Pseudo-Label
In semi-supervised learning, a pseudo-label is a predicted annotation generated by a teacher model for unlabeled data, used as a substitute for ground-truth labels to supervise the training of a student model. It is one of the core techniques in semi-supervised learning.
Timeline (last 90 days)
The idea of turning pseudo-labels from targets into reward references could potentially be extended to other semi-supervised sequence generation tasks, such as machine translation or speech translation.
Using pseudo-labels as reward references for GRPO rather than as SFT training targets can significantly improve ASR models' robustness to label noise.
SFT forces the model to incorporate incorrect pseudo-labels into the policy's support set, destroying reward variance and causing the model to lose the ability to distinguish right from wrong.
The value of pseudo-labels lies not in being imitated as ground truth, but in serving as an evaluation signal that guides the model's exploration.
The risk of pseudo-labels lies in confirmation bias—erroneous predictions get reinforced; therefore in practice novelty detection (e.g., Mahalanobis distance) is usually combined to judge whether the distribution has drifted, and pseudo-label updates are enabled only when the drift is significant and prediction confidence is high.
Under domain mismatch conditions, relying solely on confidence-based filtering is insufficient to guarantee pseudo-label quality, because models may also exhibit high confidence on incorrect samples.
Common strategies for addressing error accumulation include: setting confidence thresholds to filter low-quality pseudo-labels, mixing in real annotated data as anchor points, and introducing reward models or verifiers for secondary screening.
Traditional pseudo-labeling suffers from confirmation bias: repeatedly using pseudo-labels from the same model causes biases to accumulate and amplify with each iteration.
All Facts (8)
The value of pseudo-labels lies not in being imitated as ground truth, but in serving as an evaluation signal that guides the model's exploration.
50%UnverifiedUsing pseudo-labels as reward references for GRPO rather than as SFT training targets can significantly improve ASR models' robustness to label noise.
50%UnverifiedSFT forces the model to incorporate incorrect pseudo-labels into the policy's support set, destroying reward variance and causing the model to lose the ability to distinguish right from wrong.
50%UnverifiedThe risk of pseudo-labels lies in confirmation bias—erroneous predictions get reinforced; therefore in practice novelty detection (e.g., Mahalanobis distance) is usually combined to judge whether the distribution has drifted, and pseudo-label updates are enabled only when the drift is significant and prediction confidence is high.
50%UnverifiedUnder domain mismatch conditions, relying solely on confidence-based filtering is insufficient to guarantee pseudo-label quality, because models may also exhibit high confidence on incorrect samples.
50%UnverifiedCommon strategies for addressing error accumulation include: setting confidence thresholds to filter low-quality pseudo-labels, mixing in real annotated data as anchor points, and introducing reward models or verifiers for secondary screening.
50%UnverifiedTraditional pseudo-labeling suffers from confirmation bias: repeatedly using pseudo-labels from the same model causes biases to accumulate and amplify with each iteration.
50%UnverifiedThe idea of turning pseudo-labels from targets into reward references could potentially be extended to other semi-supervised sequence generation tasks, such as machine translation or speech translation.
50%