How code can build on code / how the architecture impact of learned intelligence impacts future architectural impacts. We start with the same wiring but early experiences change the trajectory for future architectural changes.
-
-
Replying to @markcannon5 @TonyZador and
Okay, thanks. That's really the central question we're after here - how much prior knowledge is prewired from genomic info and how much is learned
1 reply 0 retweets 4 likes -
Replying to @WiringTheBrain @TonyZador and
No problem. To me it's purely information that builds a neural architecture that facilitates intelligence creation. Based on experience etc.
1 reply 0 retweets 1 like -
Replying to @markcannon5 @TonyZador and
That's the position Bengio and
@ylecun are arguing for - as few priors as possible and mainly for meta-learning rather than specific knowledge.@GaryMarcus argues for more prior knowledge pre-wired in from the genome2 replies 4 retweets 10 likes -
Replying to @WiringTheBrain @markcannon5 and
Free Recall Retweeted Free Recall
But does the genome really encode "knowledge?" The Nativists assume, yet never explain, how knowledge is directly encoded in the genome. This preformist assumption contradicts the self-organized and pattern-forming dynamics of real developmental systems:https://twitter.com/freerecall/status/1204039449371242497?s=19 …
Free Recall added,
Free Recall @freerecallWhat are the origins of cognition? Do infants possess innate knowledge? Or is knowledge actively constructed over development? Smith (1999) on the necessity of understanding the developmental process to explain the origins of knowledge: https://onlinelibrary.wiley.com/doi/abs/10.1111/1467-7687.00062 … pic.twitter.com/fAj2m5z5VV1 reply 5 retweets 7 likes -
Replying to @freerecall @WiringTheBrain and
it's a great question. In some cases, the genome encodes pretty specific stuff. Dams beavers build are hardwired And if a mouse that builds a short burrow is raised by long-burrowing moms, it still builds short burrows. Cool stuff from Hoekstra:https://hoekstra.oeb.harvard.edu/
6 replies 4 retweets 24 likes -
Replying to @TonyZador @freerecall and
How much of this can be via genome-encoded reward fxn? For mouse who likes short borrows “ugh, I hate wide open spaces” & “mmm, I’ve burrowed a just bit and now I’m so satisfied”. Vs. pre-coded program of actions. If so, still lots of learning involved, just w/ specific cost fxn.
3 replies 0 retweets 4 likes -
Replying to @AdamMarblestone @freerecall and
i dont understand this distinction. Hopi's expts show that genes determine whether the burrow is short or long. One mech to encode this in the genome involves a reward function on the length of the burrow (and a bit more). probably need more than "stop now" for birdspic.twitter.com/WspWDX3iAM
2 replies 0 retweets 1 like -
Replying to @TonyZador @freerecall and
It doesn’t contradict the genomic encoding at all. But it puts a different emphasis as far as the need to *additionally* have a quite powerful RL system operating, to get the right behavior. It suggests a different AI emphasis: meta learning or evolving very specific reward fxns.
2 replies 0 retweets 1 like -
Replying to @AdamMarblestone @TonyZador and
Psychologists call this stuff "motivation"; it' part of the answer, but inot the whole answer. What I don't understand is why some folks are comfortable ascribing fairly detailed innate structure to motivations (or loss functions) but not to other aspects of cognition.
3 replies 0 retweets 2 likes
I’m comfortable ascribing innate structure to a ton else. But loss functions are nice and compact, and we know there is learning... so what I’m less comfortable with is AI people using too-generic and math-y end to end losses rather than diverse ethologically specific losses.
-
-
Replying to @AdamMarblestone @GaryMarcus and
The issue is that if you assume specific losses, which only act in specific settings, things start to just look like programs (in the computer sense.) Not a problem unless you want to eliminate explicit ops over variables, binding, etc. from cognition: but some seem to want that.
1 reply 0 retweets 1 like -
Replying to @recursus @GaryMarcus and
It looks like a developmental biology program! Also I should say that I do understand why people study simple end to end loss functions: to isolate other aspects of the machine learning problem and to keep things simple. So I don’t really mean that as a criticism of ML.
1 reply 0 retweets 0 likes - 6 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.