New Study Finds a Single Neuron Is a Surprisingly Complex Little Computer (singularityhub.com)



In the study, the team found it took a five- to eight-layer neural network,ย 

or nearly 1,000 artificial neurons, to mimic the behavior of a single biological neuron from the brainโ€™s cortex.

์ด ์—ฐ๊ตฌ์—์„œ ์—ฐ๊ตฌํŒ€์€ ๋‡Œ์˜ ํ”ผ์งˆ์—์„œ ๋‚˜์˜ค๋Š” ๋‹จ์ผ ์ƒ๋ฌผํ•™์  ๋‰ด๋Ÿฐ์˜ ํ–‰๋™์„ ํ‰๋‚ด๋‚ด๊ธฐ ์œ„ํ•ด 5์ธต ๋‚ด์ง€ 8์ธต์˜ ์‹ ๊ฒฝ๋ง ๋˜๋Š”ย 

๊ฑฐ์˜ 1,000๊ฐœ์˜ ์ธ๊ณต ๋‰ด๋Ÿฐ์ด ํ•„์š”ํ•˜๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค.




1๊ฐœ์˜ ๋‰ด๋Ÿฐ์„ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” 1000๊ฐœ์˜ ์ธ๊ณต๋‰ด๋Ÿฐ์ด ํ•„์š”



๋˜๋Š”


1๊ฐœ์˜ ๋‰ด๋Ÿฐ์„ ๋ชจ๋ฐฉํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š”ย 5์ธต ๋‚ด์ง€ 8์ธต์˜ ์‹ ๊ฒฝ๋ง ํ•„์š”






์ธ๊ฐ„ ๋‰ด๋Ÿฐ ์ด ๊ฐœ์ˆ˜ ์•ฝ 860 ์–ต๊ฐœ ~ 1000 ์–ต๊ฐœย 


์ธ๊ณต ๋‰ด๋Ÿฐ์œผ๋กœ ์ด๊ฒƒ์„ ๊ตฌํ˜„ํ•˜๊ธฐ ์œ„ํ•ด ์ตœ์†Œย  86 ์กฐ๊ฐœย  ~ย  100 ์กฐ๊ฐœ ํ•„์š”











To computationally compare biological and artificial neurons, the team asked: How big of an artificial neural network would it take to simulate the behavior of a single biological neuron?


First, they built a model of a biological neuron (in this case, a pyramidal neuron from a ratโ€™s cortex). The model used some 10,000 differential equations to simulate how and when the neuron would translate a series of input signals into a spike of its own.


They then fed inputs into their simulated neuron, recorded the outputs, and trained deep learning algorithms on all the data. Their goal? Find the algorithm that could most accurately approximate the model.


์ƒ๋ฌผํ•™์  ๋‰ด๋Ÿฐ๊ณผ ์ธ๊ณต ๋‰ด๋Ÿฐ์„ ๊ณ„์‚ฐ์ ์œผ๋กœ ๋น„๊ตํ•˜๊ธฐ ์œ„ํ•ด ์—ฐ๊ตฌํŒ€์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ๋ฌผ์—ˆ๋‹ค.ย 

ํ•˜๋‚˜์˜ ์ƒ๋ฌผํ•™์  ๋‰ด๋Ÿฐ์˜ ํ–‰๋™์„ ํ‰๋‚ด๋‚ด๋ ค๋ฉด ์–ผ๋งˆ๋‚˜ ํฐ ์ธ๊ณต ์‹ ๊ฒฝ๋ง์ด ํ•„์š”ํ• ๊นŒ์š”?

๋จผ์ €, ๊ทธ๋“ค์€ ์ƒ๋ฌผํ•™์  ๋‰ด๋Ÿฐ์˜ ๋ชจ๋ธ์„ ๋งŒ๋“ค์—ˆ์Šต๋‹ˆ๋‹ค. (์ด ๊ฒฝ์šฐ, ์ฅ์˜ ํ”ผ์งˆ์—์„œ ๋‚˜์˜ค๋Š” ํ”ผ๋ผ๋ฏธ๋“œ ๋‰ด๋Ÿฐ). ์ด ๋ชจ๋ธ์€ ๋‰ด๋Ÿฐ์ด ์ผ๋ จ์˜ ์ž…๋ ฅ ์‹ ํ˜ธ๋ฅผ ์ž์‹ ์˜ ์ŠคํŒŒ์ดํฌ๋กœ ๋ณ€ํ™˜ํ•˜๋Š” ๋ฐฉ๋ฒ•๊ณผ ์‹œ๊ธฐ๋ฅผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๊ธฐ ์œ„ํ•ด ์•ฝ 10,000๊ฐœ์˜ ๋ฏธ๋ถ„ ๋ฐฉ์ •์‹์„ ์‚ฌ์šฉํ–ˆ์Šต๋‹ˆ๋‹ค.

๊ทธ๋Ÿฐ ๋‹ค์Œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋‰ด๋Ÿฐ์— ์ž…๋ ฅ์„ ๊ณต๊ธ‰ํ•˜๊ณ , ๊ฒฐ๊ณผ๋ฅผ ๊ธฐ๋กํ•˜๊ณ , ๋ชจ๋“  ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ํ›ˆ๋ จ์‹œ์ผฐ์Šต๋‹ˆ๋‹ค.ย 

๊ทธ๋“ค์˜ ๋ชฉํ‘œ๋Š”? ๋ชจํ˜•์„ ๊ฐ€์žฅ ์ •ํ™•ํ•˜๊ฒŒ ๊ทผ์‚ฌํ•  ์ˆ˜ ์žˆ๋Š” ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์ฐพ์Šต๋‹ˆ๋‹ค.













1๊ฐœ์˜ ๋‰ด๋Ÿฐ์„ ํ‰๋‚ด๋‚ด๊ธฐ ์œ„ํ•ด์„œ๋Š” 1000๊ฐœ์˜ ์ธ๊ณต ๋‰ด๋Ÿฐ์ด ํ•„์š” OR 5~8์ธต์˜ ์‹ ๊ฒฝ๋ง ํ•„์š”

๋…ธ๋“œ(์› 1๊ฐœ) X 1000 = ์ƒ๋ฌผํ•™์  ๋‰ด๋Ÿฐ 1๊ฐœ







They increased the number of layers in the algorithm until it was 99 percent accurate at predicting the simulated neuronโ€™s output given a set of inputs. The sweet spot was at least five layers but no more than eight, or around 1,000 artificial neurons per biological neuron. The deep learning algorithm was much simpler than the original modelโ€”but still quite complex.


๊ทธ๋“ค์€ ์ผ๋ จ์˜ ์ž…๋ ฅ์ด ์ฃผ์–ด์กŒ์„ ๋•Œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜๋œ ๋‰ด๋Ÿฐ์˜ ์ถœ๋ ฅ์„ 99% ์ •ํ™•ํ•˜๊ฒŒ ์˜ˆ์ธกํ•  ์ˆ˜ ์žˆ์„ ๋•Œ๊นŒ์ง€ ์•Œ๊ณ ๋ฆฌ์ฆ˜์˜ ์ธต ์ˆ˜๋ฅผ ๋Š˜๋ ธ๋‹ค. ์ƒ๋ฌผ ๋‰ด๋Ÿฐ 1๊ฐœ๋‹น ์ตœ์†Œ 5๊ฐœ ์ธต์—์„œ 8๊ฐœ, ๋˜๋Š” ์•ฝ 1,000๊ฐœ์˜ ์ธ๊ณต ๋‰ด๋Ÿฐ์ด ๊ฒ€์ถœ๋˜์—ˆ๋‹ค. ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜์€ ์›๋ž˜ ๋ชจ๋ธ๋ณด๋‹ค ํ›จ์”ฌ ๊ฐ„๋‹จํ–ˆ์ง€๋งŒ ์—ฌ์ „ํžˆ ์ƒ๋‹นํžˆ ๋ณต์žกํ–ˆ๋‹ค.



Indeed, after removing these features, the team found they could match the simplified biological model with but a single-layer deep learning algorithm.


์‹ค์ œ๋กœ ์ด๋Ÿฌํ•œ ํŠน์ง•์„ ์ œ๊ฑฐํ•œ ํ›„ ์—ฐ๊ตฌํŒ€์€ ๋‹จ์ˆœํ™”๋œ ์ƒ๋ฌผํ•™์  ๋ชจ๋ธ์„ ๋‹จ์ธต ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์ผ์น˜์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๋ฐœ๊ฒฌํ–ˆ๋‹ค.


https://www.youtube.com/watch?v=3LQLCqHT5Ws&t=27s


Also, the model neuron is from a ratโ€™s brain, as opposed to a humanโ€™s, and itโ€™s only one type of brain cell. Further, the study is comparing a model to a modelโ€”there is, as of yet, no way to make a direct comparison to a physical neuron in the brain. Itโ€™s entirely possible the real thing is more, not less, complex.

๋ชจํ˜• ๋‰ด๋Ÿฐ์€ ์ฅ์˜ ๋‡Œ์—์„œ ๋‚˜์˜จ ๊ฒƒ์œผ๋กœ, ์ธ๊ฐ„์˜ ๋‡Œ์™€๋Š” ๋ฐ˜๋Œ€๋กœ, ๋‡Œ์„ธํฌ์˜ ํ•œ ์ข…๋ฅ˜์ผ ๋ฟ์ž…๋‹ˆ๋‹ค. ๊ฒŒ๋‹ค๊ฐ€, ์ด ์—ฐ๊ตฌ๋Š” ๋ชจ๋ธ์„ ๋ชจ๋ธ์— ๋น„๊ตํ•˜๊ณ  ์žˆ๋‹ค.ย 

์•„์ง๊นŒ์ง€๋Š” ๋‡Œ์˜ ๋ฌผ๋ฆฌ์  ๋‰ด๋Ÿฐ๊ณผ ์ง์ ‘ ๋น„๊ตํ•  ๋ฐฉ๋ฒ•์ด ์—†๋‹ค. ์ง„์งœ๊ฐ€ ๋” ๋ณต์žกํ•  ์ˆ˜๋„ ์žˆ๊ณ  ๋œย ๋ณต์žกํ•  ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค.



But itโ€™s hard to argue with millions of years of evolutionary experimentation. So far, following the brainโ€™s blueprint has been a rewarding strategy. And if this work is any indication, future neural networks may well dwarf todayโ€™s in size and complexity.


๊ทธ๋Ÿฌ๋‚˜ ์ˆ˜๋ฐฑ๋งŒ ๋…„์— ๊ฑธ์นœ ์ง„ํ™”๋ก ์  ์‹คํ—˜์„ ๋ฐ˜๋ฐ•ํ•˜๊ธฐ๋Š” ์–ด๋ ต์Šต๋‹ˆ๋‹ค. ์ง€๊ธˆ๊นŒ์ง€๋Š” ๋‘๋‡Œ์˜ ์ฒญ์‚ฌ์ง„์„ ๋”ฐ๋ฅด๋Š” ๊ฒƒ์ด ๋ณด๋žŒ ์žˆ๋Š” ์ „๋žต์ด์—ˆ์Šต๋‹ˆ๋‹ค. ๊ทธ๋ฆฌ๊ณ  ์ด ์ž‘์—…์ด ์–ด๋–ค ์ง•ํ›„๋ผ๋ฉด, ๋ฏธ๋ž˜์˜ ์‹ ๊ฒฝ๋ง์€ ํฌ๊ธฐ์™€ ๋ณต์žก์„ฑ์ด ์˜ค๋Š˜๋‚ ๋ณด๋‹ค ํ›จ์”ฌ ์ž‘์•„์งˆ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.















์š”์•ฝย 


1. ์ฅ ๋‹จ์ผ ๋‰ด๋Ÿฐ 1๊ฐœ๋ฅผ ๊ตฌํ˜„ย - 1000๊ฐœ์˜ ์ธ๊ณต๋‰ด๋Ÿฐ OR 5~8์ธต ์‹ฌ์ธต ์‹ ๊ฒฝ๋ง ํ•„์š”

2. ๋‰ด๋Ÿฐ์˜ ๊ตฌ์กฐ๋ฅผ ๋‹จ์ˆœํ™” ์‹œํ‚ค๋‹ˆ ๋‹จ์ธต ๋”ฅ๋Ÿฌ๋‹ ์•Œ๊ณ ๋ฆฌ์ฆ˜๊ณผ ์ผ์น˜ํ•œ๋‹ค๋Š” ์‚ฌ์‹ค ๋ฐœ๊ฒฌ

3. ์•„์ง ๋‹จ์ผ ๋‰ด๋Ÿฐ์„ ๋Œ€์ƒ์œผ๋กœ ํ•œ ์‹คํ—˜์ด๊ณ , ๊ทธ๋งˆ์ €๋„ ์ƒ์ฅ์˜ ๋‰ด๋Ÿฐ, ๋‰ด๋Ÿฐ๊ณผ ๋‰ด๋Ÿฐ์ด ์—ฐ๊ฒฐ๋œ ์‹œ๋ƒ…์Šค ๊ตฌ์กฐ์™€ ์ธ๊ฐ„ ๋‰ด๋Ÿฐ์— ๋Œ€ํ•œ ์ •๋ณด๋Š” ๋ฏธ์ง€์ˆ˜

4. ์—ญ๊ณตํ•™์€ ๋งค์šฐ ์˜๋ฏธ์žˆ๋Š” ์ „๋žต์ด๋ฉฐ, ๋ฏธ๋ž˜์—๋Š” ์‹ ๊ฒฝ๋ง์˜ ํฌ๊ธฐ์™€ ๋ณต์žก์„ฑ์„ ์˜ค๋Š˜๋‚  ๋ณด๋‹ค ๋‹จ์ˆœํ™” ์‹œํ‚ฌ ์ˆ˜ ์žˆ๋‹ค.

(1๊ฐœ์˜ ์ธ๊ณต ๋‰ด๋Ÿฐ์„ ์ƒ๋ฌผํ•™์  ์ธ๊ณต ๋‰ด๋Ÿฐ๊ณผ ๋˜‘๊ฐ™์€ ๋ฉ”์ปค๋‹ˆ์ฆ˜์œผ๋กœ ๊ตฌํ˜„ ๊ฐ€๋Šฅํ•˜๊ฑฐ๋‚˜ ๋” ๋›ฐ์–ด๋‚˜๊ฒŒ ๋งŒ๋“ค ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ณผ ์‹œ)