SCI: A Metacognitive Control for Signal Dynamics

TL;DR

This work reframes interpretability as a real-time, controllable state rather than a static property, proposing the Surgical Cognitive Interpreter (SCI), a closed-loop metacognitive layer that wraps a stochastic predictor. SCI uses a reliability-weighted, multi-scale signal representation P(t,s), a knowledge-guided interpreter ψ_Θ, and a Lyapunov-guided controller to minimize interpretive error ΔSP while safeguarding stability, including a bounded human-in-the-loop gain. Empirically, SCI demonstrates emergent metacognition by allocating substantially more compute to ambiguous inputs (up to ~3.6×–3.8×) and yields ΔSP as a usable safety signal with AUROC in the 0.63–0.86 range across MNIST, MIT-BIH, and bearings; it matches fixed-ensemble accuracy with lower average computation. The framework unifies signal processing, cognitive modeling, and control theory to deliver robust, auditable, and safer AI suitable for safety-critical settings, with clear paths for causal extensions and broader modality deployment.

Abstract

Modern deep learning systems are typically deployed as open-loop function approximators: they map inputs to outputs in a single pass, without regulating how much computation or explanatory effort is spent on a given case. In safety-critical settings, this is brittle: easy and ambiguous inputs receive identical processing, and uncertainty is only read off retrospectively from raw probabilities. We introduce the Surgical Cognitive Interpreter (SCI), a lightweight closed-loop metacognitive control layer that wraps an existing stochastic model and turns prediction into an iterative process. SCI monitors a scalar interpretive state SP(t), here instantiated as a normalized entropy-based confidence signal, and adaptively decides whether to stop, continue sampling, or abstain. The goal is not to improve accuracy per se, but to regulate interpretive error ΔSP and expose a safety signal that tracks when the underlying model is likely to fail. We instantiate SCI around Monte Carlo dropout classifiers in three domains: vision (MNIST digits), medical time series (MIT-BIH arrhythmia), and industrial condition monitoring (rolling-element bearings). In all cases, the controller allocates more inference steps to misclassified inputs than to correct ones (up to about 3-4x on MNIST and bearings, and 1.4x on MIT-BIH). The resulting ΔSP acts as a usable safety signal for detecting misclassifications (AUROC 0.63 on MNIST, 0.70 on MIT-BIH, 0.86 on bearings). Code and reproducibility: https://github.com/vishal-1344/sci

Figures (6)

• Figure 1: Closed-loop SCI architecture. Raw signals $X(t)$ are decomposed and composed into a reliability-weighted feature canvas $P(t,s)$; a knowledge-guided interpreter $\psi_\Theta$ emits predictions and rationales consumed by an SP evaluator producing $SP(t)$ and control error $\Delta SP$. A controller, bounded by safeguards and a human-gain budget $\lambda_h$, updates $\Theta$ and weights, closing the interpretability loop.
• Figure 2: Static XAI vs. SCI under a regime change. Colored bullets below each panel show $\kappa=[\kappa_1,\kappa_2,\kappa_3,\kappa_4]$, with the component(s) that trigger adaptation highlighted when $|\Delta SP|>\gamma$. Dashed vertical lines mark regime-change events prompting Eq. (1) updates. Markers use distinct shapes and colors so they remain distinguishable under common color-vision deficiencies and in grayscale.
• Figure 3: Decomposition and SP anatomy. Left: $\Pi$ produces semantic components used in $P(t,s)$. Right: $SP(t)$ aggregates calibrated components $\kappa_1$–$\kappa_4$ via convex weights $w \in \Delta^3$. All $\kappa$ components are encoded with distinct colors and marker shapes (and line styles) so they remain distinguishable under common color-vision deficiencies and in grayscale.
• Figure 4: SCI module pipeline used in our experiments. The base predictor and feature decomposition (M1--M3) are standard; SCI adds an interpreter, SP evaluator, controller, and buffer (M4--M7). In the present work, $SP$ is instantiated via entropy, but the structural decomposition into reliability-weighted features and safeguards remains applicable to richer instantiations.
• Figure 5: SCI as a metacognitive controller. Left: MNIST digit classification. Right: MIT--BIH arrhythmia detection. For each domain we plot the empirical distribution of inference steps under the SCI controller, decomposed into samples that are ultimately classified correctly (green) and incorrectly (red). On MNIST, errors receive roughly $3.6\times$ more steps than correct predictions (mean $2.84$ vs. $10.31$), while on MIT--BIH errors receive about $1.4\times$ more steps (mean $14.24$ vs. $19.86$). On the bearings dataset (not shown), SCI spends $5.83$ steps on correct windows and $22.14$ on the rare mistakes ($\approx 3.8\times$ more). These patterns indicate that SCI reallocates computational budget as a function of difficulty rather than applying a fixed inference graph to all inputs.