In machine learning tasks like Generalized Zero-Shot Learning (GZSL) or Generalized Category Discovery (GCD), a model is trained on a set of "seen" classes but must identify both seen and "unseen" classes at test time. Naturally, these models develop a heavy bias toward the categories they encountered during training. When presented with a novel object, the model often misclassifies it as a visually similar seen category simply because it is more "familiar" with that label's distribution.
To solve this, researchers have developed a suite of calibration techniques designed to level the playing field, ensuring that unseen classes have a fair chance of being predicted.
Inference Time Score Adjustment
The most direct way to mitigate bias is to adjust the model's output scores (logits) during the inference phase. If a model is consistently overconfident in seen classes, we can manually suppress those scores to allow unseen classes to emerge.
Ali Cheraghian and Yongqin Xian et al. describe a method called Calibrated Stacking, which subtracts a fixed penalty from the prediction scores of seen classes. By shifting the decision boundary, the model requires a much higher degree of visual certainty to pick a "seen" label over an "unseen" one. Similarly, Haonan Lin et al. propose Self-Debiasing Calibration (SDC), which uses the output of a frozen, biased pre-trained model to surgically subtract bias from the active model's logits.
While additive shifts are common, they can be sensitive to the magnitude of the logits. Joshua Feinglass et al. and Debasmit Das et al. argue that multiplicative scaling is more stable. By multiplying seen-class scores by a factor less than one, the correction scales with the model's confidence, maintaining better performance across a wider range of hyperparameters.
Training with Calibration Losses
Rather than fixing the bias after the model is trained, many researchers advocate for "baking" fairness into the training objective itself. This involves introducing loss functions that penalize the model for becoming too specialized in seen categories.
Ehsan Elhamifar proposes a self-calibration loss that introduces a margin during training, explicitly penalizing high scores for seen classes while encouraging the model to reserve probability mass for unseen ones. Ling Shao et al. and Huadong Sun et al. use similar category calibration losses to shift probability mass toward unseen classes, ensuring the model doesn't collapse its feature space around only known data.
In the context of incremental learning, Cheng-Hsun Lei et al. show that using binary cross-entropy instead of standard softmax can reduce bias. By treating each class activation as an absolute value rather than a relative probability, the model avoids the "winner-take-all" dynamic that typically favors newly learned or frequently seen classes.
Geometric and Distributional Alignment
A more sophisticated approach involves reshaping the underlying feature space or aligning the statistical distributions of the scores. The goal is to ensure that the "shape" of the data for seen and unseen classes looks similar to the model.
Feng Li et al. propose statistical distribution alignment, which penalizes discrepancies between the mean and variance of prediction scores for seen and unseen classes. If the scores for seen classes are consistently higher or more varied, the model is biased; matching these statistics forces the model to treat both sets of classes with equal numerical weight.
Geometrically, Rui Ma et al. use prototypes from known base classes as anchors to nudge novel samples toward their true centers. By using the well-defined geometry of seen classes to "calibrate" the sparsely defined space of unseen classes, the model can more accurately estimate where a new category should reside. Boris N. Oreshkin et al. complement this by rescaling the distance metrics of seen class prototypes, effectively pushing them further away in the embedding space to make room for novel categories.
Novelty Detection as a Gatekeeper
Finally, some methods use a two-step process: first determining if an object is "novel," then applying calibration. Jiannan Ge et al. and Runyu Ding et al. employ binary indicators or classification heads to distinguish between base (seen) and novel (unseen) instances. When the system detects a novel object, it triggers a suppression mechanism that down-weights all seen-class scores, preventing the model's inherent bias from drowning out the correct, albeit unfamiliar, label.
Key Foundational Work
An Empirical Study and Analysis of Generalized Zero-Shot Learning for Object Recognition in the Wild (2016) — contributed 2 concept elements to this direction.
Generalized Zero-Shot Learning with Deep Calibration Network (2018) — contributed 1 concept elements to this direction.