Abstract
Recursive Bayesian Classification (RBC) is a fundamental framework for sequential evidence accumulation in fields ranging from radar tracking to Brain-Computer Interfaces (BCI). A critical challenge in RBC is the design of a stopping criterion that optimally balances decision speed and accuracy. Traditional methods, such as fixed confidence thresholds or entropy-based uncertainty measures, often fail to account for the geometric evolution of the posterior distribution. This paper introduces a novel stopping criterion derived from information geometry. By proposing a "bent-in" decision boundary within the probability simplex, the authors enable faster decisions when no strong competitors exist. Experimental results on synthetic data and real-world BCI datasets demonstrate that this approach significantly improves the speed-accuracy trade-off, achieving up to a 12% reduction in decision time without compromising performance.
Introduction
In sequential data acquisition, we use Recursive Bayesian Classification (RBC) to update our posterior distribution as new evidence arrives. However, every data point comes with a cost—be it time, energy, or user fatigue. Deciding exactly when to stop and commit to a classification is the "stopping problem."
Existing stopping criteria generally fall into two categories:
- Confidence-based (M1): Stop when the maximum posterior probability exceeds a threshold (e.g., $p > 0.95$). These are often too slow, especially if the initial prior is far from the true class.
- Uncertainty-based (M3): Stop when the entropy of the distribution falls below a threshold. These can be "fragile"; in multi-class problems, quickly ruling out several unlikely candidates can cause entropy to plummet, triggering a premature stop even if the remaining top candidates are still closely tied.
The authors argue that these methods ignore the geometric trajectory of the posterior as it moves through the probability simplex.
Geometric Analysis of Posterior Progression
The core of the proposed method lies in analyzing the "posterior progression" as a path through an $n$-dimensional simplex (the space of all possible probability distributions). The authors prove that, on average, the posterior moves along a "central path" toward the correct vertex.
As shown in the figure above, the trajectory of the posterior tends to move from the center (uniform distribution) toward the target vertex. Understanding this geometry allows for the design of boundaries that are more sensitive to the relative strength of competing hypotheses.
The Proposed MP Criterion
The authors propose a new criterion, denoted as MP, which utilizes a decision boundary that "bends" toward the center of the simplex. Mathematically, this is implemented using a modified delta divergence.
A key practical outcome of this geometric approach is a "Top-Two" rule: the system stops when the difference between the two most likely classes ($p_{j1} - p_{j2}$) exceeds a specific threshold. This ensures a decision is only made when a clear leader emerges from its closest competitor, regardless of the number of other "loser" classes.
As illustrated in the figure above, the MP boundary (red) "tilts" toward the uniform distribution compared to the standard M1 boundary (blue). This allows the system to stop much earlier when there is no "close competitor," while maintaining high accuracy requirements when two classes are neck-and-neck.
Experimental Evaluation and Results
The researchers validated the MP criterion using two distinct approaches:
- Synthetic Simulations: 5,000 Monte Carlo simulations using lognormal evidence distributions mapped the speed-accuracy trade-off across 3 to 10 classes. The MP method consistently provided a superior Pareto front, allowing for faster decisions at the same accuracy level as conventional methods.
- Real-World BCI Data: The method was tested on data from 10 participants using an RSVP Keyboard, an EEG-based typing system. The MP criterion completed the typing tasks while saving approximately 10-12% of the required sequences compared to the standard confidence method (M1). In practice, this translates to significant time savings for users.
Conclusion
By moving beyond simple scalar thresholds and considering the information geometry of the probability simplex, the authors have developed a more efficient stopping criterion for RBC. The "bent-in" boundary of the MP criterion offers a robust solution that adapts to the presence or absence of competing hypotheses, providing a significant performance boost in time-sensitive applications like BCI. Future work will explore higher-order curves for the stopping boundary and analytical models for specific evidence distributions.
Summary updated