[FRIAM] AI

[Prof David West]

Marcus made a comment recently about constructing an AI plus robotic body that provided the AI with sensory inputs comparable to a human being. It made me wonder about feasibility of such an idea.

The average human body has about 100 billion nerve endings generating electrical impulses

The average human (sex, weight, height sensitive) has about 30 trillion cells emitting ultra-weak biophotons; increasingly shown to play a role in inter-cellular communication

It is extremely difficult to compare something like FLOPS for the brain, but best estimates suggest an average of 43 teraFLOPS, and up to 430 teraFLOPS for peak situations. Computers are capable of 1.1 exaFLOPS. But the brain uses 20 watts of power and the computer megawatts.

Taking into account synaptic delay and refactory delay, each nerve ending could send a signal to the brain, or the brain could ‘process’ those signals at a rate between 10 Hz (cortex) to 1,000 Hz elsewhere. Also assume that the biophotons work mostly locally and maybe 1 percent actually end up triggering something akin to a nerve signal so, until we know more, it is unlikely that more than 30,000 to 300,000 additional signals reach the brain – less than noise, given what we know now. But that might change significantly in the future, especially as we learn more about quantum effects in the brain in general.

The brain could receive 5 trillion discrete signals per second, but “pre-processing” reduces that to between 50 (average) and 500 million (peak) signals per second.

.02-.03 percent of those signals are symbolic- originating in a phoneme, lexeme, word, number.

Between .22 and 12.3 of the “non-symbolic” signals process by the brain have a mediating effect on symbolic processing, in the human brain. Some of this can be simulated by an AI. Take sarcasm as an example: humans use a lot of non-symbolic signals to detect sarcasm with a success rate of about 95%. AI’s must rely on context, on explicit labeling of training material, and, if available sound or images that can be analyzed. With a success rate of about 80%.

Currently, an AI can simulate/emulate/equate to the roughly .02-.03 percent of the signal processing  done by the human brain, i.e., that directly related to symbolic inputs. It can also deal with, roughly 80% (based on the sarcasm example) of the mediating non-symbolic signals (between .22 and 12.3 percent of signals processed by the brain.

These numbers suggest, to me, that an AI is capable of simulating/emulating/equating-to about 1 to 15% of human brain signal processing. Of course, the human brain has all kinds of help elsewhere in the body, synthesizing, attenuating (reducing), and “pre-processing” signals. An AI has none of that help.

So, it seems to me, that an AI must necessarily be a true idiot-savant for language manipulation and pattern recognition (image, sound).

Only if we define human intelligence as nothing more than human abilities with language and visual/auditory pattern recognition can we say that artificial intelligence meets or exceeds (only in terms of speed) human intelligence.

I used AI to generate all the numbers in the above.

davew