Great points about the cost of labeling and all the tinkering happening around the architecture which seems eerily similar to "feature engineering". Sure, the networks today can memorize inputs at an unimaginable scale, but generalization issues still persist.
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
You can also read Sutton's post as advocacy of "meta-methods that can find and capture this arbitrary complexity [of the world]" -- like the ideas you explored in the COG project, and others are trying to work out today with the new round of baby-machines:https://twitter.com/generuso/status/1107649917797965824 …
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Agree. If humans do task-oriented design, training data preparation and parameter tuning, we cannot claim that computational power was the key. This approach works for specialized systems but for task agnostic ones we need to add a -yet to be discovered- model of intelligence
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Great points, knowledge makes a big difference when designing better models. Even in case of one-shot learning models, knowledge rules. If we had infinite computation power to search the space of all possible designs, maybe, but we don't have that luxury.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
*Rich
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
In addition to all the other very good points you made, thank you for bringing environmental impact into the conversation.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Great post!
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Not only a handful examples may be sufficient for humans to learn something useful, but also those examples are acquired one by one, meaning that human learning is more incremental than statistical learning where all examples are thrown in at once.
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
-
-
Thank you for this!
Hvala. Twitter će to iskoristiti za poboljšanje vaše vremenske crte. PoništiPoništi
-
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.
