sound waveform
Since it can only convey one-dimensional information, the two-dimensional image seen by the eyes faces an encoding problem, which is language .
Chinese encodes the person where the signal transmitter is located as "I" , and encodes the person where the receiver is located as "you" , which is hieroglyph .
English encodes the signal transmitter as "I" and the receiver as "you", which is pinyin .
(PS: English must be the text of secondary civilization , not the text of main civilization [grin] Because English encoding method does not meet the characteristics of 90% of human information dependent on eyes)
This encoding is very rough compared with the computer H264 encoding .
H264 encoding can fully decode the image on the receiving end, while human language requires listener to based on prior information to brain [laugh]
H264 also has a prior information, which is the encoding protocol itself, which is based on discrete cosine transform (DCT), similar to the expansion of two-dimensional images on Fourier series .
The different terms of Fourier series are mathematically linearly independent of , forming a set of base in .
Human text can actually be regarded as "a set of bases" .
Chinese is a two-dimensional text, and the description in English is very simple.
It has 26 letters represent different pronunciations. These 26 letters temporary can be considered as linearly irrelevant to .
Use prime to give these 26 letters numbers:
1, A: 2, B: 3, C: 5, D: 7, E: 11, F: 13, G: 17.
2, H:19, I:23, J:29, K:31, L:37, M:41, N:43.
3, O:47, P:53, Q:59, R:61, S:67, T:71.
4, U: 73, V: 79, W: 83, X: 89, Y: 97, Z: 101.
Then, "I" (I) This visual information encoding on math is 23 [grin]
"you" (you) This visual information in math is 97*47*73 [grin]
Because prime numbers cannot be factorization , the factorization of the product of multiple prime numbers is unique in when the order is ignored, so this encoding is actually unique.
human brain processing large amount of information from eyes results , and ultimately a set of linearly irrelevant -based , as well as their coefficient combinations.
If underlying mechanism of human brain is also digital signal like computers, then this set of coefficient is rational number .
If the underlying mechanism of the human brain is the analog signal , it is a real number, but you can choose a rational number that is closest to it to : because rational number is dense on the real set.
When the human brain processes the visual information seen by the eyes, it obtains a set of rational numbers: those who have learned the real-variable function know that there is actually no difference between rational numbers and integer , both countable .
(rational numbers m/n can be regarded as two-dimensional integer pairs (m, n). According to the diagonal rule of , it can correspond to natural numbers one by one)
brain visual center This set of rational numbers processed by rational numbers , of course, it can be factorized into the product of and a set of prime numbers .
Now to pass this set of information through your mouth, you only need to find the different pronunciations of for different prime numbers for !
people have a limited pronunciation range (85-1100 Hz) and a limited hearing range (20-20000 Hz). Just find a suitable frequency for these commonly used coded prime numbers on the intersection of these two ranges.
In order to reduce the interference between different sounds (letter ), this frequency should be an prime number .
The sound intensity of people speaking does not change much. It can be considered that people's language is FM wave, not Amplitude modulation wave.
Why do deep learning generate ? People can't understand ?
Because the computer model cannot speak , it cannot communicate directly with people [covering your face]
I think that the current deep learning framework has problems writing. It should not use the computer double floating point number , but should use 2 int integers (rational numbers composed of ) to represent the weight .
This way, the network feature trained in this way, directly use the 2 integers on the numerator and denominator to factorize , and then give it a corresponding sound frequency .
As long as the computer can read data trained from a large number of pictures through sound , it is easy for humans to crack "language" , such as .
Deep learning
From the above analysis in this article, pinyin text is indeed closer to mathematical than pictograms.
Will you be scolded if you ask for a praise? [cover your face]
Perhaps 20 years later this article is a paper that will open up the world [grin]
