quick thread, any chance of a comment @jeffclune ? It's generally accepted that we need models which learn from small amounts of data. But, the way to get there is to learn the learning agents themselves, which require will require ENORMOUS amounts of data.https://twitter.com/blake_camp_1/status/1217486172588347392 …
-
-
I don't disagree with your suggestion than we can do better than evolution, but i do disagree with your assumption that meta-learning is essentially evolution. You can endow agents with any innate priors you want in the inner loop. In fact, we already do this....
-
convolution, attention, hierarchical layers, recurrent connections. All of these can be regarded as innate priors which are used to accelerate optimization in the inner loop. So, it is not true that we are starting from a primordial soup, as you suggest.
- Još 7 drugih odgovora
Novi razgovor -
-
-
Humans aren't generally intelligent. Our behaviours evolved to support a specific embodiment following a specific lifestyle. The phase space of goals we can achieve is essentially zero.
-
that is just semantics. I'd say humans aren't "universally intelligent", but human intelligence is more general than anything we have now. So yeah, my definition of generality is based on human generality, not anything more.
Kraj razgovora
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.