ActSub: Subspaces and Subprotocols

• ActSub is a concept comprising three frameworks: activation subspaces for out-of-distribution detection, active subspace methods for dimension reduction, and coinductive asynchronous session subtyping in program verification.
• In deep learning, ActSub applies singular value decomposition to partition neural network activations into decisive and insignificant subspaces for effective OOD detection.
• For computational science and concurrency, ActSub methods leverage eigen-decomposition and coinductive rules to enable robust dimension reduction and modular verification of concurrent protocols.

ActSub refers to three distinct but unrelated concepts in contemporary mathematics and computer science: (1) Activation Subspaces for out-of-distribution (OOD) detection in deep learning, (2) the classical Active Subspace methodology and its global extension for dimension reduction in computational science, and (3) coinductive asynchronous session subtyping mechanisms in concurrent program verification. Each of these contexts leverages "subspace" or "subprotocol" structures for principled dimensionality or behavioral reduction. The following sections provide a comprehensive overview of each ActSub construction, highlighting foundation, methodology, and principal results as reported in the relevant literature.

1. Activation Subspaces for Out-of-Distribution Detection in Deep Networks

Recent advances introduce Activation Subspaces (ActSub) as a post-hoc out-of-distribution detector that exploits the singular value decomposition (SVD) geometry of a trained classifier’s final linear head. The core idea is to decompose the network’s penultimate activation vector $a\in\mathbb{R}^n$ into components aligned with decisive and insignificant subspaces with respect to the classifier output (Zöngür et al., 29 Aug 2025).

Decisive and Insignificant Subspaces

Given a classifier with a final linear layer $W\in \mathbb{R}^{c\times n}$, its SVD is $W=U\Sigma V^\top$. Let the first $k$ right singular vectors $V_{dec}$ span the decisive subspace, and the remaining $n-k$ $V_{insig}$ span the insignificant subspace. Orthogonal projections yield $a_{dec}=P_{dec}a$, $a_{insig}=P_{insig}a$, with $P_{dec}=V_{dec}V_{dec}^\top$, $W\in \mathbb{R}^{c\times n}$0. The index $W\in \mathbb{R}^{c\times n}$1 is selected such that $W\in \mathbb{R}^{c\times n}$2 over training samples.

OOD Detection Scoring

Two complementary scores are computed:

• Insignificant-subspace score $W\in \mathbb{R}^{c\times n}$3: computes the average top-$W\in \mathbb{R}^{c\times n}$4 cosine similarity of $W\in \mathbb{R}^{c\times n}$5 with $W\in \mathbb{R}^{c\times n}$6 from training, defined as $W\in \mathbb{R}^{c\times n}$7. This is effective at detecting Far-OOD samples.
• Decisive-subspace score $W\in \mathbb{R}^{c\times n}$8: given a shaped activation $W\in \mathbb{R}^{c\times n}$9, computes the energy over the classifier logits restricted to the decisive subspace, $W=U\Sigma V^\top$0, which excels at Near-OOD.

The final OOD score is $W=U\Sigma V^\top$1 for some exponent $W=U\Sigma V^\top$2 fixed by validation.

Empirical Performance and Analysis

On standard benchmarks (ImageNet-1k, OpenOOD splits, CIFAR10/100), ActSub consistently achieves superior FPR@95%TPR and AUROC compared to previous post-hoc OOD detectors. Table 1 from (Zöngür et al., 29 Aug 2025) demonstrates the method’s improved average AUROC and reduced average FPR over competitors such as SCALE. The approach is robust to the choice of basis (SVD preferred), requires only a small cache of training activations ($W=U\Sigma V^\top$31–10%), and is nearly insensitive to the hyperparameter $W=U\Sigma V^\top$4. Limitations include a reliance on a linear classifier head and lack of demonstrated extension to pixel-wise or structured tasks.

Method	iNat AUROC/FPR	SUN AUROC/FPR	Places AUROC/FPR	Textures AUROC/FPR	Avg. AUROC/FPR
SCALE	98.17/9.50%	95.02/23.27%	92.26/34.51%	97.37/12.93%	95.71/20.05%
ActSub	98.51/7.19%	96.40/16.73%	94.40/25.64%	97.66/11.01%	96.74/15.14%

A plausible implication is that separating activation components by their task-specific and task-irrelevant properties, as implied by the classifier head’s geometry, provides a principled approach to improving OOD detection in both near- and far-distribution regimes.

2. Classical and Global Active Subspace Methods in Dimension Reduction

The term ActSub also denotes the Active Subspace methodology, a foundational technique for discovering low-rank structure in high-dimensional parameter spaces underpinning scalar-valued functions $W=U\Sigma V^\top$5. The method identifies a small subspace (“active subspace”) in which variations of $W=U\Sigma V^\top$6 are maximized (Yue et al., 2023).

Gradient-Based Active Subspace

The approach defines the covariance matrix of gradients

$W=U\Sigma V^\top$7

where $W=U\Sigma V^\top$8 is an input distribution. Eigenspectrum decomposition identifies the $W=U\Sigma V^\top$9-dimensional subspace spanned by the leading eigenvectors $k$0, with $k$1. Surrogate models $k$2 are then constructed, yielding effective dimension reduction in applications where $k$3 is tractable.

Global Active Subspace (GAS)

A generalization replaces gradients with expected finite differences, defining the global covariance matrix $k$4, where $k$5 is the coordinatewise divided difference. The eigenspace of $k$6 defines the global active subspace, capturing sensitivity over larger scales and removing dependence on differentiability or low-noise gradients.

Error Analysis and Practical Algorithm

Rigorous error bounds decompose approximation error into contributions from subspace projection, eigenvector estimation, Monte Carlo integration, and surrogate regression. Sample complexities depend on eigengaps, domain size, and function regularity. GAS demonstrates superior robustness to noise and non-differentiability in empirical studies; classical ActSub and GAS converge to similar subspaces under smoothness, but GAS maintains performance in discontinuous or highly noisy settings.

Setting	Classical AS	GAS
Differentiable, low-noise	Effective	Effective
Noisy gradients	Degrades	Robust
Nonsmooth/ridge	Fails	Accurate direction

These results substantiate GAS as a more broadly applicable tool for dimension reduction where gradient information is unreliable or intractable (Yue et al., 2023).

3. Coinductive Asynchronous Session Subtyping in Program Logics

In the context of verification of concurrent programs, ActSub refers to the integration of asynchronous session subtyping (“subprotocols”) into separation-logic-based systems such as Actris 2.0 (Hinrichsen et al., 2020).

Session-Typed Protocols and Subprotocols

Session types $k$7 define protocols for typed message-passing channels. Actris 2.0 introduces a coinductive subprotocol relation $k$8, meaning that $k$9 can safely implement $V_{dec}$0. This subtyping relation is equipped with rules supporting reflexivity, transitivity, dualization, concatenation, and crucially, asynchronous swaps (allowing !–? exchanges when order does not affect dependencies).

Logical Integration and Proof Principles

Subprotocol reasoning is embedded in the Iris framework, allowing the proof of functional correctness of programs that interleave message-passing with shared memory, locks, and higher-order state. Soundness theorems confirm that well-typed programs under such protocols are safe (do not get stuck) and satisfy their specified postconditions whenever they terminate.

Key Examples and Expressiveness

Expressive capacity of ActSub-style subprotocols is seen in, for instance, channel-based load-balancing mappers: worker protocols and client protocols are not immediate syntactic duals but are related by subprotocols under repeated input/output swaps. Mechanized Coq proofs exploit ActSub rules and automation to discharge verification obligations in realistic concurrent pipelines.

Theoretical Highlights

Central results include Theorem 4.1 of (Hinrichsen et al., 2020): adequacy of Iris with subprotocols, soundness of coinductively defined swap rules, and preservation properties for ghost state updates reflecting protocol transitions.

A plausible implication is that ActSub-style asynchronous subtyping equips program logics with the expressive power to reason modularly about channel protocols in the presence of asynchronous buffering and real-world concurrency constructs, both at the specification and mechanized proof level.

4. Comparative Tabulation and Domain Separation

Context	ActSub Role	Reference
OOD detection in deep learning	SVD-based activation split	(Zöngür et al., 29 Aug 2025)
Computational science, UQ	Eigenspace-based dimension reduction	(Yue et al., 2023)
Concurrency logic, program proof	Coinductive asynchronous session subtyping	(Hinrichsen et al., 2020)

Each ActSub framework is foundational in its domain but shares the underlying principle of exploiting subspace or subtyping structure for tractable reduction of analysis, inference, or specification.

5. Limitations and Future Directions

• The OOD detection ActSub is currently restricted to linear classifier heads and requires access to a cache of training activations. Extensions to segmentation or structured outputs are indicated as future work (Zöngür et al., 29 Aug 2025).
• Global Active Subspace (GAS) assumes the function is sufficiently regular for finite-difference structure to reveal meaningful subspaces; estimation error is governed by eigengap and sample size, with further extension possible to more general surrogate modeling settings (Yue et al., 2023).
• ActSub in program logics is mechanized and sound for HeapLang but has not been extended beyond binary session types and protocols expressible in Iris. Broadening to multiparty sessions or more general process calculi remains open (Hinrichsen et al., 2020).

In summary, ActSub encapsulates a vital methodological advance in each of its respective fields, leveraging subspace geometry, covariance structure, or coinductive protocol relations to enable robust analysis, reliable inference, or modular verification.