TRIPROBE: Probing Task Separability Beyond Classification for XAI

Scientific Impact Assessment

Core Contribution. TriProbe proposes a multi-level diagnostic framework for explainable AI that traces *where* class separability is gained or lost across a learning pipeline, rather than reporting only downstream accuracy. It decomposes a multi-class problem into binary sub-tasks and applies three probes—Foundational (raw inputs and hand-crafted features), Latent (autoencoder encoder representations), and Final (penultimate classifier layer)—each scored with the Maximum Fisher's Discriminant Ratio (F1) from Ho & Basu's data-complexity literature. The stated goal is a "stage-consistent" diagnostic tool that localizes bottlenecks and identifies the most-affected class pairs. The authors are explicit that the novelty lies in *integration* of established components (Fisher ratio, autoencoders, diagnostic probing), not in any new metric or model.

Methodological Rigor. This is the paper's weakest dimension. The evaluation rests on a single dataset (Roshambo, 10 subjects, 3 gestures) and a single modality (sEMG). There is no comparison against alternative XAI/probing tools (e.g., diagnostic classifiers, Shapley, integrated gradients—all cited but none benchmarked), so the claimed advantages over single-stage methods are asserted rather than demonstrated. The "validation" of F1 is largely circular: F1 rankings are shown to agree with binary confusion matrices and with one prior paper's qualitative finding that Paper–Scissors is the hardest pair. That agreement is reassuring but weak evidence, since F1 and a linear/MLP separator both capture roughly the same second-order class-separation structure—consistency is nearly guaranteed. No error bars, no cross-subject variance analysis, no statistical significance testing, and no ablation of design choices (e.g., why per-class autoencoders, why 392 latent dims, sensitivity to feature set). Several methodological choices (class-specific AEs trained on single classes; frozen encoders) are asserted to "avoid bias" without empirical support. The F1 metric is also known to be limited to axis-aligned single-feature separability, ignoring multivariate/nonlinear structure—a significant conceptual limitation for latent-space analysis that the paper does not address.

Potential Impact. The underlying need—diagnosing *why* pipelines fail, not just *whether*—is real and broadly felt, especially in physiological-signal domains where data collection is expensive and the paper's "early-warning at data-collection stage" use case is genuinely appealing. However, the contribution is a lightweight methodological recipe rather than a tool, dataset, or theoretical result that others will readily adopt. Because it uses only off-the-shelf components and a single narrow demonstration, its likely influence is modest: it may be cited within the sEMG/biosignal ML community as an interpretability heuristic, but it does not establish benchmarks, release code, or provide theory that would anchor a research line. The claim of architecture-agnostic, cross-domain applicability (speech, EEG/MEG) is plausible but entirely untested.

Timeliness & Relevance. XAI beyond attribution and "probing" of representations are active, current topics, so the framing is timely. Yet the specific angle—Fisher-ratio separability tracing—is a fairly classical idea repackaged for the XAI narrative, so it rides a trend more than it advances a frontier.

Strengths & Limitations.

Other observations. Reproducibility is partial: the AE architecture, feature list, and preprocessing are specified, and the dataset is public, but no code is released and training details for the MLP body and per-class AE selection ("satisfactory reconstruction quality") are underspecified. The work is low-resource (single GPU, small dataset), lowering the barrier for others to extend it. There is essentially no refutation or replication value in the strong sense—the agreement with [16] is a light confirmation of an already-known qualitative observation. The paper does not overturn any prior claim.

Overall, TriProbe is a competent, clearly-motivated proof-of-concept that packages known tools into a reasonable diagnostic workflow, but its narrow evaluation, absence of baselines, and reliance on a limited separability metric cap its likely scientific impact at incremental.