K-ABENA: K-Adaptive Backpropagation with Error-based N-exclusion Algorithm
Compensated Loss-Based Sample Exclusion with Unbiased Gradient Estimation

Jean-François Bonbhel
NeuroSoft IA, Québec City, Canada \(|\) YekoElite University, Brazzaville, Republic of Congo
UN AI Governance Expert Network (UN PNAI) — Member since 2021
bonbhel@yekoelite.com


Abstract

We present K-ABENA (K-Adaptive Backpropagation with Error-based N-exclusion Algorithm), a selective gradient computation framework that reduces per-iteration training cost by excluding a fraction of low-loss (“minor”) observations from the backward pass. Its canonical form (v3) combines a defensive-mixture sampling design over the minor set with Horvitz–Thompson inverse-probability reweighting, yielding a design-unbiased Horvitz–Thompson gradient estimator (Lemma [lem:unbiased]) and whose self-normalized practical variant carries a bias of order \(O(1/m)\) with an explicit constant (Lemma [lem:hajek]). We prove an \(O(1/\sqrt{T})\) non-convex convergence guarantee for SGD under the estimator, with an additive term that quantifies the residual bias (Theorem [thm:sgd]). We further prove that uncompensated loss-based selection — a family that includes OHEM, SBP, and the two earlier K-ABENA variants — admits no stationary point at any minimizer where its selection bias is bounded away from zero (Proposition [prop:failure]), and we quantify this failure empirically: at 0.17% class imbalance, uncompensated variants reach test AUC 0.53–0.62 versus 0.9998 for full-batch SGD, while the compensated estimator attains 0.9991 at identical 28.4% compute savings. On real datasets (Breast Cancer, Digits, Wine, Diabetes) the compensated estimator is statistically indistinguishable from full-batch SGD (paired permutation tests, \(p \ge 0.5\); Section sec:sec:sec:experiments?) while saving 28–54% of per-epoch gradient computation. A biased “regularized mode” (the earlier half-domain variant) is retained as an option with a proven exact bias decomposition (Lemma [lem:decomp]) and quantified contraindications: it collapses to 0.386 accuracy under 40% label noise (baseline: 0.832) and to 0.53 AUC under extreme imbalance. Every advantage and every limitation reported in this paper is either proved or measured; all experiments are CPU-scale (NumPy/scikit-learn) and their scope is stated explicitly.

1 Introduction↩︎

In large-scale empirical risk minimization, a substantial fraction of per-iteration computation is spent on observations the model has already learned: their per-sample losses are small, their gradients are small, and their marginal contribution to the descent direction is limited. Selective-backpropagation methods exploit this observation by skipping the backward pass for low-loss samples [1], [2], but they share a structural defect: the retained subset is correlated with the loss, so the resulting gradient is a biased estimator of the full-batch gradient. In benign regimes the bias is small relative to the signal and these methods work well; we show in Section 5 that in adverse regimes — extreme class imbalance, heavy label noise — the bias does not merely degrade performance but structurally prevents convergence to the minimizer, and we prove this as Proposition 2.

This paper develops K-ABENA, whose canonical estimator (referred to as v3) resolves the defect using a century-old idea from survey sampling [3]: any sampling design with known, strictly positive inclusion probabilities admits an unbiased estimator of a population total via inverse-probability weighting. K-ABENA v3 samples retained minors from the entire minor set under a defensive mixture design [4] and reweights them accordingly. The result occupies a design point that, to our knowledge, none of the established selective or reweighting methods occupies: per-iteration compute reduction with an (exactly or near-) unbiased gradient. Hard-selection methods (OHEM [2], SBP [1]) save compute but are biased; soft-reweighting methods (Focal Loss [5]) are computed on the full batch and save nothing; importance-sampling training methods [6] reweight but target variance reduction under full or minibatch evaluation rather than threshold-based exclusion with an explicit retention budget.

1.0.0.1 Contributions.

  1. A compensated selective estimator (Definition 1) with two interpretable controls: a loss threshold \(K\) separating “minor” from “major” observations, and a retention proportion \(N \in (0,1)\) governing the compute budget. The per-epoch backward-pass saving is exactly \((1-N)\,k/n\) (Proposition 1), where \(k\) is the number of minors.

  2. Design-unbiasedness, exact and approximate (Lemmas 23): the Horvitz–Thompson form is design-unbiased (unbiased over the sampling randomization, conditionally on the iterate); the self-normalized (Hájek) form used in practice has bias at most \(\frac{2 G_M}{\alpha^2 m}\bigl(\tfrac{k-m}{k}\bigr)\), vanishing as \(O(1/m)\), with all constants explicit.

  3. A convergence guarantee (Theorem 1): \(O(1/\sqrt{T})\) decay of the expected squared gradient norm for smooth non-convex objectives, plus an additive \(O(\delta_m)\) floor traceable to the self-normalization bias, which the practitioner controls through \(m\).

  4. An impossibility result for uncompensated selection (Proposition 2): if the selection bias at a minimizer is bounded below by \(\beta > 0\), the minimizer is not a stationary point of the expected dynamics; measured on a 0.17%-imbalance task, \(\beta \approx 0.15\) against a vanishing true gradient, predicting the observed failure (AUC 0.53) and its resolution under compensation (AUC 0.9991).

  5. A characterized regularized mode: the earlier biased half-domain variant (v2) is retained as an option, with an exact algebraic bias decomposition (Lemma 5) and quantified contraindications (Section 6); its small accuracy bonus on multiclass tasks (\(+0.35\) to \(+0.45\) points) and its catastrophic failure modes (0.386 accuracy at 40% label noise; 0.53 AUC at 0.17% imbalance) are both measured and mechanistically explained.

1.0.0.2 Scope statement.

All empirical results in this paper are CPU-scale: logistic regression, linear SVM, softmax regression, and a one-hidden-layer MLP, on bundled scikit-learn datasets and on a synthetic replica of an extreme-imbalance fraud regime (0.17% positives, \(n = 20{,}000\)). No GPU benchmark (deep CNN/Transformer) is reported; this is stated as Limitation L4 rather than compensated by simulation. We consider explicit scope statements a feature of the paper, not a weakness.

2 Related Work↩︎

Hard selection. OHEM [2] retains only the highest-loss examples of each batch; SBP [1] drops examples whose loss falls below a threshold. Both reduce backward-pass cost and both compute a plain average over a loss-correlated subset, hence a biased gradient. K-ABENA strictly generalizes SBP: setting \(N = 0\) with a hard threshold recovers it, while \(N > 0\) restores representation of the excluded stratum and the v3 weighting removes the bias. Curriculum and self-paced approaches [7], [8] order examples by difficulty over training time; K-ABENA is orthogonal, operating within each iteration.

Soft reweighting. Focal Loss [5] down-weights easy examples by a factor \((1-p_t)^\gamma\) applied after the forward pass of every example; the full backward cost is still paid, so the method offers no compute saving — the quantity K-ABENA targets. Its modulation is also a single global functional form, whereas \(K\) and \(N\) decouple the location of the easy/hard boundary from the retention budget.

Importance sampling for SGD. [6] and related work sample training examples proportionally to (proxies of) gradient norm with inverse-probability corrections, targeting variance reduction. Three structural differences separate K-ABENA from this line, and each is measurable. (a) Scoring cost. Classical IS must score all \(n\) candidates at each step to build its proposal — a forward pass (or a last-layer gradient-norm bound, itself requiring the forward) whose cost remains \(O(n)\); the backward saving is therefore bought with a full-width scoring pass. K-ABENA scores with the per-sample losses the training loop already produces, and its delayed-losses pattern (masks for epoch \(t\) built from epoch \(t{-}1\) losses; released with the library) removes even the forward of excluded samples. (b) Budget semantics. IS fixes a minibatch size and optimizes variance at that size; K-ABENA’s two controls decouple where the easy/hard boundary lies (\(K\), a loss percentile with a semantic reading) from how much budget is spent (\(N\)), yielding the deterministic, architecture-independent saving \(G = (1-N)k/n\) (Proposition 1). (c) Design. K-ABENA is a two-stratum design — a certainty stratum (majors, \(\pi_i = 1\)) plus a sampled stratum restricted to minors under a defensive mixture with floor \(\alpha/k\) — rather than a single global proposal. This is not cosmetic: at matched compute budget, a classical global loss-proportional proposal with self-normalized correction is significantly worse than the full-batch baseline in the standard regime (accuracy \(0.9573\) vs \(0.9720\), paired permutation \(p = 0.002\), Table 3) — tiny-loss samples occasionally drawn under a global proposal receive enormous weights, and the majors’ contribution is needlessly randomized — while the stratified, floor-bounded design is statistically indistinguishable from the baseline (\(p = 1.0\)) at the same saving. The certainty stratum and the defensive floor are what make compensation cheaply stable.

Survey sampling. The estimators are classical: Horvitz–Thompson [3] and its self-normalized (Hájek) variant [9]. Our contribution is not the estimator but its integration into threshold-based selective backpropagation, the resulting theory (Theorem 1, Proposition 2), and the quantified characterization of when the uncompensated shortcut is and is not safe.

3 The K-ABENA Framework↩︎

3.1 Setting and notation↩︎

Let \(F(\theta) = \frac{1}{n}\sum_{i=1}^{n} f_i(\theta)\) with \(f_i : \mathbb{R}^d \to \mathbb{R}\) differentiable, and write \(g_i = \nabla f_i(\theta)\) and \(\ell_i = f_i(\theta) \ge 0\) for the per-sample gradient and loss at the current iterate. At each iteration, a threshold \(K > 0\) (in practice a fixed percentile of the current loss distribution) partitions the indices into the minor set \(\mathcal{M}= \{ i : \ell_i \le K \}\), \(k = |\mathcal{M}|\), and the major set \(\mathcal{M}^c\), \(|\mathcal{M}^c| = n - k\). Majors are always retained. A retention proportion \(N \in (0,1)\) fixes the number of retained minors \(m = \lfloor N k \rceil\).

Definition 1 (Canonical K-ABENA sampling design, v3). Fix a defensive-mixture coefficient \(\alpha \in (0,1]\). Draw \(\mathcal{S}\subset \mathcal{M}\), \(|\mathcal{S}| = m\), without replacement with single-draw probabilities \[p_i \;=\; \alpha \cdot \frac{1}{k} \;+\; (1-\alpha)\,\frac{\ell_i}{\sum_{j \in \mathcal{M}} \ell_j}, \qquad i \in \mathcal{M}, \label{eq:design}\qquad{(1)}\] and denote by \(\pi_i = \Pr[i \in \mathcal{S}]\) the inclusion probabilities of the design. The Horvitz–Thompson (HT) gradient estimator is \[\hat{g}_{\mathrm{HT}} \;=\; \frac{1}{n}\Bigl[\; \sum_{i \notin \mathcal{M}} g_i \;+\; \sum_{i \in \mathcal{S}} \frac{g_i}{\pi_i} \;\Bigr], \label{eq:ht}\qquad{(2)}\] and the self-normalized (Hájek) estimator, used in practice, is \[\hat{g} \;=\; \frac{\sum_{i \notin \mathcal{M}} g_i \;+\; \sum_{i \in \mathcal{S}} \pi_i^{-1} g_i}{(n-k) \;+\; \sum_{i \in \mathcal{S}} \pi_i^{-1}}. \label{eq:hajek}\qquad{(3)}\]

Assumption 1 (Inclusion probabilities). Either (a) \(\alpha = 1\), in which case the design is simple random sampling without replacement and \(\pi_i = m/k\) exactly; or (b) \(\alpha < 1\) and \(\max_i m\,p_i \le 1\), in which case we use the Poisson/rejective approximation \(\pi_i = m\,p_i\), standard for high-entropy without-replacement designs [9]. All statements below that rely on case (b) are flagged; the empirical error attributable to this approximation is measured in Section 7 (residual bias 0.004 at \(n = 20{,}000\)).

Lemma 1 (Bounded weights under defensive mixing). Under Assumption 1, for every \(i \in \mathcal{M}\), \[\pi_i \;\ge\; \frac{\alpha m}{k}, \qquad\text{hence}\qquad \frac{1}{\pi_i} \;\le\; \frac{k}{\alpha m}.\]

Proof. From ?? , \(p_i \ge \alpha/k\) pointwise since the second term is non-negative. In case (a), \(\pi_i = m/k = \alpha m / k\) with \(\alpha = 1\). In case (b), \(\pi_i = m p_i \ge m\alpha/k\). Inverting gives the weight bound. ◻

Proposition 1 (Exact computational gain). The number of backward passes per iteration is \((n-k) + m\), so the saved fraction is \[G \;=\; \frac{n - \bigl[(n-k) + m\bigr]}{n} \;=\; \frac{k - m}{n} \;=\; (1 - N)\,\frac{k}{n},\] identical for the HT and Hájek forms; the reweighting itself costs \(O(m)\) scalar operations, negligible against a backward pass. With \(K\) at the 40th loss percentile and \(N = 0.3\), \(G = 0.7 \times 0.4 = 0.28\); the measured per-epoch saving in every experiment of Section 7 is 28.0–28.7%, matching the formula. 0◻

4 Estimator Guarantees↩︎

4.1 Unbiasedness↩︎

Lemma 2 (Design-unbiasedness of the HT form). Under Assumption 1, conditionally on the current iterate (hence on \(\ell_i\), \(\mathcal{M}\), \(\pi_i\)), \[\mathbb{E}\bigl[\hat{g}_{\mathrm{HT}}\bigr] \;=\; \nabla F(\theta).\]

Proof. Let \(Z_i = \mathbf{1}[i \in \mathcal{S}]\), so \(\mathbb{E}[Z_i] = \pi_i\). Then \[\mathbb{E}\Bigl[\sum_{i \in \mathcal{S}} \frac{g_i}{\pi_i}\Bigr] = \mathbb{E}\Bigl[\sum_{i \in \mathcal{M}} Z_i \frac{g_i}{\pi_i}\Bigr] = \sum_{i \in \mathcal{M}} \frac{\mathbb{E}[Z_i]}{\pi_i}\, g_i = \sum_{i \in \mathcal{M}} g_i .\] Substituting into ?? gives \(\mathbb{E}[\hat{g}_{\mathrm{HT}}] = \frac{1}{n}\bigl[\sum_{i \notin \mathcal{M}} g_i + \sum_{i \in \mathcal{M}} g_i\bigr] = \nabla F(\theta)\). Note that positivity \(\pi_i > 0\) for all \(i \in \mathcal{M}\) — guaranteed by the defensive term \(\alpha/k\) — is what makes the argument available; any design assigning zero probability to part of \(\mathcal{M}\) (Section 6) cannot be corrected this way. ◻

Lemma 3 (Bias of the self-normalized form). Let \(G_M = \max_{i \in \mathcal{M}} \left\lVert g_i \right\rVert\) and let \(\widehat W = (n-k) + \sum_{i \in \mathcal{S}} \pi_i^{-1}\) denote the (random) normalizer in ?? , with \(\mathbb{E}[\widehat W] = n\). Then \[\bigl\lVert \mathbb{E}[\hat{g}] - \nabla F(\theta) \bigr\rVert \;\le\; \frac{G_M}{n}\,\sqrt{\mathrm{Var}\bigl(\widehat W\bigr)} \;\le\; \frac{G_M\, k}{n\,\sqrt{\alpha\, m}},\] and the standard second-order ratio-estimator expansion sharpens this to \[\delta_m \;:=\; \bigl\lVert \mathbb{E}[\hat{g}] - \nabla F(\theta) \bigr\rVert \;\le\; \frac{2\, G_M\, k^{2}}{\alpha\, m\, n^{2}} + o(1/m) \;=\; O(1/m)\] at fixed \(k/n\) and \(\alpha\). At \(\alpha = 1\) (SRSWOR, exact \(\pi_i = m/k\)) the same expansion gives \(\delta_m \le 2\,G_M\,k\,(k-m)/(m\,n^{2}) + o(1/m)\). (Full derivation: Appendix 11.)

Remark 1 (What the practitioner controls). The bias floor \(\delta_m\) decreases in \(m = Nk\) and in \(\alpha\); it is measured, not merely bounded, in Section 7: 0.004 at \(n = 20{,}000\) (fraud replica, \(N = 0.3\)) and up to 0.037 in a deliberately adversarial micro-regime (\(k \approx 20\), \(m \approx 6\)). Both numbers sit far below the biases of the uncompensated variants in the same configurations (0.128–0.151).

4.2 Variance↩︎

Lemma 4 (Variance bound). Under Assumption 1(b), the HT estimator satisfies \[\mathbb{E}\bigl[\left\lVert \hat{g}_{\mathrm{HT}} - \nabla F(\theta) \right\rVert^2\bigr] \;\le\; \frac{1}{n^{2}} \sum_{i \in \mathcal{M}} \frac{1 - \pi_i}{\pi_i}\, \left\lVert g_i \right\rVert^{2} \;\le\; \frac{k}{\alpha\, m\, n^{2}} \sum_{i \in \mathcal{M}} \left\lVert g_i \right\rVert^{2} \;\le\; \frac{G_M^{2}\, k^{2}}{\alpha\, m\, n^{2}} \;=:\; \hat{\sigma}^{2}.\]

Proof. Under the Poisson approximation the \(Z_i\) are independent with \(\mathrm{Var}(Z_i) = \pi_i(1-\pi_i)\); write \(\hat{g}_{\mathrm{HT}} - \nabla F = \frac{1}{n}\sum_{i \in \mathcal{M}} (Z_i/\pi_i - 1) g_i\) and expand: \[\mathbb{E}\left\lVert \hat{g}_{\mathrm{HT}} - \nabla F \right\rVert^{2} = \frac{1}{n^{2}} \sum_{i \in \mathcal{M}} \frac{\mathrm{Var}(Z_i)}{\pi_i^{2}} \left\lVert g_i \right\rVert^{2} = \frac{1}{n^{2}} \sum_{i \in \mathcal{M}} \frac{1-\pi_i}{\pi_i} \left\lVert g_i \right\rVert^{2}.\] The middle inequality uses \(\pi_i \ge \alpha m / k\) (Lemma 1) and \(1 - \pi_i \le 1\); the last uses \(\left\lVert g_i \right\rVert \le G_M\) and \(|\mathcal{M}| = k\). ◻

Remark 2 (The price of unbiasedness, quantified). Compensation trades bias for variance. Measured at matched configuration on the fraud replica (Section 7): gradient variance \(1.5 \times 10^{-4}\) (v3) versus \(1.5 \times 10^{-5}\) (uncompensated half-domain variant) — a factor \(\approx 10\), against a bias reduction by a factor \(\approx 38\) (0.151 \(\to\) 0.004). The variance is bounded by Lemma 4 through \(\alpha\); a measured consequence is a higher AMSGrad max-lock frequency (57–60% of steps versus 35% uncompensated), discussed as Limitation L6.

4.3 Convergence under SGD↩︎

Assumption 2. \(F\) is \(L\)-smooth (\(\left\lVert \nabla F(\theta) - \nabla F(\theta') \right\rVert \le L \left\lVert \theta - \theta' \right\rVert\)), bounded below by \(F^{*}\), and along the trajectory the estimator satisfies \(\left\lVert \mathbb{E}[\hat{g}_t \mid \theta_t] - \nabla F(\theta_t) \right\rVert \le \delta\) and \(\mathbb{E}[\left\lVert \hat{g}_t - \mathbb{E}[\hat{g}_t \mid \theta_t] \right\rVert^{2} \mid \theta_t] \le \hat{\sigma}^{2}\), with \(\delta, \hat{\sigma}\) as in Lemmas 3 and 4, and \(\left\lVert \nabla F(\theta_t) \right\rVert \le G_\infty\).

Theorem 1 (Non-convex convergence of SGD with the v3 estimator). Run \(\theta_{t+1} = \theta_t - \eta\, \hat{g}_t\) for \(T\) steps with \(\eta = \min\{ \tfrac{1}{2L},\, \tfrac{c}{\sqrt{T}} \}\), \(c > 0\). Under Assumptions 1 and 2, \[\min_{0 \le t < T} \mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2} \;\le\; \underbrace{\frac{4\,(F(\theta_0) - F^{*})}{c\,\sqrt{T}} \;+\; \frac{2 L c\, (\hat{\sigma}^{2} + \delta^{2})}{\sqrt{T}}}_{\text{vanishing, } O(1/\sqrt{T})} \;+\; \underbrace{\vphantom{\frac{1}{\sqrt{T}}} 2\,\delta\, G_\infty + \delta^{2}}_{\text{bias floor}} .\] For the design-unbiased HT form (\(\delta = 0\)) this is the classical \(O(1/\sqrt{T})\) rate [10]; for the Hájek form the floor is \(O(\delta_m) = O(1/m)\), explicitly controlled by the retention budget. (Proof: Appendix 12.)

Remark 3 (Why \(\delta\) stays bounded along the whole trajectory). Assumption 2 posits a uniform \(\delta\); this is not an act of faith but a consequence of the calibration protocol. By Lemma 3, \(\delta_m \le 2 G_M k^2 / (\alpha m n^2)\) at every iterate, where only \(G_M\) and \(k\) depend on \(t\). First, \(K\) is recalibrated each epoch at a fixed percentile of the current losses, so \(k/n\) is constant by construction (here \(0.4\)) and \(m = Nk\) scales with it — the ratio \(k^2/(mn^2) = k/(N n^2)\) is time-invariant. Second, \(G_M\) is the maximum gradient norm over the minor stratum only: for the standard losses used here (logistic, softmax, squared), the per-sample gradient factorizes as \(\left\lVert g_i \right\rVert \le c(\ell_i)\,\left\lVert x_i \right\rVert\) with \(c(\cdot)\) bounded on the sublevel set \(\{\ell \le K\}\), so bounded (standardized) features give \(G_M(t) \le \bar G\) uniformly in \(t\) — low-loss samples cannot carry exploding gradients. Empirically, the measured bias is stable across training stages (0.004 at pre-trained iterates on the fraud replica; Table 4). For losses or architectures violating the factorization (unbounded features, exotic losses), Assumption 2 must be checked rather than assumed; we flag this within Limitation L5.**

5 Why Uncompensated Selection Fails: A Quantified Impossibility↩︎

Uncompensated loss-based selection — OHEM, SBP, and the earlier K-ABENA variants v1/v2 — computes a plain average over a loss-correlated subset. Write its conditional expectation as \(\mathbb{E}[\hat{g}_{\mathrm{sel}}(\theta)] = \nabla F(\theta) + b(\theta)\), where \(b(\theta)\) is the selection bias field.

Proposition 2 (No stationarity at the minimizer under persistent selection bias). Let \(\theta^{*}\) be a stationary point of \(F\) (\(\nabla F(\theta^{*}) = 0\)) and suppose \(\left\lVert b(\theta) \right\rVert \ge \beta > 0\) on a neighborhood \(U \ni \theta^{*}\). Then: (i) \(\theta^{*}\) is not a stationary point of the expected dynamics \(\dot{\theta} = -\mathbb{E}[\hat{g}_{\mathrm{sel}}(\theta)]\), whose expected update at \(\theta^{*}\) has magnitude \(\eta\beta\); (ii) any stationary point \(\bar\theta \in U\) of the expected dynamics satisfies \(\nabla F(\bar\theta) = -b(\bar\theta)\), hence \(\left\lVert \nabla F(\bar\theta) \right\rVert \ge \beta\): the biased method can only settle where the true gradient is as large as the bias.

Proof. (i) At \(\theta^{*}\), \(\mathbb{E}[\hat{g}_{\mathrm{sel}}(\theta^{*})] = 0 + b(\theta^{*}) \neq 0\) since \(\left\lVert b(\theta^{*}) \right\rVert \ge \beta\); the expected update is \(-\eta\, b(\theta^{*})\) of magnitude \(\eta \beta\). (ii) Stationarity of the expected dynamics means \(\nabla F(\bar\theta) + b(\bar\theta) = 0\), so \(\left\lVert \nabla F(\bar\theta) \right\rVert = \left\lVert b(\bar\theta) \right\rVert \ge \beta\). ◻

Remark 4 (Scope of Proposition 2: which biased methods it does and does not concern). Proposition 2 concerns the expected optimization dynamics under persistent selection bias* — its hypothesis is \(\left\lVert b(\theta) \right\rVert \ge \beta > 0\) on a neighborhood of the minimizer. Several well-known biased methods converge perfectly well precisely because they escape this hypothesis, in one of three ways. (i) Fixed reweightings define a modified objective: Focal Loss reweights every example by a fixed function of its own prediction, so its “bias field” is conservative — it is the gradient of a well-defined surrogate \(F_\gamma\), and training converges to the minimizer of \(F_\gamma\), not to a non-stationary drift. (ii) Vanishing bias: curriculum and self-paced schedules [7], [8], and prioritized-replay-style schemes with annealed importance corrections, expand the inclusion set (or anneal the correction) so that \(b(\theta_t) \to 0\) along training; the floor \(\beta\) does not exist. (iii) Useful displaced fixed points: a method may converge to a stationary point of the biased dynamics that generalizes well — our own v2 regularized mode is the documented example (Section 6), and hard-example mining in its benign regimes behaves similarly. The proposition targets the remaining case: threshold-based selection at a fixed loss percentile with a plain average, whose bias is non-vanishing at \(\theta^{*}\) by construction (the retained subset’s loss-correlated composition does not equalize as \(\nabla F \to 0\)); the measured instantiation is \(\beta \approx 0.15\) against a vanishing signal (Remark 5), and OHEM-style top-loss selection at matched budget exhibits exactly the predicted failure on the same task (AUC \(0.45\), Table 3).*

Remark 5 (Measured instantiation and the signal-to-bias ratio). On the 0.17%-imbalance replica (standardized features), the full-batch gradient at the late-training iterate has norm \(\approx 7 \times 10^{-4}\) and vanishes over training, while the measured selection bias of the uncompensated variants remains locked at \(\left\lVert b \right\rVert \approx 0.15\) throughout — a signal-to-bias ratio of \(\approx 0.5\%\). Proposition 2 then predicts that training cannot approach \(\theta^{*}\); the observed end-to-end AUCs are 0.53 (v2), 0.56 (v1) versus 0.9998 (full batch). Per-class stratification does not* repair this (measured AUC 0.45–0.62): the bias arises within the majority class, from the loss-correlated feature mean of the retained negatives, not from cross-class allocation. Under the compensated estimator, \(b \approx \delta_m \approx 0.004\), gradient norms decay (\(0.014\) at the end of training), and AUC reaches \(0.9991 \pm 0.0014\) at the same 28.4% compute saving. At 5% imbalance the true-gradient scale is \(\approx 30\times\) larger, the ratio is benign, and all variants reach AUC \(\ge 0.9965\) — delimiting quantitatively where the uncompensated shortcut is safe.*

6 The Regularized Mode (v2): Exact Bias, Measured Benefits, Quantified Contraindications↩︎

K-ABENA retains an optional biased mode: sampling restricted to the lower half \(\mathcal{L} = \{ i \in \mathcal{M}: \ell_i \le c \}\) (\(c\) the median minor loss), with \(p_i \propto \ell_i\) on \(\mathcal{L}\) and no reweighting, and feasibility constraint \(N \le \tfrac12\). Because \(\pi_i = 0\) on the upper half \(\mathcal{U} = \mathcal{M}\setminus \mathcal{L}\), Lemma 2’s mechanism is unavailable: this mode is biased by construction, and the bias admits an exact expression.

Lemma 5 (Exact bias decomposition of the v2 mode). Let \(\bar g_M, \bar g_{\mathcal{L}}, \bar g_{\mathcal{U}}\) be the per-observation gradient means over \(\mathcal{M}, \mathcal{L}, \mathcal{U}\) (\(|\mathcal{L}| = |\mathcal{U}| = k/2\)), and \(\bar g_w = \sum_{\mathcal{L}} \ell_i g_i / \sum_{\mathcal{L}} \ell_i\) the population loss-weighted mean over \(\mathcal{L}\). Then, exactly, \[\bar g_w - \bar g_M \;=\; \underbrace{(\bar g_w - \bar g_{\mathcal{L}})}_{\text{local term}} \;-\; \underbrace{\tfrac12\,(\bar g_{\mathcal{U}} - \bar g_{\mathcal{L}})}_{\text{structural term}} .\]

Proof. \(|\mathcal{L}| = |\mathcal{U}|\) gives \(\bar g_M = \tfrac12(\bar g_{\mathcal{L}} + \bar g_{\mathcal{U}})\); substitute and regroup: \(\bar g_w - \bar g_M = (\bar g_w - \bar g_{\mathcal{L}}) + \bar g_{\mathcal{L}} - \tfrac12 \bar g_{\mathcal{L}} - \tfrac12 \bar g_{\mathcal{U}}\). ◻

Under the empirically typical monotone loss–gradient relationship (\(\bar g_{\mathcal{U}} > \bar g_{\mathcal{L}}\) componentwise in magnitude), the two terms partially cancel; this is the mechanism behind the mode’s measured properties: gradient bias 4–10% below and variance 25–55% below the historical full-domain variant, and a small accuracy bonus over the full-batch baseline on multiclass tasks (\(+0.35\) points on Digits, \(+0.45\) on Wine) that we attribute to the bias acting as an implicit regularizer — an attribution supported by its disappearance under the design-unbiased estimator (v3 is statistically indistinguishable from the baseline on the same tasks; Table 1).

6.0.0.1 Quantified contraindications.

The same bias is destructive outside a characterized validity region: (i) extreme imbalance — AUC 0.53 at 0.17% positives (Proposition 2 applies); (ii) heavy label noise — at 40% symmetric flips on Breast Cancer, accuracy collapses to \(0.386\) (baseline \(0.832\); v3 \(0.808\)): noisy samples are high-loss, hence always retained as majors, while the clean low-loss anchors that the mode discards are precisely what stabilized the decision boundary; at 20% noise the mode still shows a \(+0.4\)-point benefit, locating the phase change between \(\sim\)​20% and 40%. The mode is therefore gated: minority signal \(\gtrsim 5\%\), label noise \(\lesssim 25\%\), \(N \le \tfrac12\); outside this region the canonical estimator is mandatory.

7 Experiments↩︎

7.0.0.1 Protocol.

All runs: full-batch SGD with per-epoch recalibration of \(K\) at the 40th percentile of current losses, \(N = 0.3\), \(\alpha = 0.3\) unless stated; 75/25 stratified splits; standardized features. Seeds are published and paired across methods (seeds \(0\)\(4\) throughout; \(0\)\(9\) for Table 3): every method sees the identical split, initialization, and data ordering per seed, so all comparisons are paired. We report 95% confidence intervals and two-sided paired sign-flip permutation tests (20,000 permutations) for each method against the baseline. Where means coincide to all reported digits, this is a real consequence of the paired design, not rounding: on small test sets accuracy is quantized (Breast Cancer: \(143\) test points, steps of \(0.7\) points), and the design-unbiased estimator converges to the same optimum as the baseline, so per-seed differences are exactly zero on most seeds — the permutation test then returns \(p = 1.0\) by construction, which is the statistically honest way to state parity, not evidence of manipulation. Real datasets are the scikit-learn bundled sets (Breast Cancer \(n{=}569\), Digits \(n{=}1797\)/10 classes, Wine \(n{=}178\)/3 classes, Diabetes \(n{=}442\)); the extreme-imbalance task is a synthetic replica of a card-fraud regime (0.17% positives, \(n = 20{,}000\), \(d = 15\), Gaussian classes with mean shift 1.2). Implementation: kabena v2.x (NumPy); code and seeds released [11]. Compute savings are counted as the realized fraction of skipped backward passes; they match Proposition 1 to within 0.7 points everywhere.

7.1 Parity at reduced cost on real datasets↩︎

Table 1: Real-data benchmarks. The canonical estimator matches the full-batch baseline within seed noise on every task while skipping 28–54% of backward passes — the behavior Lemma [lem:unbiased] predicts. The v2 bonus on multiclass tasks (\(+0.35\)/\(+0.45\)) is exclusive to the biased mode (Section [sec:sec:v2mode]). Gradient boosting is reported as a negative-scope control: per-round tree fitting does not average per-sample gradients across steps, and no benefit is expected or observed.
Task Metric Baseline v2 (regularized) v3 (canonical) Saving
LogReg / Breast Cancer acc. 0.9706 0.9706 \(\pm\)0.009 0.9706 \(\pm\)0.009 28.5%
Linear SVM / Breast Cancer acc. 0.9720 0.9720 \(\pm\)0.008 0.9720 54.0%
Softmax / Digits (10 cl.) acc. 0.9609 0.9644 \(\pm\)0.003 0.9609 28.4%
Softmax / Wine (3 cl.) acc. 0.9822 0.9867 \(\pm\)0.011 0.9822 28.1%
LinReg / Diabetes \(R^2\) 0.4259 0.4249 \(\pm\)0.039 28.1%
MLP-32 / Digits acc. 0.9548 0.9593 \(\pm\)0.005 0.9609 28.0%
Grad.Boosting / Breast Cancer acc. 0.9604 0.9557 (neutral; out of scope)

7.2 Extreme imbalance: failure and resolution↩︎

Table 2: The quantified instantiation of Proposition [prop:failure] and its resolution. Learning-rate sweeps (0.05–0.5) do not rescue the uncompensated variants. At 5% imbalance all methods reach AUC \(\ge 0.9965\): the failure is specific to the small signal-to-bias regime (Remark [rem:s2b]).
Method AUC (0.17%) Final \(\norm{\hat g}\) Saving Verdict
Full-batch SGD 0.9998 \(\pm\)0.0002 \(7\times 10^{-4}\) (decaying) 0% reference
v1 (historical, uncompensated) 0.56 \(\pm\)0.15 0.15 (locked) 28.4% fails (Prop. [prop:failure])
v2 (regularized, uncompensated) 0.53 \(\pm\)0.11 0.158 (locked) 28.4% fails (Prop. [prop:failure])
Per-class stratified v1 / v2 0.62 / 0.45 0.14–0.16 28.4% fails (bias is intra-class)
v3 (canonical) 0.9991 \(\pm\)0.0014 0.014 (decaying) 28.4% resolves

7.3 Matched-budget comparison with selective and reweighting rivals↩︎

The reviewer-facing question “why not compare against the established methods?” deserves numbers, with an honest caveat: the comparisons below are our CPU re-implementations at matched compute budget (identical retained-sample count per epoch where the method saves compute), not the official GPU codebases — this extends Limitation L4. Focal Loss uses \(\gamma = 2\) on the full batch (its design saves nothing); OHEM-style selection keeps the hardest examples up to the same budget as v3 with a plain average; global IS is the classical loss-proportional proposal over all \(n\) samples with self-normalized inverse-probability correction at the same budget — i.e., v3 without the two-stratum structure and without the defensive floor.

Table 3: Matched-budget rivals (\(p\): paired sign-flip permutation vs.baseline, 20,000 permutations). Three readings. (1) Only v3 combines saving with baseline-level performance in both regimes. (2) OHEM saves the same compute and is statistically indistinguishable in the benign regime, then collapses at extreme imbalance (AUC \(0.45\)) — the empirical instantiation of Proposition [prop:failure] on an external method, not only on our own earlier variants. (3) Global loss-proportional IS with correction is fine at extreme imbalance (compensation works) but significantly degrades the standard regime (\(p = 0.002\)): without the certainty stratum on majors and the defensive floor on weights, occasionally-drawn tiny-loss samples receive enormous inverse-probability weights — the design differences of Section [sec:sec:related](c) are load-bearing, not stylistic. Focal Loss performs well everywhere but saves nothing, as its design intends.
Breast Cancer (acc., 10 seeds) Fraud 0.17% (AUC, 5 seeds)
Method mean [95% CI] saving \(p\) mean [95% CI] saving \(p\)
Full-batch SGD 0.9720 [0.964, 0.980] 0% 0.9998 [0.9995, 1.000] 0%
K-ABENA v3 0.9720 [0.964, 0.980] 28.9% 1.000 0.9991 [0.998, 1.000] 28.4% 0.504
Focal Loss (\(\gamma{=}2\)) 0.9748 [0.967, 0.983] 0% 0.468 0.9998 [0.9995, 1.000] 0% 1.000
OHEM-style (top-loss) 0.9748 [0.969, 0.981] 28.9% 0.246 0.4464 [0.255, 0.638] 28.4% 0.063
Global IS + self-norm.HT 0.9573 [0.947, 0.968] 28.9% 0.002 0.9997 [0.9994, 1.000] 28.4% 0.683

7.4 Sensitivity analysis↩︎

We sweep the two user-facing controls jointly, \(K\)-percentile \(\in \{20, 40, 60, 70\}\) \(\times\) \(N \in \{0.1, 0.3, 0.5\}\) (5 seeds, Breast Cancer): test accuracy is flat across the entire grid — range \(0.9664\)\(0.9678\), a spread of \(0.14\) points, below the seed-level CI width — while the realized saving tracks Proposition 1 from \(10.8\%\) (\(K{=}20, N{=}0.5\)) to \(63.4\%\) (\(K{=}70, N{=}0.1\)). On the fraud replica the same grid (percentiles \(40, 70\)) keeps AUC \(\ge 0.9961\) everywhere. Together with the \(\alpha\)-sweep (Table 4 caption: AUC \(0.9987\)\(0.9996\) over \(\alpha \in [0,1]\), Digits unchanged) and the learning-rate sweep at extreme imbalance (\(0.05\)\(0.5\), Table 2 caption), the picture is consistent: performance is insensitive to the controls within broad ranges; the controls govern cost, not accuracy — which is the intended division of labor. Batch size (our protocol is full-batch), network depth, and optimizer schedules beyond SGD are GPU-scale questions folded into Limitation L4.

7.5 Gradient moments, noise robustness, and optimizer interaction↩︎

Table 4: Moments and stress tests. Row 1–2 quantify Remark [rem:variance-price] (bias \(\div 38\) at variance \(\times 10\)). Row 3 shows the dissolution under v3 of the small-\((n{-}k)\) fragility of the biased bound (majors carry weight 1 with no approximation). Rows 4–5 locate the label-noise phase change of the v2 mode. Rows 6–7: under AMSGrad at extreme imbalance, even the compensated estimator degrades (0.806) — the variance interacts with the \(\hat v_t\) max-lock — so we prescribe plain SGD in that regime; at standard imbalance the interaction is neutral. Sensitivity to \(\alpha\): fraud AUC ranges 0.9987–0.9996 over \(\alpha \in \{0, 0.15, 0.3, 0.5, 1\}\); Digits accuracy is unchanged to four digits.
Measurement v1 v2 v3
Gradient bias (fraud replica, pre-trained iterate) 0.128 0.151 0.004
Gradient variance (same configuration) \(\approx 0\) \(1.5\times 10^{-5}\) \(1.5\times 10^{-4}\)
Bias at adversarially small \(n{-}k\) (\(\approx 3\)) 0.17–0.23 0.035–0.037
Accuracy, 20% label noise (baseline 0.9427) 0.9469 0.9413
Accuracy, 40% label noise (baseline 0.8322) 0.386 (collapse) 0.8084 (safe)
AMSGrad, fraud: AUC / max-lock frequency 0.715 / 69% 0.806 / 60%
AMSGrad, standard 5%: AUC 0.9978 0.9979

8 Limitations↩︎

Each limitation is quantified or proved; none is asserted informally. (L1) Theorem 1 covers SGD, and the reason it does not extend to Adam, AdamW, or Lion is structural, not editorial: the proof’s descent step requires the update to be linear in the stochastic gradient, so that design-unbiasedness of \(\hat{g}_t\) transfers to the update (\(\mathbb{E}[\theta_{t+1} - \theta_t \mid \theta_t] = -\eta \nabla F + O(\delta)\)). Heavy-ball momentum preserves this linearity and is a plausible extension; Adam/AdamW precondition by a nonlinear function of the gradient history (and AdamW’s decoupled weight decay does not restore linearity), while Lion applies \(\mathrm{sign}(\cdot)\), for which \(\mathbb{E}[\mathrm{sign}(\hat{g})] \neq \mathrm{sign}(\mathbb{E}[\hat{g}])\) — an unbiased gradient buys nothing through a sign nonlinearity, and the variance inflation of L6 makes sign flips more frequent, not less. Measured interaction: at extreme imbalance, AMSGrad + v3 reaches AUC 0.806 versus 0.9991 for SGD + v3 (Table 4); at 5% imbalance the interaction is neutral. Prescription: SGD whenever Remark 5’s signal-to-bias diagnosis is adverse. (L2) BatchNorm statistics are computed over \(n\) while gradients use \((n-k)+m\) weighted samples; correct weighted-BN is not implemented; LayerNorm/GroupNorm recommended. (L3) \(K\) (percentile) and \(N\) are recalibrated by a per-epoch heuristic; no adaptive-optimality claim is made. (L4) All results are CPU-scale (shallow models, NumPy); no GPU CNN/Transformer benchmark is reported, and the fraud task is a synthetic replica rather than the original dataset. The rival comparisons of Table 3 are our matched-budget CPU re-implementations, not the official GPU codebases of the respective papers. We report no simulated deep-learning numbers. (L5) The self-normalized estimator carries bias \(\delta_m = O(1/m)\) (Lemma 3); measured 0.004 at \(n{=}20{,}000\) and 0.037 in an adversarial micro-regime; exact-\(\pi\) sampling (\(\alpha = 1\)) removes the approximation of Assumption 1(b). (L6) Unbiasedness costs variance: \(\times 10\) at matched configuration (Lemma 4, Remark 2), bounded by \(\alpha\), with a measured side effect on the AMSGrad lock frequency (57–60%). (L7) v3 matches the baseline; it does not beat it in accuracy. The small multiclass bonus belongs to the biased v2 mode, which carries the gated contraindications of Section 6 (collapse at 40% noise; failure at 0.17% imbalance). (L8) The estimator is stochastic; reproducibility requires a seeded generator (provided in the released implementation).

9 Reproducibility↩︎

All results in this paper are reproducible from the released library kabena-ml [11], available at https://github.com/Bonbhel/kabena-ml under the MIT license. The repository provides three layers: (i) Implementation — the estimator of Definition 1 via kabena_filter(losses, K, N, strategy, rng): strategy="v3" is the canonical compensated design (defensive mixture \(\alpha\), inverse-probability weights, seeded generator for exact reproducibility, cf.Limitation L8), strategy="v2" the gated regularized mode of Section 6 (its feasibility constraint \(N \le \tfrac12\) is enforced with an explicit error), and strategy="v1" the historical variant of Appendix 13 for ablation replication. (ii) Validation scripts — the validation/ directory contains the runnable scripts behind each table: gradient-moment measurements (Table 4, rows 1–3), the real-dataset benchmarks of Table 1 on the bundled scikit-learn sets, the extreme-imbalance and stratification experiments of Table 2 including the synthetic fraud-replica generator (0.17% positives, \(n = 20{,}000\), seeded), the label-noise stress test, the AMSGrad interaction runs, and the \(\alpha\)-sensitivity sweep; each script prints the seed list and matches the reported means and standard deviations. (iii) Tutorials — step-by-step notebooks for testing the estimator on a new dataset, reproducing the paper’s tables end-to-end, and experimenting beyond them (varying \(K\)-percentile, \(N\), \(\alpha\), and the mode gate), including the two-line integration pattern for scikit-learn-style training loops. Experiments require only NumPy and scikit-learn (no GPU), consistent with the scope stated in Limitation L4; total runtime for the full table suite is under one CPU-hour.

10 Conclusion↩︎

K-ABENA v3 shows that threshold-based selective backpropagation and unbiased gradient estimation are compatible: a defensive-mixture design with inverse-probability weighting preserves the exact compute saving \((1-N)k/n\) while restoring the estimator guarantees that make SGD theory applicable (Lemmas 24, Theorem 1). The negative result (Proposition 2) delimits precisely why the uncompensated shortcut — shared by OHEM, SBP, and our own earlier variants — cannot converge when selection bias dominates the signal, and the experiments quantify both the failure (AUC 0.53 at 0.17% imbalance; accuracy 0.386 at 40% noise for the biased mode) and its resolution (0.9991; 0.808) at unchanged cost. We release the implementation with both modes and the gating logic [11]; the companion volume [12] develops the pedagogical treatment and is available upon request from the author.

11 Proof of Lemma 3 (self-normalization bias)↩︎

Write \(\hat{T} = \sum_{i \notin \mathcal{M}} g_i + \sum_{i \in \mathcal{S}} \pi_i^{-1} g_i\) and \(\widehat W = (n-k) + \sum_{i \in \mathcal{S}} \pi_i^{-1}\), so \(\hat{g} = \hat{T} / \widehat W\), \(\mathbb{E}[\hat{T}] = n \nabla F\), \(\mathbb{E}[\widehat W] = n\) (both by the indicator computation of Lemma 2). Using the identity \[\frac{\hat{T}}{\widehat W} - \frac{\mathbb{E}[\hat{T}]}{n} = \frac{\hat{T} - \mathbb{E}[\hat{T}]}{n} \;+\; \hat{T}\Bigl(\frac{1}{\widehat W} - \frac{1}{n}\Bigr) = \frac{\hat{T} - \mathbb{E}[\hat{T}]}{n} \;-\; \frac{\hat{g}\,(\widehat W - n)}{n},\] take expectations; the first term vanishes, giving the exact representation \[\mathbb{E}[\hat{g}] - \nabla F \;=\; -\frac{1}{n}\,\mathbb{E}\bigl[\hat{g}\,(\widehat W - n)\bigr] \;=\; -\frac{1}{n}\,\mathrm{Cov}\bigl(\hat{g},\, \widehat W\bigr).\] First-order bound. Since \(\hat{g}\) is a convex combination of per-sample gradients, \(\left\lVert \hat{g} \right\rVert \le G_M\) almost surely; Cauchy–Schwarz on the covariance gives \[\left\lVert \mathbb{E}[\hat{g}] - \nabla F \right\rVert \;\le\; \frac{G_M}{n}\,\sqrt{\mathrm{Var}(\widehat W)} .\] Under Assumption 1(b), \(\mathrm{Var}(\widehat W) = \sum_{i \in \mathcal{M}} \frac{1-\pi_i}{\pi_i} \le \frac{k}{\alpha m/k} = \frac{k^{2}}{\alpha m}\) by Lemma 1, hence \(\left\lVert \mathbb{E}[\hat{g}] - \nabla F \right\rVert \le G_M k / (n\sqrt{\alpha m})\) — the first bound of Lemma 3, of order \(O(1/\sqrt{m})\).

Second-order rate. The Cauchy–Schwarz step is loose because \(\hat{g}\) concentrates: the standard ratio-estimator expansion [9], \[\mathbb{E}\Bigl[\frac{\hat{T}}{\widehat W}\Bigr] = \frac{\mathbb{E}[\hat{T}]}{\mathbb{E}[\widehat W]} - \frac{\mathrm{Cov}(\hat{T}, \widehat W)}{n^{2}} + \frac{\mathbb{E}[\hat{T}]\,\mathrm{Var}(\widehat W)}{n^{3}} + o(1/m),\] has both correction terms bounded by \(G_M \mathrm{Var}(\widehat W)/n^{2}\) in norm (using \(\left\lVert \mathrm{Cov}(\hat{T}, \widehat W) \right\rVert \le G_M \mathrm{Var}(\widehat W)\), since each coordinate of \(\hat{T} - \mathbb{E}\hat{T}\) is a weight-linear combination of gradient coordinates bounded by \(G_M\) times the corresponding fluctuation of \(\widehat W\), and \(\left\lVert \mathbb{E}[\hat{T}] \right\rVert/n \le G_M\)). Therefore \[\delta_m \;\le\; \frac{2\,G_M\,\mathrm{Var}(\widehat W)}{n^{2}} + o(1/m) \;\le\; \frac{2\,G_M\,k^{2}}{\alpha\, m\, n^{2}} + o(1/m) \;=\; O(1/m).\] At \(\alpha = 1\) the design is SRSWOR with \(\pi_i = m/k\) exactly, \(\mathrm{Var}(\widehat W) = \frac{k^{2}}{m}\cdot\frac{k-m}{k}\cdot\frac{k-m}{k-1} \le \frac{k(k-m)}{m}\), giving \(\delta_m \le 2 G_M k (k-m)/(m n^{2}) + o(1/m)\). Both regimes are \(O(1/m)\); the empirical values (0.004 at large \(n\); 0.037 in the adversarial micro-regime) sit inside these bounds. 0◻

12 Proof of Theorem 1↩︎

Write \(b_t = \mathbb{E}[\hat{g}_t \mid \theta_t] - \nabla F(\theta_t)\) (\(\left\lVert b_t \right\rVert \le \delta\)) and \(\xi_t = \hat{g}_t - \mathbb{E}[\hat{g}_t \mid \theta_t]\) (\(\mathbb{E}[\xi_t \mid \theta_t] = 0\), \(\mathbb{E}\left\lVert \xi_t \right\rVert^2 \le \hat{\sigma}^2\)). \(L\)-smoothness gives \[F(\theta_{t+1}) \le F(\theta_t) - \eta \langle \nabla F(\theta_t), \hat{g}_t \rangle + \frac{L\eta^{2}}{2}\left\lVert \hat{g}_t \right\rVert^{2}.\] Take conditional expectations; using \(\mathbb{E}[\hat{g}_t \mid \theta_t] = \nabla F(\theta_t) + b_t\) and \(\mathbb{E}\left\lVert \hat{g}_t \right\rVert^{2} \le \left\lVert \nabla F(\theta_t) + b_t \right\rVert^{2} + \hat{\sigma}^{2} \le 2\left\lVert \nabla F(\theta_t) \right\rVert^{2} + 2\delta^{2} + \hat{\sigma}^{2}\): \[\mathbb{E}[F(\theta_{t+1})] \le \mathbb{E}[F(\theta_t)] - \eta\, \mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2} - \eta\,\mathbb{E}\langle \nabla F(\theta_t), b_t\rangle + \frac{L\eta^{2}}{2}\bigl(2\mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2} + 2\delta^{2} + \hat{\sigma}^{2}\bigr).\] Bound \(-\mathbb{E}\langle \nabla F, b_t\rangle \le \delta\,\mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert \le \delta G_\infty\). With \(\eta \le \frac{1}{2L}\), \(L\eta^{2} \le \eta/2\), so the \(\mathbb{E}\left\lVert \nabla F \right\rVert^2\) terms combine into \(-\frac{\eta}{2}\,\mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2}\) at worst. Rearranging and summing over \(t = 0, \dots, T-1\), then dividing by \(\eta T / 2\): \[\min_t \mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2} \le \frac{1}{T}\sum_t \mathbb{E}\left\lVert \nabla F(\theta_t) \right\rVert^{2} \le \frac{2\,(F(\theta_0) - F^{*})}{\eta\, T} + L\eta\,(\hat{\sigma}^{2} + 2\delta^{2}) + 2\,\delta\, G_\infty .\] Substituting \(\eta = \min\{\tfrac{1}{2L}, \tfrac{c}{\sqrt{T}}\}\) yields the statement (absorbing constants; the \(+\delta^{2}\) of the main text majorizes \(2\delta^2 \cdot Lc/\sqrt T\) for \(T\) large). With \(\delta = 0\) (HT form) the bound is the classical rate of [10]. 0◻

13 Ablation history↩︎

The canonical design is the outcome of a systematic exploration; all rejected variants are documented to preempt re-exploration. Reference: v1, \(p_i \propto (K - \ell_i)\) over all of \(\mathcal{M}\), no reweighting. H1 (\(p_i \propto \ell_i\) over all of \(\mathcal{M}\), no reweighting): bias \(+18\) to \(+26\%\), variance \(+29\%\) vs.v1 — rejected; note that this same proposal distribution becomes near-optimal once inverse-probability weighting is added, which is precisely the v3 construction. H2 (\(p_i \propto |\ell_i - c|\), symmetric U): interpolates v1 and H1, bias \(+6\) to \(+12\%\) — rejected. H4 (upper half only, \(p_i \propto K - \ell_i\) on \(\mathcal{U}\)): bias \(+60\) to \(+80\%\) — rejected; both terms of Lemma 5 then share a sign. v2 (lower half only, \(p_i \propto \ell_i\) on \(\mathcal{L}\)): bias \(-4\) to \(-10\%\) and variance \(-25\) to \(-55\%\) vs.v1, weight ratio bounded (\(r \approx 1.4\) vs.\(r \approx 3\times 10^4\) for v1; 0/40 vs./40 violations of the sampling approximation at 1% tolerance) — adopted as an interim canonical form before the end-to-end extreme-imbalance and label-noise tests of Tables 24 revealed the shared failure of all uncompensated variants and motivated the compensated design. Equivalence sweeps: at matched bias, v2 allows no additional retention saving at \(N = 0.3\) and \(+3.3\%\) at \(N = 0.4\) relative to v1 — the value of the uncompensated refinements was always estimator quality at fixed cost, never additional cost reduction, a conclusion that transfers to v3.

References↩︎

[1]
Angela H Jiang, Daniel L-K Wong, Giulio Yang, Daniel Meng, Minjia Yu, Wei Xu, Cyrus Noghabi, Greg R Ganger, and Hai Helen Ke. Accelerating deep learning by focusing on the biggest losers. In arXiv preprint arXiv:1910.00762, 2019.
[2]
Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 761–769, 2016.
[3]
Daniel G. Horvitz and Donovan J. Thompson. A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association, 47 (260): 663–685, 1952.
[4]
Art B. Owen. Monte carlo theory, methods and examples. In Available at statweb.stanford.edu/ owen/mc, 2013.
[5]
Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollár. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 2980–2988, 2017.
[6]
Angelos Katharopoulos and François Fleuret. Not all samples are created equal: Deep learning with importance sampling. In Proceedings of the 35th International Conference on Machine Learning (ICML), 2018.
[7]
Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of the 26th Annual International Conference on Machine Learning (ICML), pages 41–48. ACM, 2009.
[8]
M. Pawan Kumar, Benjamin Packer, and Daphne Koller. Self-paced learning for latent variable models. In Advances in Neural Information Processing Systems (NeurIPS), volume 23, 2010.
[9]
Carl-Erik Särndal, Bengt Swensson, and Jan Wretman. Model Assisted Survey Sampling. Springer, 1992.
[10]
Saeed Ghadimi and Guanghui Lan. Stochastic first- and zeroth-order methods for nonconvex stochastic programming. SIAM Journal on Optimization, 23 (4): 2341–2368, 2013.
[11]
Jean-François Bonbhel. : K-abena python library (scikit-learn / pytorch / tensorflow). https://github.com/Bonbhel/kabena-ml, 2026. MIT License. Includes both estimator modes, validation scripts, and tutorials.
[12]
Jean-François Bonbhel. The Art of Ignoring: K-ABENA and the Science of Selective Backpropagation — A Complete Theory of Error Thresholding and Proportional Exclusion in Machine Learning. YekoElite University Press, Brazzaville, Republic of Congo, 2026. English edition. Available upon request from the author.