Evidence ratio classifier: A one-pass model for fast incremental learning

We introduce the Evidence Ratio Classifier (ERC), a one-pass learning rule in which every weight is a closed-form ratio of empirical co-occurrence counts. Training therefore reduces to a single linear scan of the data, and inference to a few table look-ups followed by a winner-takes-all comparison; no back-propagation, learning-rate tuning, or other iterative optimisation is required. ERC is benchmarked on four standard tabular data sets— Adult Income , Mushroom , Iris , and Credit-Card Fraud . On Adult ( training instances) it reaches accuracy versus for a variational Bayesian neural network and for Gaussian Naive Bayes, while using roughly fewer multiply–add operations at inference. Across the other tasks ERC matches or exceeds Naive Bayes and stays within one percentage point of the variational baseline, with lower arithmetic cost. Because ERC stores, for every input pattern , the empirical conditional probability , its lookup table is self-explanatory: each entry quantifies the data support for the rule . This intrinsic transparency is attractive wherever auditability is required, e.g. credit scoring and anti-money-laundering (Basel III; EU AI Act, Art. 13), medical decision support (FDA SaMD guidance), insurance underwriting, or judicial risk assessment. Model updates consist solely of incrementing integer counters; no gradients or floating-point arithmetic are involved. Such integer-only updates execute on processors that lack a hardware FPU or must remain fully deterministic, as found in ultra-low-power micro-controllers, hard real-time UAV and robotic controllers, or certified safety-critical systems (DO-178C avionics, CENELEC EN 50,128 railway, IEC 62,304 medical implants). ERC thus combines accuracy, explainability, and hardware efficiency in a single, making it deployable on resource–constrained hardware. Classic gradient-based network (left) versus ERC (right). ERC converts one-pass co-occurrence counts into conditional-probability weights and predicts with a winner-takes-all rule—no back-prop, no floating-point math. • Introduce Evidence Ratio Classifier: single-pass Bayesian learning, no back-prop. • Training uses 30 fewer ops; inference needs look-ups vs 8-epoch VBNN. • Sparse motif topology found on-the-fly via support–selectivity pruning. • Each weight stores , giving intrinsic, audit-ready explanations. • Counter-only updates fit integer-only MCUs, FPGAs and neuromorphic accelerators.