Semantic Deviation Index (SDI)

• Semantic Deviation Index (SDI) is a metric that quantifies divergence from canonical states using normalized Euclidean or regression-based error.
• In compute-first networking, SDI distinguishes impactful state changes from irrelevant variations, enhancing task success and reducing update frequency.
• In computational linguistics, SDI assesses polysemy by measuring deviations in contextual embeddings, aiding in the detection of lexical ambiguity.

The Semantic Deviation Index (SDI) is a metric designed to quantify the degree to which a given data representation, state, or distribution diverges from a canonical or decision-relevant baseline by measuring semantically-relevant variation. The metric is instantiated in multiple domains: in compute-first networking (CFN) to quantify information staleness relevant to task offloading decisions, and in computational linguistics as the “polysemy index” to measure semantic ambiguity in unlabeled corpora. In both contexts, SDI operationalizes the principle that not all changes are equally meaningful—offloading policies or linguistic senses typically hinge on specific, non-linear regions of the latent state space.

1. Mathematical Definition and Core Formulation

In compute-first networking as per “Decision-Aware Semantic State Synchronization in Compute-First Networking” (Qi et al., 3 Jan 2026), SDI at time $t$, denoted as $SDI_t$, is defined as the normalized Euclidean distance between the most recent semantic state vector $z_t = E_\theta(x_t)$ (a latent embedding of raw resource state $x_t$ via a trained encoder $E_\theta$) and its last communicated (cached) value $\hat{z}_t$ at the Access Point (AP): $$SDI_t = \frac{\|z_t - \hat z_t\|_2}{\|\hat z_t\|_2 + \epsilon}$$ where $\epsilon > 0$ avoids degeneracy. This normalization enables SDI to compare relative state changes irrespective of scale.

In computational linguistics (Sproat et al., 2019), SDI is instantiated as the minimum sum-of-squares deviation from a best single-peak isotonic–antitonic regression fit to the angularly-binned radial mass of context embeddings of a term, post multidimensional scaling (MDS) or PCA. Explicitly, for a sequence $A = (A_1,\ldots, A_B)$ of radial masses in $B$ angular bins, the SDI for a word $w$ is

$$SDI(w) := \min_{p} \left( \sum_{b=1}^{p}(A_b - f_b^{\uparrow})^2 + \sum_{b=p}^B (A_b - f_b^{\downarrow})^2 \right)$$

where $f_b^{\uparrow}$ is the isotonic (monotone-increasing) fit up to bin $p$, and $f_b^{\downarrow}$ is the antitonic (monotone-decreasing) fit after $p$.

In both cases, SDI provides a scalar measure of the semantic non-conformity or potential for meaning-altering change.

2. Rationale and Decision-Awareness

SDI is designed to capture the distinction between mere data freshness (as measured by age-of-information, AoI) and semantic relevance to task-critical decisions. In CFN, AoI reflects only elapsed time since the last update and cannot discriminate between state variations that affect the control policy from changes that are operationally irrelevant. SDI, in contrast, operates in a latent space constructed specifically to encode decision boundaries: small changes in raw metrics that leave offloading decisions invariant yield low SDI, while even minor state perturbations at decision thresholds produce pronounced SDI spikes, maximizing protocol responsiveness to meaningful events (Qi et al., 3 Jan 2026).

In lexical ambiguity detection, SDI quantifies how contextually clustered or dispersed a term’s usage is in its semantic manifold. A single broad peak corresponds to unambiguous usage (low SDI); multiple peaks indicate polysemous behavior (high SDI) (Sproat et al., 2019). This suggests that SDI’s application generalizes to quantifying semantic structure in any high-dimensional or latent domain where critical transitions reside on a non-linear manifold.

3. Protocols and Algorithms

Compute-First Networking (CFN) Protocol

The service node (SN) in each time slot $t$ executes the following protocol (Qi et al., 3 Jan 2026):

1. Measure $x_t$; compute $z_t = E_\theta(x_t)$.
2. Calculate $SDI_t$.
3. Assemble input $s_{sn,t} = [z_t; SDI_t; QoS_t]$ with uplink congestion feature $QoS_t$.
4. Compute update probability $p_{up} = \pi_{sn}(s_{sn,t})$ (output of a learned MLP).
5. If $p_{up} > 0.5$, transmit $z_t$ to AP and update cached semantic; else, retain current cache.

for each time t: x_t = measure_raw_state() z_t = E_theta(x_t) SDI_t = norm(z_t - zhat) / (norm(zhat) + eps) s_sn = concat(z_t, SDI_t, QoS_t) p_up = pi_sn(s_sn) if p_up > 0.5: send(z_t) zhat = z_t

Polysemy Index Protocol

Given a target term $w$, the protocol (Sproat et al., 2019) is:

1. Collect contextual terms within $k$-token windows.
2. Construct co-occurrence matrix for these terms, compute symmetric distances.
3. Apply MDS/PCA to obtain 2D embedding.
4. Convert embedding to polar coordinates; bin angular coordinates.
5. Sum radial masses in bins; perform isotonic–antitonic regression.
6. SDI(w): Minimum sum-of-squares error from best monotonic peak fit.

4. Policy Optimization and Thresholding

Thresholding on SDI is typically not static. In SenseCFN (Qi et al., 3 Jan 2026), the SDI is supplied as a feature to a policy network ($\pi_{sn}$), which is trained via centralized training with distributed execution (CTDE). The policy is optimized using a compound loss that penalizes both communication overhead (average $p_{up}$) and task failures. Imitation-learning labels, generated offline from AoI and queue thresholds ($\tau_{max}, \delta_{warn}$), provide supervision for when an update is necessary, allowing the SDI-utilizing agent to learn an optimal reactiveness, rather than adhering to a hand-tuned threshold.

The AP, when making offloading decisions, consumes cached semantic state, its AoI, estimated delay, and local state. Including AoI and cached semantic together allows policy networks to attenuate the influence of stale information, replacing abrupt, heuristic AoI gating with a learned, continuous modulation (Qi et al., 3 Jan 2026).

5. Experimental Results and Functional Impact

In compute-first networking applications, SDI-based policies exhibit marked advantages under variable load conditions (Qi et al., 3 Jan 2026):

• Task Success Rate: Near-100% at moderate to high loads ($\lambda_{in}\leq55$ Hz); 99.6% at saturation ($\lambda_{in}=55$ Hz), outperforming content-aware (81.3%) and AoI-based (∼74%) baselines.
• Update Frequency: For $\lambda_{in}\leq40$ Hz, average update rates are reduced by 69–96% over baselines. Even at load, updates remain adaptively controlled.
• Update Interval Distribution: SDI-driven protocols yield broad, load-adaptive inter-update intervals, contrasting with the periodicity of conventional strategies.
• Semantic Latent Dimension: Representational power plateaus at $d_{sem}\geq3$; $d_{sem}=1$ yields only 70% task success, while $d_{sem}\geq2$ achieves ∼100%.

A plausible implication is that SDI facilitates robust performance and communication overhead reduction, especially in conditions prone to decision boundary crossings or bursty workloads.

6. Applications Beyond Networking and Ambiguity Detection

The SDI principle—triggering actions only on semantically consequential changes—generalizes to multiple domains:

• Control-plane synchronization: e.g., routing table updates, distributed cache coherence, where updates are transmitted only when the projected effect on routing or cache consistency exceeds SDI-based thresholds.
• Cluster-level orchestration: Defining a collective SDI across multiple heterogeneous service nodes enables coordinated consistency in distributed edge clusters (Qi et al., 3 Jan 2026).
• Semantic analysis in language: The polysemy index application demonstrates SDI’s utility in language engineering, providing a natural, data-driven ranking of ambiguous terms for lexicon expansion or disambiguation (Sproat et al., 2019).

7. Limitations and Assumptions

The SDI methodology requires certain domain assumptions:

• Markov Approximation in CFN: Encoder window $w=1$ suffices when raw features encode history; non-Markovian workloads may degrade performance.
• Offline Labeling Bias: Use of fixed thresholds to generate training labels may suboptimally constrain the learned policy. Reinforcement-based adaptation could enhance flexibility.
• Single SN–AP Topology: Current results are for point-to-point architectures. Multi-agent CFN will require SDI to be extended to multi-agent latent manifolds.
• Domain Shift: Encoders and policies may require transfer learning or continual calibration under workload or traffic regime shifts.
• Computational Complexity in Linguistics: Distance matrix and MDS steps scale quadratically or cubically in contextual vocabulary size; sparsification or subsampling are practical necessities (Sproat et al., 2019).

Overall, SDI provides a formal, domain-adapted metric for quantifying meaningful deviation in semantic and latent spaces, guiding decision-aware updates in networking, language, and beyond.