Current evaluations of learning pipelines usually reduce to downstream accuracy, leaving the underlying reasons for task success or failure unexplored. TriProbe introduces a multi‑level probing framework that provides an explainable diagnosis of task separability. Instead of treating models as black boxes, it tracks how separability evolves across input space, learned features, and final classifier.
TriProbe decomposes a multi‑task problem into binary subtasks and applies three complementary probes:
- Foundational Probe assesses separability directly on the raw inputs;
- Latent Probe measures separability within the model’s internal feature representations;
- Final Probe evaluates whether tasks remain separable at the classifier output.
The separability metric is the Maximum Fisher's Discriminant Ratio, defined as $$MFD = \frac{(\mu_1 - \mu_2)^2}{\sigma_1^2 + \sigma_2^2}$$ where $\mu_i$ and $\sigma_i^2$ denote the mean and variance of class $i$. This principled metric quantifies conflicts or bottlenecks at each level.
Experiments on the Roshambo sEMG benchmark demonstrate that TriProbe can uncover hidden breakdowns, guiding data collection, validation, and architecture design to address the identified issues.
Review