The interactive two-timescale model (M4) provides the most general account of naturalistic learning, though the additive-forgetting model (M3) fits better at six of eight interval subsets¶

Claim · All claims
Evidence · 1 study · associational r2 · q2
Currency · newest study 2022 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? For Lost in Migration, M4 gave the most accurate out-of-sample predictions overall; for Ebb and Flow, M3 and M4 were equivalently accurate. → Aakriti Kumar 2022 q2 i? M3 outperformed the interactive model at six of the eight retention-interval subsets, so M4's superiority rests on the widest intervals. → Aakriti Kumar 2022

Evidence¶

Aakriti Kumar 2022¶

Aakriti Kumar, Aaron S. Benjamin, Andrew Heathcote and Mark Steyvers. (2022). Comparing models of learning and relearning in large-scale cognitive training data sets. npj Science of Learning 7:24. https://doi.org/10.1038/s41539-022-00142-x

q2 · i? · associational · r2

Out-of-sample predictive evaluation (RMSE on 20% held-out data, 80/20 split) of four models fit to individual users' gameplay; Table 1 reports e.g. RMSE 3.55 for M4 vs 3.56 for M3 on all Lost in Migration data, with 19,463 and 19,694 participants per game.

"For lost in migration, M 4 provides the most accurate predictions overall. For ebb and flow, models M 3 and M4 provide equivalently accurate predictions. Taken together, M 4, which uniquely allows for a degree of interactivity between the two timescales of learning, provides the most general account of performance."

Discussion¶

Related Claims¶

• Losses after long breaks in naturalistic practice are rapidly offset by very little practice: a single gameplay can restore prior performance levels even after delays of hundreds of days — related
• Warm-up decrements are ubiquitous and substantial in naturalistic learning: the two-timescale model M2 improves out-of-sample prediction across the board over the baseline model M1 — related
• Models able to weight performance by recency fit better on the Assistments and KDD datasets, without explicit memory-decay terms being necessary — related
• Performance loss after breaks in practice grows with delay: a 100-day delay yields a 20% reduction and a 400-day delay a 50% reduction in first-gameplay performance — related
• Repeated retrieval practice alters the rate at which learned items are forgotten, according to learn-to-criterion studies and forgetting-curve analyses the chapter reviews — related
• No consistent learning effect of repeated TrackFX gameplay was found across children, though 37% showed positive performance slopes and improvement reached an asymptote after six sessions — related