Identifying transfer learning in the reshaping of inductive biases
Anna Székely, Wigner Research Centre for Physics // Budapest University of Technology, Hungary; Balázs Török, Mozalearn Ltd., Hungary; Dávid Gergely Nagy, Gergő Orbán, Wigner Research Centre for Physics, Hungary; Mariann M. Kiss, Dezső Németh, Eötvös Lóránd University, Hungary; Karolina Janacsek, University of Greenwich, United Kingdom
Posters 3 Poster
Pacific Ballroom H-O
Sat, 27 Aug, 19:30 - 21:30 Pacific Time (UTC -8)
Transfer learning is a critical hallmark of human intelligence that has been frequently pitted against the capacities of artificial learning agents. Yet, the computations relevant for transfer learning have been little-investigated in humans. Here we follow an analytical paradigm that allows tracking learning in individuals and to identify signatures of the transfer of knowledge. From a Bayesian learning perspective, updating the prior over possible in- ventories that can be recruited for interpreting data is the key for efficient transfer of knowledge. We investigate two consequences of this computation: 1, Expediting the acquisition of new internal models; 2, Flexible paral- lel maintenance of multiple internal models. We reverse- engineered the internal models of individuals in an implicit sequence learning paradigm from their responses. Participants were trained on a non-trivial visual stimulus sequence for multiple days, which was followed by the introduction of a new sequence. We found that acquisition of the new sequence was considerably sped up by earlier exposure but this enhancement was specific to individuals showing signatures of abandoning initial inductive biases. Enhancement of learning was reflected in building up a new internal model. We found internal models were automatically switched when the sequences were interchanged.