Confidence-Normalized Pooling Rule

• The confidence-normalized pooling rule is a strategy that aggregates instance predictions by weighting them with learned or computed confidence scores for improved reliability.
• It employs mathematical formulations like softmax weighting and power pooling, enabling adaptive thresholding and flexible aggregation across diverse datasets.
• This approach has shown significant gains in areas such as semi-supervised sound event detection and digital pathology by leveraging all available information and ensuring explainability.

The confidence-normalized pooling rule encompasses a class of pooling and aggregation strategies in weakly supervised and multiple-instance learning scenarios, where individual prediction scores are combined using weights determined by learned or computed confidence estimates. This paradigm enables models to leverage all available data, including uncertain or unlabeled instances, by explicitly modulating the contribution of each prediction based on a principled measure of reliability. Such pooling approaches have demonstrated empirical superiority in domains such as semi-supervised sound event detection and digital pathology, outperforming naive mean or max-pooling by enabling flexible, explainable, and robust instance aggregation.

1. Mathematical Formulation of Confidence-Normalized Pooling

Confidence-normalized pooling aggregates instance-level predictions $p_i$ using weights $c_i$ that capture either the model's learned confidence (as in semi-supervised sound event detection) or the model's estimated certainty (e.g., via MC-dropout). The general form for combining $N$ predictions is: $Z = \frac{\sum_{i=1}^N c_i p_i}{\sum_{j=1}^N c_j}$ where $c_i \in (0,1)$ represents the confidence associated with prediction $p_i$ (Liu et al., 2020, Gildenblat et al., 2020).

A related variant involves a learned exponent $n$ (power-pooling), producing a nonlinear aggregation: $$y_c = \frac{\sum_i y_f(i)^{n+1}}{\sum_i y_f(i)^{n}}$$ recovering mean-pooling ($n=0$), linear pooling ($n=1$), and max-pooling as $n \to \infty$.

Further generalization applies softmax weighting over confidences: $$w_i = \frac{\exp(\tau c_i)}{\sum_j \exp(\tau c_j)}, \qquad Z = \sum_i w_i p_i$$ incorporating temperature $\tau > 0$ to modulate sensitivity (Gildenblat et al., 2020).

2. Learning and Estimating Confidence

In confidence-normalized pooling, $c_i$ can be learned as an auxiliary output of the model (as in C-SSED) or estimated from predictive uncertainty (as in Certainty Pooling for MIL):

Learned Confidence in C-SSED:

A parallel confidence head predicts $c_i = \sigma(W_c h_i + b_c)$ for each time frame, where $h_i$ is the feature vector and $\sigma$ is the sigmoid. To prevent trivial solutions ($c_i=0$ everywhere), a hint-penalty loss $L_c = -\sum_i \log c_i$ is imposed, balancing fidelity and informativeness. Confidence modulates the reliance on ground-truth ($t_f(i)$) versus predicted ($y_{fs}(i)$) labels via: $y'_{fs}(i) = (1-c_i) t_f(i) + c_i y_{fs}(i)$ During self-training, confidence weights framewise contributions to the retraining loss, ensuring unreliable pseudo-labels are down-weighted and rare true negatives retained.

Certainty in MIL via MC-dropout:

Certainty $c_i$ is computed as the inverse of the predictive standard deviation: $c_{i} = \frac{1}{\sigma_{i}+\epsilon}$ where $\sigma_i$ is the standard deviation over $T$ MC-dropout stochastic forward passes, and $\epsilon$ is a small constant. These certainties define per-instance weights for pooling, or in the "hard" variant, the instance with maximal $c_i p_i$ is selected as the representative bag score (Gildenblat et al., 2020).

3. Empirical Behavior and Threshold Flexibility

Confidence-normalized pooling enables adaptive instance-thresholding within the aggregation process. In power pooling, the effective threshold is $\theta = n/(n+1)$, which can be tuned or learned through gradient-based optimization, permitting transitions between mean-like and max-like pooling. This is critical for tasks where events vary in duration: brief events benefit from near-max pooling, while extended phenomena are better captured via mean pooling (Liu et al., 2020).

Confidence weighting in self-training avoids the typical exclusion of low-probability (and potentially informative) frames; all data contribute with their influence scaled according to confidence. This approach has empirically reduced error rates by up to 34% over baseline models in semi-supervised sound event detection, and by 20% when coupled with self-training (Liu et al., 2020).

4. Applications in Sound Event Detection and Multiple Instance Learning

The confidence-normalized pooling rule is central to recent advances in semi-supervised sound event detection (C-SSED, by Liu et al.) and multiple-instance learning (Certainty Pooling):

• C-SSED (Sound Event Detection): Combines trainable power pooling with confidence-normalized weighting, achieving superior error rates and F$_1$ compared to linear pooling and attention. The method retains all data, modulating their influence via learned confidence, and dynamically selects pooling nonlinearity per dataset/class (Liu et al., 2020).
• Certainty Pooling (MIL): Aggregates instance scores in large, low-evidence bags using certainty weights, enhancing robustness and explainability. Statistically, certainty pooling delivers the highest bag and instance AUCs in both synthetic (MNIST bags) and real-world pathology datasets (Camelyon16) (Gildenblat et al., 2020).

Pooling Method	Sound Event Detection F$_1$	MIL Bag-level AUC	Instance-level AUC
Attention/Average	32.04%	0.56–0.82	0.63–0.75
Linear	34.27%	0.72–0.90	0.70–0.72
Power/Certainty	37.04%	0.88–0.93	0.80–0.77

5. Implementation and Training Strategies

Training with a confidence-normalized pooling rule involves:

• Initialization of model (student/teacher), pooling exponent $n$, and confidence head.
• Forward propagation of input instances with computation of scores and confidences.
• Aggregation via weighted pooling (linear, power, or certainty pooling).
• Application of multitask losses: clip-level, frame-level, hint penalty, and consistency losses.
• Joint optimization of model parameters and pooling operator (including $n$).
• In MIL, MC-dropout is used to estimate instance certainties, with backpropagation weighted accordingly.

Pseudocode for MIL Certainty Pooling with MC-dropout (adapted from (Gildenblat et al., 2020)):

for each bag B_m: for i in B_m: for t in 1...T: p_i^{(t)} = model(x_i) # with dropout mu_i = mean(p_i^{(t)}) sigma_i = std(p_i^{(t)}) c_i = 1/(sigma_i + epsilon) w_i = softmax(tau * c_i) Z_m = sum(w_i * mu_i) loss = Loss(Z_m, Y_m) update model

6. Robustness, Explainability, and Broader Impact

Confidence-normalized pooling enhances robustness to noisy and ambiguous instances by down-weighting those with high uncertainty, reducing the risk of spurious activations from outlier instances. The explicit use of confidence as pooling weights provides interpretable decision pathways, as aggregated outputs can be traced to highly certain component predictions, facilitating explainability in domains where trust and traceability are crucial (e.g., medical imaging).

The technique has been recommended for a wide class of multiple-instance learning problems, particularly where evidence ratios are low and bag sizes are large. The effectiveness of learned confidence heads and certainty weighting generalizes to self-training and pseudo-labeling pipelines, making it suitable for weakly labeled data beyond sound event detection and pathology (Liu et al., 2020, Gildenblat et al., 2020).

7. Limitations, Hyperparameter Sensitivity, and Extensions

Computing instance certainties via MC-dropout imposes additional inference overhead; common settings use $T=10$–30 stochastic passes. Hyperparameters such as temperature $\tau$, $\epsilon$, dropout fraction, and number of MC draws require empirical tuning. If instance certainties are uniformly low, pooling reduces to the mean, yielding no deleterious effect.

Extensions include defining $c_i$ via entropy or mutual information, learning parametric uncertainty-to-weight mappings, or gating classical attention weights by certainty for hybrid pooling. The approach conceptually unifies hard-max and soft-mean pooling, providing a tunable middle ground for both robustness and explainability.

In summary, the confidence-normalized pooling rule integrates principled confidence metrics into permutation-invariant pooling, yielding substantial performance and interpretability gains over standard approaches in weakly supervised, semi-supervised, and multiple-instance learning contexts (Liu et al., 2020, Gildenblat et al., 2020).