What's missing in this paper is the notion of self-supervised learning: https://medium.com/intuitionmachine/the-paradigm-shift-of-self-supervised-learning-744a6819ce08 … We have to ween ourselves away from the misconstrued idea of unsupervised learning.
-
-
-
I think self-sup learning is cool and important! But do you think it is a dominant form of learning in animals other than humans? Do you think it is how colts can walk hours after birth?
- Još 5 drugih odgovora
Novi razgovor -
-
-
Interesting that you avoid most of the discussion in the last 70 years from Bateson, Kuo, Hinde, Gottlieb, and others from developmental psychobiology. That brain connectivity may be present at birth but it has to come from some ontogenetic process shaping synaptic pruning.
-
And considering the dynamic and interactive nature of synaptic overproduction and pruning, and the organismal feedback from behavior and physiological onto these processes, some of the statements about being "encoded" in the genome seem a bit hyperbolic.
- Još 4 druga odgovora
Novi razgovor -
-
-
Thanks for the intriguing paper! Between the extreme cases of C. elegans and humans, there must be a transition from hard-coded to rule-encoded connectomes. Has there been work to identify organisms nearer the transition, or to place this transition in evolutionary history?
-
One might imagine that eg drosophila, with 100k neurons, is near the transition. It has some identifiable neurons that are same across individuals but probably some randomness of connectivity
- Još 1 odgovor
Novi razgovor -
-
-
This may be also of interest on gene differences between how mice and human brains operatehttps://www.nature.com/articles/s41586-019-1506-7 …
-
This is a tour de force, and really important for eg work on human disease. But it is not easy to infer from these genetic differences what makes humans functionally special. I suspect the expression differences for any pair of mammals separated by as many Myrs are comparable.
Kraj razgovora
Novi razgovor -
-
-
Nice points, but misses entire fields that have been doing what you advocate: Indirect encoding (eg HyperNEAT), neuroevolution (evolves architecture, often w/ intralife learning) https://www.nature.com/articles/s42256-018-0006-z …. The AI-GA paper makes many similar points to yours https://arxiv.org/abs/1905.10985
-
thanks! looking forward to reading it.
- Još 2 druga odgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.