Ronen Tamari@rtk254

Very interesting work, obj.-oriented reps. + sequential application of learned rules - arch. biased towards both (1) learning of specialized modules as well as (2) their reuse. Compare w/ @robert_csordas https://t.co/JhcXKkD5O6 where Transformers found to succeed at 1 but not 2 https://t.co/RYIf8l19em