Dual-Criterion Neuron Selection

• Dual-criterion neuron selection is a method that combines statistical discriminability and functional magnitude filters to robustly identify high-impact neurons.
• It balances sensitivity and specificity by intersecting complementary criteria, reducing redundancy and controlling performance degradation.
• The approach is used in diverse domains, from AI model editing and interpretability in vision to biological receptor selection, illustrating its broad applicability.

Dual-criterion neuron selection encompasses a class of methods in computational neuroscience, artificial intelligence, and systems biology that identify neuron (or unit) subsets in deep or biological networks by requiring simultaneous satisfaction of two complementary selection criteria. This approach enables high specificity or interpretability while controlling undesirable side effects such as performance degradation, redundancy, or loss of diversity. Recent advances operationalize dual-criterion neuron selection through precise, quantifiable metrics and are implemented in leading model editing, interpretability, and biological modeling frameworks.

1. Conceptual Foundations and Scope

Dual-criterion neuron selection refers to strategies that constrain neuron selection to the intersection of distinct filters, each designed to ensure a different aspect of selection quality. The two most common paradigms are:

• Statistical discriminability: The neuron’s activations distinguish contrasting conditions (e.g., high vs. low trait, effective vs. redundant, positive vs. negative relevance).
• Functional magnitude, reuse, or error minimization: The neuron’s effect is material in driving measurable changes (large mean activation change, substantial relevance score impact, high downstream reuse).

This methodology has found empirical support as an optimal trade-off between sensitivity and specificity, both in machine learning (e.g., LLM editing, explainable AI, unsupervised trace scoring) and multicellular biological systems (e.g., olfactory receptor selection to enforce both monoallelicity and population diversity).

2. Mathematical Formulations and Algorithms

The formalism of dual-criterion neuron selection is context-dependent but typically instantiates the following structure.

A. DPN-LE (Dual Personality Neuron Localization and Editing) Paradigm

Given paired sets of input samples grouped by contrasting trait or label:

• For each layer $l$ and neuron $j$ in a transformer’s MLP block:
  1. Compute groupwise means $\mu_{high}^{(j)}$, $\mu_{low}^{(j)}$.
  2. Compute pooled variance $\sigma_{pooled}^{(j)}$.
  3. Calculate Cohen’s $d$: $$d_{l}^{(j)} = (\mu_{high}^{(j)} - \mu_{low}^{(j)}) / \sigma_{pooled}^{(j)}$$.
  4. Calculate mean activation difference: $s_{l}^{(j)} = \mu_{high}^{(j)} - \mu_{low}^{(j)}$.

• Selection requires $|d_{l}^{(j)}| > \tau_d$ (statistical filter) and $|s_{l}^{(j)}|$ above a quantile threshold $j$0 (magnitude filter).

B. LRP-based Interpretability (Vision, XAI)

In relevance-propagation-based neuron selection:

• For each neuron, compute the mean squared error (MSE) and symmetric mean absolute percentage error (SMAPE) between its single-neuron LRP map and the global LRP heatmap.
• Assign a joint score $j$1.
• Neurons are selected by ranking $j$2 (lowest denote highest relevance and faithfulness) (Bhati et al., 2024).

C. NEX: Explore–Exploit Scoring

Detects "E-phase" (exploration) neurons as those newly recruited during chain-of-thought generation and credits only those neurons reused in the subsequent "X-phase" (exploitation):

• Novelty (spike) score and subsequent reuse are jointly required for a neuron to be positively weighted (Chen et al., 5 Feb 2026).

3. Empirical Results and Performance Trade-offs

The dual-criterion approach is empirically validated in multiple domains:

Model/Domain	First Criterion	Second Criterion	Reduction in Intervened Neurons	Capability/Accuracy Impact
DPN-LE (LLMs)	$j$3 (effect size)	$j$4 top $j$5 quantile	1,400 (0.5%) vs 21,000 (NPTI)	<8% reasoning drop (vs ≥16% prior), strong control
LRP (VGG16)	MSE on heatmap	SMAPE on heatmap	Top-5 per layer (from dense)	6.3% SMAPE, improved sharpness over dense heatmap
NEX (CoT LMs)	New neuron in E-phase	Reuse in X-phase	Only productive neurons scored	$j$6 held-out accuracy corr., ensemble boost

Dual-criterion DPN-LE isolates fewer, more precise neurons for trait editing, achieving ~96.7% reduction in intervention set size over single-metric approaches while maintaining $j$79.1/10 trait control and sharply reducing average capability drops on GSM8K, HotpotQA, and TriviaQA (Zheng et al., 30 Apr 2026). In LRP-based interpretability for computer vision, the dual error metric selection yields visualizations that better match global model decisions, with sharper and less spurious attention regions (Bhati et al., 2024). NEX shows that only neurons both novel and reused (dual criteria) contribute to higher accuracy, outperforming single-spike or reuse-only heuristics in model selection and trace curation (Chen et al., 5 Feb 2026).

4. Theoretical Justification and Optimization Principles

Dual-criterion selection achieves synergy by balancing the strengths and limitations of each individual filter:

• Statistical (e.g., effect size $j$8): Ensures neuron differences are robust to noise, but may admit many low-impact neurons.
• Magnitude/relevance (e.g., $j$9 or heatmap error): Prioritizes high-effect neurons, but these may be noisy or spurious.
• Dual application: Their intersection yields high-precision, minimal-cardinality intervention sets, improves stability, and avoids performance collapse—e.g., excluding neurons that are “statistically significant but negligible” and “large-magnitude but unreliable”.

The same logic underpins dual-objective optimization in biological neuron selection, as exemplified by olfactory receptor expression. Here, regulatory mechanisms enforce both monoallelic expression (single-allele per sensory neuron, maximizing specificity) and ensemble diversity (population‐level entropy, maximizing coverage), using layered physical and kinetic constraints, and are best described as navigating a Pareto front in the space of two biological objectives (Tian et al., 2015).

5. Algorithmic Implementations and Hyperparameterization

Dual-criterion routines are characterized by three stages:

1. Scoring: Compute per-neuron scores for each of the two criteria (Cohen’s $\mu_{high}^{(j)}$0 and magnitude in DPN-LE (Zheng et al., 30 Apr 2026), error minimization in LRP (Bhati et al., 2024), E-phase novelty and X-phase reuse in NEX (Chen et al., 5 Feb 2026)).
2. Filtering: Independently threshold or rank on each criterion; select the intersection.
3. Application: Intervene (e.g., reweight activations, edit, propagate relevance) only on selected neurons.

Key hyperparameters are set empirically for balance:

• DPN-LE employs $\mu_{high}^{(j)}$1 (top 0.5% by $\mu_{high}^{(j)}$2) and $\mu_{high}^{(j)}$3 (LLaMA-3-8B) to achieve the best empirical trade-off in trait-control accuracy vs. reasoning task preservation (Zheng et al., 30 Apr 2026).
• LRP-based selection adjusts $\mu_{high}^{(j)}$4 (tradeoff) and $\mu_{high}^{(j)}$5 (cardinality) to minimize errors subject to interpretability constraints (Bhati et al., 2024).
• NEX requires a sticky transition probability of $\mu_{high}^{(j)}$6, segment length $\mu_{high}^{(j)}$7 tokens, and as little as $\mu_{high}^{(j)}$8–$\mu_{high}^{(j)}$9 unlabeled samples for robust neuron score estimation (Chen et al., 5 Feb 2026).

6. Generalizations, Broader Applications, and Theoretical Limits

Dual-criterion neuron selection generalizes across modalities and functions:

• Personality editing: Manipulating specific behavioral traits in LLMs with minimal global interference (Zheng et al., 30 Apr 2026).
• Factual correction: Contrasting pre- and post-retrieval representations; identifying neurons distinctly encoding new vs. old facts (suggested application).
• Toxicity and stylistic control: Isolating neurons with distinct profiles in toxic vs. safe, or passive vs. active-voice generations.
• Biological diversity enforcement: Balancing monoallelicity and diversity in olfactory receptor expression (Tian et al., 2015).
• Model interpretability: Enabling precise visualizations or explanations by filtering neurons that uniquely and reliably contribute to target outputs (Bhati et al., 2024).
• Chain-of-thought selection and ranking: Improving ensemble performance and trace reliability by scoring based on both innovative recruitment and effective reuse (Chen et al., 5 Feb 2026).

Where only a single criterion is used, redundancy, instability, overfitting, or performance collapse are frequently observed. Dual-criterion frameworks shift the selection process toward a Pareto-efficient regime, especially when system constraints or task objectives are in competition.

This suggests the dual-criterion methodology represents a general principle of high-precision intervention and interpretability in distributed neural systems, with broad implications for machine learning, neuroscience, and biological regulation.