GTF
Generalized Teacher Forcing; an RNN training technique for dynamical system reconstruction that blends ground-truth observations and model predictions at a tunable mixing ratio as inputs, reducing exposure bias and serving as an attractor regularizer in chaotic systems.
Timeline (last 90 days)
The research team introduced Generalized Teacher Forcing (GTF) to stabilize DEER, making its iterative process controllable and convergent again.
Compared to conventional teacher forcing, GTF can reduce exposure bias (the distribution mismatch caused by relying on ground-truth data during training and on the model's own predictions during inference).
GTF mixes true observations and the model's own predictions in proportion via a tunable mixing coefficient before feeding the result into the next step, biasing toward true data in the early stages of training and gradually increasing the proportion of the model's self-predictions as training progresses.
All Facts (3)
The research team introduced Generalized Teacher Forcing (GTF) to stabilize DEER, making its iterative process controllable and convergent again.
50%UnverifiedCompared to conventional teacher forcing, GTF can reduce exposure bias (the distribution mismatch caused by relying on ground-truth data during training and on the model's own predictions during inference).
50%UnverifiedGTF mixes true observations and the model's own predictions in proportion via a tunable mixing coefficient before feeding the result into the next step, biasing toward true data in the early stages of training and gradually increasing the proportion of the model's self-predictions as training progresses.
50%