Auxiliary Confidence Losses

• Auxiliary confidence losses are supplemental objective functions that adjust model confidence outputs to improve calibration, error detection, and robustness.
• They employ methods like entropy regularization, TCP regression, and confidence minimization to balance primary and auxiliary training objectives.
• Empirical evaluations show marked improvements in weak supervision, OOD detection, and meta-learning, yielding enhanced selective prediction and performance.

Auxiliary confidence losses are supplemental objective functions introduced in model training to manipulate, estimate, or regulate the network’s confidence output. These losses are used in various regimes, including weak supervision, confident pseudo-label selection, out-of-distribution (OOD) detection, selective classification, and joint optimization via meta-learning to adaptively balance primary and auxiliary objectives. They target improvement in calibration, error detection, robustness, or convergence by shaping the network’s confidence structure, often via entropy regularization, regression to true class probability, or minimization of overconfidence on adversarial or uncertain inputs.

1. Formal Definitions and Main Approaches

Auxiliary confidence losses generally supplement the primary supervised or weakly supervised objective with one or more terms that either (a) encourage the model to produce calibrated or conservative confidence values; or (b) act as instance-level weighting factors for pseudo-label–based training.

Notable formulations include:

• Instance-level entropy-weighted losses: Used for Learning from Label Proportions (LLP), these combine bag-level supervisory signal with an auxiliary pseudo-label loss, weighted by functions of entropy to select only high-confidence instances (Ma et al., 2024).
• Confidence estimation via regression: Auxiliary networks are trained to predict the true class probability (TCP), using a regression loss (typically MSE) between the auxiliary network’s output and the main classifier’s softmax probability assigned to the true label. This differs from simply using the max softmax output (MCP), providing more faithful error detection (Corbière et al., 2020).
• Confidence minimization on uncertainty sets: For OOD detection and conservative prediction, an auxiliary loss penalizes confident predictions on auxiliary (outlier/uncertain) data by minimizing cross-entropy to the uniform distribution, aligning the network’s output with maximal entropy when faced with unknowns (Choi et al., 2023).
• Adaptive mixing via bi-level optimization: In the context of noisy supervision or knowledge distillation, instance-level mixing weights for auxiliary confidence-related losses are meta-learned to minimize downstream validation loss (Sivasubramanian et al., 2022).

2. Entropy-based Weighting and Pseudo-labeling

Filtering or emphasizing examples based on the model’s confidence is critical in weakly supervised settings with unreliable labels. In "Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions" (L²P-AHIL), confidence is quantified using a dual entropy-based weight (DEW):

• Bag-level entropy ($\omega^{\mathrm{b}}$): Measures deviation of the within-bag distribution of per-class predictions from the ideal (sorted) proportion. Defined as a Gaussian-shaped function of the entropy gap, encourages high $\omega^{\mathrm{b}}$ where the model’s prediction distribution concentrates according to the bag’s ground-truth proportions.
• Instance-level entropy ($\omega^{\mathrm{i}}$): Monotonically penalizes instance-level predictions with high entropy. Perfectly confident predictions get weight close to 1; highly uncertain predictions are down-weighted exponentially with respect to entropy.

The combined weight $$\omega_{i,j} = \omega^{\mathrm{b}}_{i,c} \cdot \omega^{\mathrm{i}}_{i,j}$$ directly modulates an auxiliary cross-entropy loss over strong augmentations using hard pseudo-labels. The total objective is a sum of the bag-level proportion loss and the auxiliary instance-level loss, balanced by a coefficient $\alpha$ (Ma et al., 2024).

3. Auxiliary Confidence Estimation Networks

Accurate confidence quantification is essential for selective prediction and failure detection. "Confidence Estimation via Auxiliary Models" introduced the ConfidNet auxiliary model: a small network regressing the main classifier’s TCP $\operatorname{TCP}_F(x,y)$, with parameters optimized by minimizing mean squared error relative to the pre-computed ground-truth probabilities (Corbière et al., 2020).

The key distinction is the use of the ground-truth label (for TCP) rather than the predicted label (for MCP) as the regression target. Different architectures (MLP for classification, multi-scale heads for segmentation) are employed depending on the primary task. In unsupervised domain adaptation (UDA), this auxiliary confidence map guides high-confidence pseudo-label assignment for target-domain self-training.

This auxiliary loss realizes superior performance in both failure prediction (lower FPR, higher AUPR) and UDA semantic segmentation (improved mIoU), demonstrating the concrete utility of accurate, regression-based confidence estimation.

4. Confidence Minimization for Robustness and OOD Detection

Auxiliary losses that minimize confidence on outlier or uncertain data provide a mechanism for calibrated and selective prediction, particularly in safety-critical environments or under distribution shift.

In the Data-Driven Confidence Minimization (DCM) framework (Choi et al., 2023):

• The auxiliary loss is a cross-entropy between the model’s prediction and the uniform distribution applied to an “uncertainty dataset” (ideally containing near-OOD or adversarial examples).
• The overall objective is

$$\mathcal{L}_{\mathrm{DCM}}(f) = \mathcal{L}_{\mathrm{CE}}(f,\mathcal{T}) + \lambda \mathcal{L}_{\mathrm{conf}}(f,\mathcal{U}),$$

where $\mathcal{T}$ is the in-distribution set and $\mathcal{U}$ is the held-out/uncertainty set.

This formulation is theoretically justified: under mild assumptions, it provably yields a model where maximum-softmax-probability (MSP) on OOD inputs is strictly lower than on in-distribution examples, enabling thresholding-based separation. Empirically, DCM markedly reduces OOD false positive rates and boosts calibration and risk-coverage metrics for selective classification compared to prior approaches (Choi et al., 2023).

5. Adaptive Mixing and Meta-learning of Auxiliary Losses

The presence of multiple auxiliary losses—possibly representing different sources of weak supervision, noisy pseudo-labels, or confidence regularizers—necessitates optimal weighting. The AMAL framework (Sivasubramanian et al., 2022) introduces bi-level meta-learning to adaptively mix primary and $K$ auxiliary losses at an instance level.

• The training objective is

$$\mathcal{L}_{\mathrm{train}}(\theta,\Lambda) = \frac{1}{N} \sum_{i=1}^N \left[\lambda_{p,i}\mathcal{L}_p(x_i, y_i; \theta) + \sum_{k=1}^K \lambda_{a_{k,i}} \mathcal{L}_{a_k}(x_i; \theta)\right],$$

where $\Lambda$ are learned mixing weights.

• Weights $\Lambda$ are optimized to minimize a validation loss after a look-ahead SGD step on $\mathcal{L}_{\mathrm{train}}$; gradients are backpropagated through the update. This procedure allows AMAL to adapt loss composition dynamically, automatically down-weighting noisy or uninformative auxiliary signals depending on their downstream impact.

AMAL has demonstrated improved test accuracy in knowledge distillation, rule-denoising, and noisy-label scenarios, outperforming fixed coefficients and previous self-supervision meta-learning baselines (Sivasubramanian et al., 2022).

6. Empirical Performance and Application Domains

Auxiliary confidence losses have demonstrated concrete improvements across diverse problem settings:

Setting	Main Auxiliary Loss Paradigm	Impact/Results
LLP (L²P-AHIL)	Entropy-weighted instance loss	≤3% accuracy drop as bag size ↑; 40–70% drop in baselines (Ma et al., 2024)
Failure detection (ConfidNet)	TCP regression (MSE)	AUPR ↑53.7% (CIFAR-10/VGG-16); FPR@95↓ by 4pp over MCP (Corbière et al., 2020)
UDA segmentation (ConDA)	TCP regression + adversarial loss	mIoU gains: 45.5→49.9% (GTA5→Cityscapes) (Corbière et al., 2020)
OOD/selective (DCM)	Minimize uniform-entropy on OOD	FPR@95 drop by 6.3pp (CIFAR-10) and 58.1pp (CIFAR-100) (Choi et al., 2023)
KD/rule-based (AMAL)	Meta-learned instance mixing	Top-1 accuracy ↑2–4% vs KD; automatic down-weighting of noisy signals (Sivasubramanian et al., 2022)

In all cases, the key contribution is an auxiliary objective that either calibrates, estimates, or differentially weights examples by their confidence, leading to improved generalization, robustness, and error detectability.

7. Practical Considerations and Hyperparameter Selection

• Balancing coefficients ($\alpha$, $\lambda$): Empirically, settings of 0.5 for auxiliary loss weights generally perform well; excessive emphasis can risk overfitting to pseudo-label noise (Ma et al., 2024, Choi et al., 2023).
• Entropy-to-weight smoothness ($\beta$ in L²P-AHIL): Must be tuned according to label space size (e.g., β=1 for 10-class, β=5 for 100-class). Too small—good pseudo-labels underweighted; too large—admits low-confidence labels (Ma et al., 2024).
• Frequency of meta-updates: For AMAL, updating every 5–10 epochs is sufficient for stable gains without major computational overhead (Sivasubramanian et al., 2022).
• Uncertainty set design (DCM): Inclusion of near-OOD or representative hard examples is critical. Filtering of ID examples from $\mathcal{U}$ is unnecessary if the theoretical assumptions hold (Choi et al., 2023).

• "Forming Auxiliary High-confident Instance-level Loss to Promote Learning from Label Proportions" (Ma et al., 2024)
• "Confidence Estimation via Auxiliary Models" (Corbière et al., 2020)
• "Conservative Prediction via Data-Driven Confidence Minimization" (Choi et al., 2023)
• "Adaptive Mixing of Auxiliary Losses in Supervised Learning" (Sivasubramanian et al., 2022)