by Tessler, Tsivids, Madeano, Harper and Tenenbaum (Arxiv 2021)
The authors want to understand how language enables the accumulation of knowledge across generations. The approach is specifically to compare learning curves between (a) multiple generations, where each generation receives two lives at a game and can pass their knowledge to the next generation via language, and (b) a single learner who can continually learn with the same total number of lives.
As stated above, the idea is to compare individual players against a “multi-generational player” i.e. a sequence of players who pass written information to the next generation. This multi-generational players looks like:
Each individual player and “multi-generational player” (sometimes referred to as chains) play 10 games, which each require a different type of problem-solving:
Messages stay within a single lineage. There are 10 parallel chains, with 8 generations (a.k.a. participants) per chain. N=80 total participants.
No analyses on the content of the inter-generational messages, and how they affect the next generation’s performance
Generations are based on chained, meaning message will persist regardless of performance it provides. Ideally, one would want messages to be passed proportionally to how successful the progenitor was, regardless of generation? Two of the strengths of cultural learning is that (a) we can select who we learn from, and (b) written instructions aren’t reliant on the weakest link in a chain of descendants.