Committee Disagreement

Committee Disagreement is a measure of predictive uncertainty calculated from the variance or entropy across the outputs of an ensemble of models (the 'committee'). It is the foundational criterion in Query-By-Committee (QBC). The algorithm selects for labeling the data instances where committee members' predictions are most divergent, under the principle that resolving this disagreement will provide maximal information to improve the ensemble.

• Operational Principle: High disagreement indicates regions of the input space where the committee's collective knowledge is uncertain or contradictory.
• Mathematical Forms: Commonly measured via vote entropy (entropy over class votes) or average Kullback-Leibler (KL) divergence between individual model predictions and the committee consensus.

Disagreement is formalized using information-theoretic or statistical measures over the committee's predictions for a given data point.

• Vote Entropy: For classification, let (V(c)) be the number of committee members voting for class (c). Vote entropy is calculated as (-\sum_c \frac{V(c)}{M} \log \frac{V(c)}{M}), where (M) is committee size. High entropy indicates a split vote.
• Average KL Divergence: (\frac{1}{M} \sum_{m=1}^{M} D_{KL}(P_m(y|x) || \bar{P}(y|x))), where (P_m) is a member's predictive distribution and (\bar{P}) is the committee average. This measures how much each member's prediction differs from the consensus.
• Simple Margin: The difference between the votes for the two most popular classes. A small margin indicates high disagreement.

Committee Disagreement functions as the acquisition function within an iterative active learning loop for data streams.

1. Stream Processing: As each new unlabeled instance (x_t) arrives, the committee generates predictions.
2. Disagreement Scoring: The chosen metric (e.g., vote entropy) computes a disagreement score for (x_t).
3. Query Decision: If the score exceeds a threshold (or is among the top-k for a batch), the system queries an oracle (e.g., human expert) for the true label (y_t).
4. Committee Update: All models in the committee are retrained or updated on the new ((x_t, y_t)) pair. This cycle directly targets the reduction of uncertainty in the most contentious regions of the feature space.

Using a committee provides distinct benefits compared to uncertainty sampling with a single model.

• Mitigates Approximation Bias: Single-model uncertainty methods (like Monte Carlo Dropout) rely on approximations of a Bayesian posterior. A diverse committee can empirically capture a broader spectrum of plausible hypotheses.
• Explicit Hypothesis Diversity: Disagreement directly measures variation between different learned models, which can be more robust than the internal uncertainty of a single, potentially overconfident, model.
• Compatibility with Non-Probabilistic Models: Committee methods can be applied to models that don't natively output calibrated probabilities (e.g., standard SVMs or decision trees) by using their discrete predictions.

Deploying Query-By-Committee in production streams involves several engineering considerations.

• Committee Construction: Members must be diverse to ensure meaningful disagreement. Techniques include training on different data bootstraps, using varied model architectures, or initializing with different random seeds. Lack of diversity leads to unanimous, uninformative agreements.
• Computational Overhead: Maintaining and performing inference with (M) models increases compute and memory costs by a factor of (M) compared to a single model. This is a key trade-off for latency-sensitive streams.
• Streaming Adaptation: In non-stationary environments with concept drift, the committee must adapt. Strategies include weighting members based on recent performance or periodically retraining/replacing underperforming members to maintain relevant disagreement on new data concepts.

Committee Disagreement connects to several key strategies and frameworks within active and continuous learning.

• Bayesian Active Learning: QBC is a practical, non-parametric approximation of a fully Bayesian approach, where the committee samples represent the hypothesis space.
• Ensemble Learning: It leverages ensemble methods not for improved predictive accuracy via averaging, but for the explicit purpose of measuring model variance.
• Drift-Aware Querying: In Stream-Based Active Learning, the disagreement threshold can be dynamically adjusted when concept drift is detected, prioritizing queries from the new data distribution.
• Batch Mode Active Learning: When selecting a batch of queries, a diversity criterion is often combined with disagreement to avoid selecting multiple highly similar, redundant points.

Query-By-Committee is the active learning framework where Committee Disagreement is the primary acquisition function. It operates by:

• Maintaining an ensemble of diverse models (the committee).
• For each unlabeled data point, having each committee member make a prediction.
• Calculating the disagreement (e.g., via vote entropy or KL divergence).
• Selecting the point with the highest disagreement for labeling, as it represents the region of the input space where the model committee is most uncertain and knowledge gain is maximal.

Uncertainty Sampling is the broader family of query strategies to which Committee Disagreement belongs. It selects points where a model's prediction is least confident. Key methods include:

• Least Confidence: Querying the instance where the model's top predicted class has the lowest probability.
• Margin Sampling: Selecting points where the difference between the top two class probabilities is smallest.
• Entropy-based Sampling: Choosing points where the predictive class distribution has the highest entropy. Committee Disagreement is an ensemble-based extension of these single-model uncertainty measures.

Bayesian Active Learning provides a probabilistic foundation for uncertainty estimation, an alternative to the committee approach. It uses:

• Bayesian Neural Networks (BNNs): Models with distributions over weights, where predictive uncertainty is derived from the posterior.
• Monte Carlo Dropout: A practical approximation that performs multiple stochastic forward passes to estimate model uncertainty.
• Acquisition functions like Bayesian Active Learning by Disagreement (BALD), which seeks data points that maximize the mutual information between model parameters and the prediction. This offers a theoretically grounded counterpart to heuristic committee measures.

Batch Mode Active Learning addresses the practical need to query labels for multiple points simultaneously. When using Committee Disagreement in batch settings, a naive top-k selection can lead to redundant queries. Advanced strategies incorporate:

• Diversity Measures: Using techniques like Core-Set Selection or k-Means++ on the uncertain points to ensure the batch covers diverse regions of the feature space.
• Submodular Optimization: Framing batch selection as maximizing a submodular function that balances high disagreement with representativeness. This is critical for efficient label acquisition in production systems where oracle throughput is limited.

Expected Model Change is an alternative, more computationally intensive acquisition function that looks beyond predictive uncertainty. It selects the data point expected to induce the largest change to the model parameters (e.g., the greatest gradient magnitude) if its true label were known. Compared to Committee Disagreement:

• Committee Disagreement measures hypothesis space disagreement.
• Expected Model Change estimates parameter space impact. It is particularly useful when the learning objective is sensitive to specific parameter updates, but requires calculating gradients for all unlabeled candidates, which can be prohibitive for large models.

Drift-Aware Querying is a specialized strategy for non-stationary data streams, directly relevant to Committee Disagreement in online settings. It modifies the query criterion to:

• Prioritize labeling instances that are not only uncertain but also likely from a new or changing data distribution (concept drift).
• Dynamically adjust the committee or uncertainty measure in response to detected drift.
• Balance exploration of new concepts with exploitation of current knowledge. This ensures the active learning system remains effective as the underlying task evolves, preventing the committee from focusing on outdated regions of disagreement.