N
N
Neogener2017-02-21 19:00:37
Machine learning
Neogener, 2017-02-21 19:00:37

What is the evidence that the machine is learning, and not collecting new data and comparing it with the already loaded?

There is a story about a neural network that analyzed frames from a video shoot and identified a cat. I heard it from Serbrant from Yandex. He called it ii. But the cats were initially loaded - photos with cats. And why is the comparison of photos and analysis of similarity with existing ones, this is training?
So what is machine learning? Without terms, in simple words, what is the logic?
For example, as a human - a biological being, I define cats by comparing them with previous cats and only in this way, and only by external signs, because I was told from childhood that "they are just like that." The rest of my training took place in the same way, they showed options, they said what it was, and then I determined the rest by similarity - in fact, they loaded the base and parameters for me, that is , I didn’t invent anything myself.
It turns out that machine learning is not learning at all, as it is commonly perceived, but all the good old database of objects with parameters that are based on when analyzing other similar ones (with a very slight difference in details )?
If you compare software and AI, then it's the same thing, but AI just sells better than "new more complex program"?

Answer the question

In order to leave comments, you need to log in

5 answer(s)
I
ipswitch, 2017-02-21
@Neogener

I, being an opponent of the concept "Aaaa, AI is coming, skynet is coming soon, aaa !!!", I always informally explain to students this way:
1. A "neural network" is just such an advanced " filter ", in fact - a sieve with a flexible grid size . Its task most often is classification , selection, separation of conditional "grains" from "chaff".
This is no "artificial intelligence"!
2. For such a " selection " we need to somehow set the criteria , how one differs from the other. Since the world is not perfect, and the grains are not perfect, and the sieve is not perfect, we also need an " error"- response threshold, threshold, tolerance, backlash, allowing you to skip "non-standard, but still grain."
3. The process of selecting coefficients (" weights " ) is essentially fine tuning, calibration , fitting this "sieve". mathematical algorithms are used , essentially statistical, probabilistic, which one way or another average the array of possible data supplied to the input in order to obtain output coefficients at which this "sieve" will work with an acceptable (evaluable, known, calculated or required for a given task ) accuracy.
That's what "learning" is all about. In fact, we are writing an algorithm that, based on the measured spread of input data variations, gives us such coefficients at which the sieve will pass most of the necessary "grains" and not pass most of the husks.
Replace these words with others. Not "training", but "selection of coefficients", not "thinks", but "filters", not "neural network", but "fuzzy classifier", and much will become simpler and clearer.

V
VoidVolker, 2017-02-21
@VoidVolker

In short, yes. Everything is exactly the way it is. And in fact - there is no real AI here anywhere in fact. Stupid imitation and beautiful words for marketing. On the other hand, what is learning? The process of acquiring new knowledge? From outside? Analysis of available?

N
nirvimel, 2017-02-21
@nirvimel

What is a cat made of?
- From the head, torso, paws and tail.
What is the chair on which the cat sits made of?
- From the seat, legs and back.
What is the picture of a cat?
- From millions of pixels (more precisely, from millions of digital values ​​that determine the color of each pixel).
What is the image of the chair made of?
- From millions of pixels, the same ones that make up the image of a cat, only arranged differently.
How to distinguish a picture of a cat from a picture of an empty chair?
- (Probably) We need to compare the pixel values ​​in different images.
And specifically what pixels to compare with what? By what formula?
- I don't know. Not a single person knows and cannot know such a formula that from a million numerical arguments (pixel color values) would produce one value at the output, a simple answer: is this picture a cat or a chair.
But how then does the human brain solve this problem?
- The human brain works on fuzzy logic . Therefore, it may seem to a person that images of familiar objects are hidden in the clouds and blots of Rorschach (false positive response of the neural network). The neurons of the brain are initially interconnected chaotically, and this structure is not able to give a definite answer about the belonging of various images to the same category. But in the course of learningthis structure is rebuilt in such a way that the answers to the question about whether the image belongs to a certain category each time (learning step) better correspond to reality. See What is the learning principle of neural networks based on?. In the process of recognition itself, there is no purposeful recall or search for information in memory (this would take too much time). The brain of an adult gives the answer "this is (not) a cat" as quickly as the brain of a newborn, only the accuracy of these answers and the percentage of errors differ. This accuracy is determined only by the structure of connections between neurons in the network, and not by the number of different cats that a person can consciously remember. Evidence of this is the fact that people who have experienced memory loss do not change the ability to recognize simple images, despite the fact that they are not even able to remember their own name.

S
Sergey, 2017-02-21
@begemot_sun

AI, learning - akin to an algorithm, customization.
Any computer runs an algorithm, so no matter what is inside, neural networks or something else. If the algorithm is able to accumulate and generalize (read self-tuning), then this is a type of AI.
Linear regression is also a private level of AI.
And all sorts of approximations with interpolations too.
And the methods for solving systems of linear and differential equations are pure AI.
Oh yes, any Kalman filtering and control theory - AI is the thing.
Only no one would think of calling it AI.
These are all algorithms.

X
xmoonlight, 2017-02-21
@xmoonlight

Training - installation of "weights" in the nodes of the neural network according to predetermined input and output parameters.
Example, learning (child):
show a box (bar) (input) -> say "this is a brick, like a cube, only elongated" (output ) a match is a cube, a feature difference is formed and a new branch from the "cube" node appears again.
We make a comparison: how much does a brick differ from other objects?
Let's say the result of comparison for the coincidence of contour lines and sound information is:
1. brick vs ball: <50%
2. brick vs cube: >=50%, (BUT! STRICTLY LESS THAN 100%)
That is, we see that the brick - not a ball at all, but more like a "non-cube") and in this regard,
Was:
1st level (1=100%): 0.5(ball), 0.5(dice).
Now:
1st level (1=100%): ball: 0.5, cube: 0.5
2nd level (0.5=100%): brick: 0.5
(3rd: 0.25=100%, etc.
) weights from any of the "nodes" of the 1st level to any "leaf" in the "tree" - always equal to one!
In this way, we gradually build a relationship tree.
This is how learning happens somehow: reference reference features (learning), comparison with those available from the knowledge "tree" and adding new knowledge to the desired branch of the tree.
The connections themselves are tied to time with a gradual subtraction of the weights of the tree nodes to zero and we get the effect of forgetting (or "release" of nodes = memory) if the information is not periodically updated. When updating information, we update the entire branch to the sheet with the maximum weight: " repetition is the mother of learning ."

Didn't find what you were looking for?

Ask your question

Ask a Question

731 491 924 answers to any question