5 ms·

> May well turn out later on that AI is best modeled for speed and detail at 11 bit resolution - after all - evolution has done a lot of work that we equate as

> May well turn out later on that AI is best modeled for speed and detail at 11 bit resolution - after all - evolution has done a lot of work that we equate as machine learning. I don't see gradient descent or back propagation in biology.

Zenst 7y ago
Maybe not yet, but for example gradient descent - https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0219004 https://journals.plos.org/plosone/article?id=10.1371/journal... As for back propagation, I'm sure there are aspects in nature in which translate - maybe the nerves system of some species. Way to see it is that evolutionary change of any form could in theory be modeled in machine learning. There may be aspects of evolution and how it operates that help refine how we do machine learning.
salty_biscuits 7y ago
Well for gradient ascent you have this. https://en.wikipedia.org/wiki/Chemotaxis https://en.wikipedia.org/wiki/Chemotaxis
ianai 7y ago
iirc, back propagation sounds similar in inspection to some of the processes seen in sleeping brains. Ie what they see in MRIs looks like what back propagation might look like from outside. But I’m really not comfortable enough with these terms to speak about them.
DoctorOetker 7y ago
The jury is still out on back-propagation (i.e. reverse accumulation automatic differentiation) within the brain (retrograde signalling from the post-synaptic neuron to the pre-synaptic neuron may still be discovered and turn out to approximate RM AD) But regarding evolution implementing back-propagation, I agree that is just bonkers, as it would require cells to back-propagate in time information about death or procreation! If our biology implements information time machines, biology would be a lot weirder than it is (and it would imply that we could use that by sequencing an organisms genome (or ensemble of them) to read from the future, and keeping it fed without allowing reproduction, and later when we want to write information back either kill or let it reproduce). That's just plain nonsense. There can't be back-propagation in evolution without biological time-travel of information. And machine learning without back-propagation is basically a genetic algorithm, so the claim that "evolution has done a lot of work that we equate as machine learning" would then reduce to a widely accepted fact that nobody disputes and is thus uninformative: "evolution has done a lot of work that we equate as genetic algorithms"
- Zenst 7y ago
  "There can't be back-propagation in evolution without biological time-travel of information." You still can't do such time-travel in computers, sure you could simulate it but on a chronological scale in reality, your not time traveling. To equate this into evolution, if you viewed a species as the model, each generation adapts to the previous generations experience, hence it feeds back into the next generation. That evolution may prevail or that whole lineage dies off. It gets down to the perspective of how you measure that time aspect and define the start point of your time perspective. In a simulation you are arbitrary resetting that time baseline on each model, so the equivalent would be to view the human as a baseline of each birth. yes it gets confusing, but to dismiss it just because you have no actual time-travel from a cosmological perspective and compare to an arbitrary time perspective that runs in a linear temporal realm of the universe is unfair. But many things are hard to equate, you could say it is like comparing apples to pears, but then there are many perspectives in which you can compare them and find them equal. But you can't take one single perspective and transpose that upon another perspective and presume a conclusion.
  - DoctorOetker 7y ago
    ? if natural selection were to use back-propagation, then killing or allowing procreation would be measurable not only in the genome right after birth of an IVF organism, but even be measurable in the half genomes of the sperm and egg cells before they met (assuming we reach some non-destructe genome sequencing technology). That would simply make actual natural selection useless. Moreover, if all cells had this capacity, and our neurons are cells too, why don't I have this skill inately? I stand by my opinion that natural selection does not employ timetravel, and hence also does not employ backpropagation EDIT: adding, one does not need timetravel for backpropagation in the brain or in a computer, just enough memory or delay bandwidth to store the intermediate values for backpropagation. i.e. the feedback is nearly instantaneous, or low bandwidth from example to scoring, but scoring in natural selection depends on how much offspring one makes before one dies, how can this inform the choice mutations??
    - DoctorOetker 7y ago

Zenst 7y ago

Maybe not yet, but for example gradient descent - https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0219004 https://journals.plos.org/plosone/article?id=10.1371/journal... As for back propagation, I'm sure there are aspects in nature in which translate - maybe the nerves system of some species. Way to see it is that evolutionary change of any form could in theory be modeled in machine learning. There may be aspects of evolution and how it operates that help refine how we do machine learning.

salty_biscuits 7y ago

Well for gradient ascent you have this. https://en.wikipedia.org/wiki/Chemotaxis https://en.wikipedia.org/wiki/Chemotaxis

ianai 7y ago

iirc, back propagation sounds similar in inspection to some of the processes seen in sleeping brains. Ie what they see in MRIs looks like what back propagation might look like from outside. But I’m really not comfortable enough with these terms to speak about them.

DoctorOetker 7y ago

The jury is still out on back-propagation (i.e. reverse accumulation automatic differentiation) within the brain (retrograde signalling from the post-synaptic neuron to the pre-synaptic neuron may still be discovered and turn out to approximate RM AD) But regarding evolution implementing back-propagation, I agree that is just bonkers, as it would require cells to back-propagate in time information about death or procreation! If our biology implements information time machines, biology would be a lot weirder than it is (and it would imply that we could use that by sequencing an organisms genome (or ensemble of them) to read from the future, and keeping it fed without allowing reproduction, and later when we want to write information back either kill or let it reproduce). That's just plain nonsense. There can't be back-propagation in evolution without biological time-travel of information. And machine learning without back-propagation is basically a genetic algorithm, so the claim that "evolution has done a lot of work that we equate as machine learning" would then reduce to a widely accepted fact that nobody disputes and is thus uninformative: "evolution has done a lot of work that we equate as genetic algorithms"

Zenst 7y ago
"There can't be back-propagation in evolution without biological time-travel of information." You still can't do such time-travel in computers, sure you could simulate it but on a chronological scale in reality, your not time traveling. To equate this into evolution, if you viewed a species as the model, each generation adapts to the previous generations experience, hence it feeds back into the next generation. That evolution may prevail or that whole lineage dies off. It gets down to the perspective of how you measure that time aspect and define the start point of your time perspective. In a simulation you are arbitrary resetting that time baseline on each model, so the equivalent would be to view the human as a baseline of each birth. yes it gets confusing, but to dismiss it just because you have no actual time-travel from a cosmological perspective and compare to an arbitrary time perspective that runs in a linear temporal realm of the universe is unfair. But many things are hard to equate, you could say it is like comparing apples to pears, but then there are many perspectives in which you can compare them and find them equal. But you can't take one single perspective and transpose that upon another perspective and presume a conclusion.
- DoctorOetker 7y ago
  ? if natural selection were to use back-propagation, then killing or allowing procreation would be measurable not only in the genome right after birth of an IVF organism, but even be measurable in the half genomes of the sperm and egg cells before they met (assuming we reach some non-destructe genome sequencing technology). That would simply make actual natural selection useless. Moreover, if all cells had this capacity, and our neurons are cells too, why don't I have this skill inately? I stand by my opinion that natural selection does not employ timetravel, and hence also does not employ backpropagation EDIT: adding, one does not need timetravel for backpropagation in the brain or in a computer, just enough memory or delay bandwidth to store the intermediate values for backpropagation. i.e. the feedback is nearly instantaneous, or low bandwidth from example to scoring, but scoring in natural selection depends on how much offspring one makes before one dies, how can this inform the choice mutations??
  - DoctorOetker 7y ago