Frozen Encoder
In multilingual representation research, the practice of directly using intermediate-layer outputs of a pretrained language model as fixed features without fine-tuning on downstream tasks, so as to exclude interference from parameter updates.
Timeline (last 90 days)
The advantages of the frozen-encoder-plus-lightweight-readout paradigm include low training cost, shared features across multiple readout layers, and the need to train only a new readout layer for each new task, but the performance ceiling is limited by the representational capacity of the encoder.
A study on Bengali idioms used six frozen encoders, 10,359 Bengali idiom items, and eight input conditions to systematically test the reading, transfer, and erasure of cross-lingual concept subspaces.
All Facts (2)
The advantages of the frozen-encoder-plus-lightweight-readout paradigm include low training cost, shared features across multiple readout layers, and the need to train only a new readout layer for each new task, but the performance ceiling is limited by the representational capacity of the encoder.
50%UnverifiedA study on Bengali idioms used six frozen encoders, 10,359 Bengali idiom items, and eight input conditions to systematically test the reading, transfer, and erasure of cross-lingual concept subspaces.
50%