Friday, September 7, 2012
Nervous system has around 10-100 billion neurons. Half of these are in cerebellum and most others in brain. Each eye has around 1 million neurons that reach brain and there may be less than 10 million primary sensory neurons in total. Electric frequency in brain could be from around zero Hz to several hundred Hz but gamma frequency that is associated with awareness is usually between 30-100 Hz. For 10-100 billion this gamma frequency could mean 300-10 000 billion signals/bits (~37-1200 GB) per second. This number would be much lower if knowing cortical column activity was enough. Cortical columns are vertical columns of ~100 neurons through cortex that usually activate together. Brain cortex has around hundred million cortical columns at most. Data flow through them could be 3-30 billion signals/bits (~0.37-3.7 GB) per second.
If neuronal activity was represented with pixels and graphic cards would have to animate them then best cards nowadays are quite close to being able to keep up with this data flow. Best Nvidia card below 500 dollars is about as powerful as the best AMD card with that price in terms of pixel and texel writing speed. These cards can write 32 billion pixels per second and 128 texels per second with memory bandwidth of 192 GB per second. This should be enough to send some screen overview of all the cortical columns even if they worked at 100 Hz. If humans had 100 billion neurons working at 100 Hz then those cards could mediate at most 16% of this activity but if humans had 10-30 billion neurons that are usually below 50 Hz then it could be enough to send some giant screen real time overview of activity from every single neuron.
This of course doesn't describe all the possible connections between neurons but graphic cards could be used as some memory or input for other processors that compared activity between certain cells/pixels by getting input from graphics card.
Connections between cells could be expressed in minimalistic binary. For example quadtree type addresses save memory compared to address expressed in numbers. Quadtree divides area to 4 and each such step takes 2 bits (1 bit system could be used to divide area to half). While 1st 2 bits express in which quarter something is then next 2 bits could divide this 25% into further 4 parts. In 5 steps it would 1 out of 1000 and every extra 5 steps would increase precision by ~1000 so 20 bit quadtree could find 1 pixel out of million. 40 bit quadtree should find 1 out of trillion which is more than enough for human brain neuron addresses. 40 bits is 5 bytes compared to 12 bytes that it would take to express trillion with decimal numbers in bytes. Each connection between 2 neurons could be expressed in 5 bytes no matter how far apart they are but each extra connection could take extra 5 bytes unless connections between neighboring cells could be simplified due to most of quadtree address being identical. If there were 10 billion memory cells each with 10 connections at most then it would take around 0,5 terabytes to represent all the memories.
Memory capacity may be more serious limitation as 2 terabyte hard drives are around 120 dollars which could fill up in less than 10 seconds if all activity got saved 100 times per second but it is not likely that all human memory cells work at the same time. Recording all sensory activity may be easier. If humans had 10 million sensory neurons then even if they worked at 100 Hz (1 Gb or 120 MB per second) it would take around half a hour to fill up the memory. Humans memory is very fuzzy compared to digital memories. Personally usual memories feel like dimly lit simplified minimalistic quickly replaying slideshows that could probably take much less than 1 MB for these couple of seconds. Human memory seems to remember mostly unusual stuff that we are not used while forgetting most monotonous repetitive experiences. If nothing special happens in some day then it's easy for humans to forget what happened on that day after weeks or months passed.
To go on more hypothetical territory i suspect axons and myelin around them are source of subjective sentience and structures like that are not found in electronics as far as i know.
I doubt anyone can produce billions of myelin type and sized structures for 500 dollars yet but electronics in terms of data processing and computing is quite close to being able to do many things that humans can. Special structures may be needed to give data analyzers ability to feel like living creatures.