
What do you think?
-
I think artificial intelligence is part of algorithm, hence algorithm has broader meaning.
I can use algorithm to count from 1 to 1000, or calculate greatest common divisor, but that's hardly called artificial intelligence.
There are some requirements for artificial intelligence, such as mimicking behaviors of intelligent beings.
-
The definition of algorithm is still a subject of academic debate (https://en.wikipedia.org/wiki/Algorithm_characterizations#1968.2C_1973_Knuth.27s_characterization).
I like Knuth's definition, which can be paraphrased: an algorithm is a predefined sequence of instructions (possibly including decisions and loops) which converts an input into the output in a finite (and reasonable) number of steps.
While the number of possible games of chess or Go is finite, it is huge (ie not reasonable), so a brute-force search to find the winning moves for these games cannot be completed by current computers within billions of years.
However, I think you could extend the notion of algorithm from one which produces the "right" answer to one that produces an "approximate" answer; solving the "traveling salesman" problem for 100 cities is not reasonable, but getting an answer which is within (say) 1% of the "right" answer has greater economic value (as is a chess program that doesn't select the "right" move, but selects a "pretty good" move in a few seconds).
Artificial Intelligence covers many fields and many techniques, so it is hard to generalize:
- Game-playing machines must produce answers within the maximum time dictated by the rules (ie reasonable). The moves don't need to be "right", just better than the other player. So these AI applications could be implemented by an "approximate" algorithm.
- A lot of AI work is done today by Artificial Neural Networks, modeled on the operation of living neurons. In principle, these can produce answers in a finite time, using highly parallel hardware (like the retina of the eye). Given a neural network and a set of weights, this could be implemented as an algorithm.
- However, when you come to self-learning networks (eg for handwriting recognition, facial recognition or playing Go), the results depend on the input training set, and its sequence of presentation. And it can evolve over time, with experience. I suggest that this is beyond the scope of an algorithm, since a finite set of steps has not been defined before you start, but is developed "on the fly".
I've heard it said that Intelligence is anything that computers can't do (yet).
By this criterion, AI can never be implemented as an algorithm... [B)]
-
An algorithm is a set of instructions. AI implies a system that can modify its algorithms in response to learned inputs rather than "givens".
The output of an algorithm will not surprise its author, who could have reached the same conclusion "manually". The distinguishing feature of intelligence is the ability to surprise the author.
You can write an algorithm that predicts the response of a stone hit by a stick. You can't write an algorithm for the response of a dog hit by a stick - the dog will have taken account of your size, the size of the stick, the known strength of its teeth, and a whole bunch of other environmental inputs, and developed its own responses, including a second-order evaluation of the power of surprise.
-
The brain's firmware does AI differently from machines. The brain uses both spatial and casual memory and logic. Casual logic is 2-D or based on cause and affect. Casual memory is memory that has a logical basis; common sense. Spatial logic is 3-D and could be defined as cause, affect, cause and/or affect, cause and affect.
Creativity, extrapolates known cause and affect; 2-D logic, to create others underlying causes and extrapolated affects. If I take a tree leaf and let it fall, it gently glides and swings back and forth downward; cause and affect. The phenomena has cause and affect by which we may be able to identify a leaf type. Instead of just a correlation, science will to try to explain the underlying cause of this cause and affect. AI may not be able to do this just yet, since this uses a unique schema.
In terms of a visual, 3-D logic can be seen as analogous to a 3-D golf ball. We can approximate this 3-D golf ball with a large number of 2-D planes/circles, at different angles, all with a common center. A 3-D or spatial memory is composed of a large number of logic planes and memories with a common theme.
As example, there are a wide range of opinions and religions connected to the concept of god. Each opinion can be seem as one of the rational planes, with all these opinions using a common center; god. Each opinion comes from a different angle, with the sum of all 2-D logic circles, equal to the whole truth about the subject; 3-D.
All these opinions are needed for humans to express 3-D. The brain does this naturally. The open minded person tends to listen to all sides, consciously, since conscious 3-D allows a unique form of intelligence to emerge; creativity.
Picture our 3-D logic ball, composed of a wide range of logic planes, each with the same center, but all at different angles. I hit the 3-D golf ball with a golf club; whack!. The ball deforms in 3-D, causing logic planes to distort out of their plane. This distortion in 3-D is the intuitive gap where an idea might appear to defy 2-D logic. It defies the 2-D logic of any logic plane, but it nevertheless follows natural logic laws in terms of 3-D.
(https://www.thenakedscientists.com/forum/proxy.php?request=http%3A%2F%2Fwww.maniacworld.com%2Fgolf-ball-hitting-steel.jpg&hash=517044f920957a6d10a6d51158a913f4)
AI could be enhanced with 3-D memory and logic. When any stimulus is inputted, the AI begins with the 2-D logic to get a stock answer. If there is no stock answer, the computer swings the golf club for a 3-D distortion, that hopefully allows the ball to end in the hole or get close to the hole. This allows an extrapolation of logic planes, adding new premises. Some of these may be useful and others will not. You go back to 2-D and decide what can be added to 2-D and then to 3-D.
I generate a lot of ideas this way. I output or write the 2-D logic planes based on underlying 3D logic. I may say different things on the same subject, each time I write, because I am tweaking many logic planes to help advance the 3-D ball. I am more covered with spatial integration. It is not as easy to write 3-D since it involves a z-axis, that needs to be experienced; gut feeling. Archimedes's eureka moment started at the z-axis; gut feeling and excitement, that and ended in a logic plane; hole in one.
I will have to share my thought dimensionality theory, which can correlate how our memory can be organized and interpreted by the brain from 0-D to 4-D. The 4-D memory is the 3-D memory with a time element. For example, falling in love integrates humans with love. These dynamics will last for a finite duration as it extrapolates to humans to a future state.
In terms of 3-D logic, and 4-D logic, sometimes the first whack at the ball causes it to go into the rough; nothing useful appears. But since there are 18 holes, one can come on strong at the end since this helps to align the centers differently. This form of AI will make computers scary, since you may never see it coming.
-
Technically artificial intelligence is normally what is called a 'heuristic', which means it's not guaranteed to work; much like humans, whereas an algorithm is defined to always do what it's supposed to.
There is a little bit of confusion possible here though, since a program to perform artificial intelligence is executing an algorithm that is always going to run the AI successfully, but the resulting AI is not guaranteed to do what you want.