PaperlayerView exact revision on arXiv

Source-linked AI summary

Measure Less, Know More: Self-Supervised Test-Time Feature Acquisition

Eeshaan Jain, Linus Bleistein, Bart Deplancke, Charlotte Bunne

arXiv:2610.03454v1cs.LGcs.AI

TL;DR

Acquiring every modality at test time can be costly, while sequential selection is difficult when the downstream task is unknown. ECHO-k uses pretrained representations as self-supervised targets for a reinforcement-learning acquisition policy; across diverse benchmarks, it reports effective budgeted acquisition, supported by theoretical analysis in a linear setting.

  • Problem

    Acquiring all modalities can be costly, and selecting views is challenging when downstream tasks are unknown and task-specific acquisition signals are unavailable.

    Source: §1 · p. 1

  • Method

    ECHO-k trains a reinforcement-learning policy to select views sequentially by using a pretrained model’s full-view representation as a self-supervised proxy target.

    Source: §1 · p. 2

  • Results

    Across image, tabular, single-cell, spatial-proteomics, and geospatial benchmarks, ECHO-k rapidly concentrates informative views early and improves budgeted acquisition.

    Source: §1 · p. 3, §6 · p. 10

  • Takeaways & Limitations

    Representation recovery provides a task-agnostic acquisition signal, and the learned policy can allocate test-time acquisition budgets instance by instance.

    Source: §1 · p. 2

    The downstream-loss guarantee is stated for a fixed predictor and loss satisfying Lipschitz conditions.

    Source: §2 · p. 3

Abstract

from arXiv · show

Recent progress in multimodal, high-dimensional learning has enabled foundation models to process heterogeneous, large-scale data. However, at test time, acquiring all features or modalities can be prohibitively costly and often redundant. Sequentially selecting informative modalities is therefore critical, yet challenging when the downstream task or prediction target is unknown. To this end, we introduce ECHO-$k$, a task-agnostic and self-supervised learning principle for modality acquisition: we use a deep model's internal pretrained representations (e.g., from a foundation model) as proxy targets that summarize cross-modal information. We provide theoretical guarantees in a stylized linear setting that motivate a reinforcement learning (RL) policy for sequential modality selection. Across task-agnostic and label-free acquisition baselines, ECHO-$k$ consistently improves budgeted downstream performance across diverse foundation-model backends. Our method provides a principled route to cost-aware test-time deployment, with implications for any multimodal system where measurements are expensive or time-constrained, and downstream tasks unknown a priori.

1 Introduction

Acquiring every modality can be too costly, while task-specific acquisition depends on labels or objectives that may not be available. ECHO-k uses pretrained representations to guide label-free sequential view selection, with theoretical motivation and reported effectiveness across diverse benchmarks.

1 Introduction

  • Acquiring every view is often infeasible because each assay, sensor, or data source can impose costs in time, money, throughput, or patient burden.The value of a view depends on which views have already been observed, making budgeted acquisition sequential and adaptive.Source: §1 · p. 1
  • Existing sequential acquisition methods commonly optimize task-specific losses and may require labels, task-specific reward design, or modeling of task uncertainty.Their policies can be tied to a particular task and dataset.Source: §1 · p. 2
  • ECHO-k uses pretrained models’ latent representations as a task-agnostic signal for selecting views that approximate the full representation.The policy is trained to make progress toward the full-view representation without task-specific supervision.Source: §1 · pp. 2–3
  • The paper analyzes self-supervised feature acquisition in a linear setting, relating representation recovery to latent regret and showing gains from conditioning on partial observations over fixed orderings.The theoretical results motivate the RL reward and long-horizon acquisition strategy.Source: §1 · p. 2
  • Across image, tabular, single-cell, spatial-proteomics, and geospatial benchmarks, ECHO-k rapidly concentrates informative views in early acquisition stages.The introduction presents this as evidence of effective representation-driven acquisition under realistic budget constraints.Source: §1 · p. 3

2 Representation-Driven Acquisition as a Markov Decision Process

The section formulates acquisition as recovering a pretrained model’s full-view representation from a budget-limited sequence of partial observations. It connects that objective to downstream loss and distinguishes training signals from the information available at deployment.

2 Representation-Driven Acquisition as a Markov Decision Process

  • The full representation is computed from view-specific encodings combined by a fixed aggregator; views may be modalities or jointly acquired feature groups.The framework also covers unimodal models whose views are acquisition units processed by a shared encoder.Source: §2 · p. 3
  • The objective minimizes squared distance between the partial and full-view representations after selecting T views, with T much smaller than K.The section defines acquisition quality through representation-recovery loss.Source: §2 · p. 3
  • Training uses unlabeled data and either a full-view representation or an available partial-view representation as its recovery target; deployment uses only the current acquisition state.The full-view representation is never accessed by the learned policy at deployment.Source: §2 · p. 3
  • For a fixed downstream predictor with Lipschitz predictor and loss, representation error bounds the increase in downstream loss; labels are used only for post-hoc evaluation.This is a sufficient condition for latent recovery to bound degradation, not a claim about arbitrary predictors.Source: §2 · p. 3
  • The theoretical question is how partial-observation acquisition compares with an oracle greedy policy that can evaluate each view using full information.The complete MDP formulation is referenced in § C.Source: §2 · p. 3

3 Theoretical Guarantees

The analysis compares a full-information oracle that greedily selects views to reduce representation error with a test-time policy that must choose from partial observations. It bounds the resulting regret through a pathwise proxy analysis and, under a Gumbel latent-correlation model and a near-isometry condition, connects that proxy to true latent reconstruction loss.

3 Theoretical Guarantees

  • The optimal test-time policy greedily minimizes conditional expected representation loss using previously observed views, unlike the full-information oracle, which can use unobserved-view interactions.The oracle uses individual view energies and cross-view interaction terms; test-time policies rely on conditional expectations given observed views.Source: §3 · p. 4

3.1 Pathwise Decomposition of the Proxy Regret

  • The pathwise proxy-regret bound charges regret only when the test-time policy selects outside the oracle-selected set, with mistake magnitude controlled by conditional deviation.The proxy objective sums conditional expected view energies over unselected views, isolating partial observability from pairwise interactions in the true loss.Source: §3 · p. 4, §3.1 · p. 5
  • Expected-regret analysis is intractable without structure on view-energy orderings; a shared discrete latent factor with Gumbel noise makes the ranking distribution tractable as a Plackett–Luce mixture.The Gumbel model assumes positive views and positive posterior probability for every latent state.Source: §3.1 · p. 5

3.2 Regret Bounds under the Gumbel Model

  • A larger utility separation γ makes view-energy rankings easier to identify from partial observations.The regret bound’s oracle-mismatch term decreases as γ increases.Source: §3.2 · pp. 5–6
  • Near-isometry of the embedding directions controls the cross-view interactions ignored by the proxy, keeping proxy loss aligned with true latent reconstruction loss.The passage reports this condition as empirically satisfied by DINOv2-family vision transformers and trained autoencoders.Source: §3.2 · p. 6
  • Theorem 3.5 bounds latent regret by separating the cost of partial observability from proxy-to-true-loss error under the stated assumptions.The oracle-mismatch term decreases with utility separation; both bound terms vanish at t=0 and t=d.Source: §3.2 · p. 6

4 Methodology

The methodology uses theoretical insights about representation recovery and partial observability to motivate ECHO-k, a long-horizon reinforcement-learning policy for self-supervised view acquisition. The policy rewards progress toward a reference representation and can account for view costs.

4 Methodology

  • Figure 2 compares the expected true-loss gap for the test-time optimal, random, and prior-based policies against the oracle.Source: Figure 2 · p. 6

4.1 From Theoretical Test-Time Policies to Learnable Policies

  • The theory motivates moving beyond greedy acquisition because it analyzes only myopic policies, while practical acquisition is a finite-horizon problem with unknown conditional view distributions.Source: §4.1 · p. 6
  • ECHO-k learns a parameterized policy to maximize cumulative discounted improvement in representation quality across an acquisition trajectory, rather than mimic a one-step oracle.Source: §4.1 · pp. 6–7
  • Training uses either the full-view representation as its reference or, when only some views are available, a noisy label-free reference from those views.Source: §4.1 · p. 7

4.2 Learning Representation-Driven Policies

  • The reward measures marginal progress toward a reference representation without downstream labels; it can upweight early steps, normalize instance errors, and account for nonuniform view costs.Source: §4.2 · p. 7
  • The policy observes partial representations, masked view values, an acquisition mask, and remaining budget, and is optimized with PPO and invalid-action masking.Source: §4.2 · p. 7

5 Experimental Results

The experiments test self-supervised acquisition across frozen encoders and foundation-model backends, with downstream labels reserved for evaluation. ECHO-k performs well across diverse tasks, tolerates incomplete reference representations, and improves on label-free adaptive baselines and representation-recovery comparisons.

5 Experimental Results

  • Across experiments, frozen representation models receive partial views, and a single policy per backbone is reused across budgets and downstream tasks, including tasks unseen during policy learning.Policies use self-supervised rewards; downstream labels are reserved for post-hoc evaluation, and results are averaged over three random seeds.Source: §5 · pp. 7–8

5.1 Evaluation on Foundation Model Backends

  • ECHO-k outperforms Variance, GIR, and Random after the first ~2–3% of hPancreas acquisitions and beats baselines on both NSCLC subtyping and grading tasks.The NSCLC comparison uses identical acquisition orders for both tasks; ECHO-k achieves favorable accuracy on both.Source: §5.1 · pp. 8–9, Figure 3 · p. 8, Figure 4 · p. 9
  • On soil-pH regression, ECHO-k beats non-oracle baselines under limited modality budgets and reaches strong predictive performance after acquiring 20–30% of views.R2 is reported as a function of the number of modalities.Source: Figure 3 · p. 8, §5.1 · p. 9

5.2 Robustness to Partially Observed Reference Data

  • With randomly masked reference views, ECHO-k degrades smoothly rather than collapsing; transcriptomics remains strong at m = 20%, while proteomics degrades more gradually.At m = 10%, transcriptomics performance is nearly unchanged; at high budgets, GIR is competitive with ECHO-k.Source: §5.2 · p. 9, Table 1 · p. 10

5.3 Comparison with Adaptive Acquisition

  • ECHO-k outperforms GSMRL-AIR, Adaptive Latent Gain, and Cluster Information Gain across every tested budget on MiniBooNE and hPancreas.Adaptive Latent Gain shares ECHO-k's state and reward but optimizes only one-step gain; ECHO-k also remains competitive with supervised policies.Source: §5.3 · p. 10, Table 2 · p. 10

5.4 Comparison of representation-recovery

  • ECHO-k achieves higher normalized representation recovery than Variance across all budgets on hPancreas and MiniBooNE, and recovery ordering matches downstream accuracy ordering.The results support the paper's stated connection between representation recovery and downstream error in Theorem 3.5.Source: §5.4 · p. 10, Table 3 · p. 10

6 Conclusion

ECHO-k uses pretrained latent representations to guide label-free, budgeted sequential view acquisition. Its theory and experiments support this framework across synthetic data and multimodal application domains.

6 Conclusion

  • ECHO-k acquires views under a cost budget so partial representations rapidly approach the full-view representation.The framework is task-agnostic and label-free, using a pretrained model’s latent representation as a self-supervised proxy.Source: §6 · p. 10
  • A linear-setting analysis characterizes the gap between oracle and feasible test-time policies, supporting a cost-aware RL algorithm for long-horizon acquisition.The algorithm learns sequential acquisition strategies.Source: §6 · p. 10
  • Experiments span synthetic data and foundation-model settings in single-cell transcriptomics, spatial proteomics, and geospatial sensing.The conclusion reports improved budgeted acquisition, transfer to unseen downstream tasks, robustness to partially observed references, and extension to heterogeneous costs.Source: §6 · p. 10

A Impact and Limitations

The section outlines ECHO-k’s potential use under resource constraints and identifies deployment boundaries, then situates its task-agnostic approach in related work and formalizes sequential acquisition.

A Impact and Limitations

  • In clinical and scientific settings, ECHO-k could prioritize measurements under limited resources, but use requires a target-specific cost model and evaluation on the intended endpoint and deployment distribution.The method preserves information encoded by the backbone, does not establish that the full-view representation is optimal for every task, and does not test explicit reference-to-deployment distribution shift.Source: A Impact and Limitations · p. 18

B Related Work

  • Prior feature-acquisition work is generally task- and data-driven; ECHO-k addresses the comparatively little-studied task-agnostic setting, with an unsupervised theoretical analysis of policy gaps.The related-work discussion identifies Li & Oliva (2021) as a close, partially supervised approach.Source: B Related Work · p. 19

C Extended Mathematical Framework

  • The framework represents each view with an encoder and aggregates observed views into a partial representation, then selects unacquired views to reduce representation-recovery loss.The ideal greedy rule is not deployable because it requires the full-view representation and candidate-view forward passes; the MDP reward measures progress toward a reference representation.Source: C Extended Mathematical Framework · pp. 19–20

D Notation

  • The notation distinguishes general acquisition quantities, linear-theory variables, and latent-correlation and spectral-misalignment terms.For example, z⋆ denotes the full-view representation, while z(Pt) denotes the partial representation induced by acquired views Pt.Source: D Notation · pp. 20–22

E.1 Proof of Lemma 3.1

  • The Lemma 3.1 proof pairs equal-sized sets of oracle-only and test-time-only selections, then uses test-time optimality and availability to compare their conditional values.It constructs a permutation between the two sets and bounds the resulting discrepancy using the conditional deviation spread.Source: §E.1 · p. 23

E.2 Proof of Proposition C.1

The supplied proof passages develop bounds on policy disagreement and representation regret, while the dataset section describes benchmarks spanning multiple acquisition settings.

E.2 Proof of Proposition C.1

  • The proof uses Lemma 3.1 and Hölder’s inequality, combines Hölder’s inequality with Lemma F.3 to obtain an upper bound, while another excerpt applies Jensen’s inequality.Source: §E.2 · pp. 23–24, §E.3 · p. 24

F.1 Technical Lemma 1: Inclusion of Mismatch Events

  • Lemma F.1 places test-time-only selections within a bad-ranking event, linking policy mismatches to the ranking of feature energies.The proof argues that a selected item outside the oracle’s set cannot be among the relevant top-ranked remaining items.Source: §F.1 · p. 25

F.2 Technical Lemma 2: Moment Bound on Conditional Deviations

  • Lemma F.2 supplies a moment bound on conditional deviations across the views still available to the test-time policy.The excerpt identifies the bound’s setting and moment order but does not provide its complete expression.Source: §F.2 · p. 25

F.3 Technical Lemma 3: Bounding Disagreement Probability

  • The disagreement analysis bounds ranking failures using posterior accuracy and Plackett–Luce selection probabilities.It conditions on latent states and relates failure to whether the highest-utility remaining view is selected early.Source: §F.3 · p. 26

F.4 Technical Lemma 4: Posterior Concentration

  • Lemma F.4 bounds posterior failure using Chernoff coefficients between latent-state view distributions.The bound uses a worst-case contraction coefficient and a positive minimum prior probability across latent states.Source: §F.4 · pp. 26–27

F.5 Technical Lemma 5: Spectral Misalignment Bound

  • Lemma F.5 expresses differences between true and proxy losses through a restricted Gram-deviation matrix, yielding a spectral-norm bound.The relevant restriction is indexed by views not jointly selected by the compared sets.Source: §F.5 · p. 27

G.1 Datasets

  • The evaluation covers six datasets and seven downstream tasks, including controlled vision and physics benchmarks and biological, clinical, and geospatial settings.The supplied dataset descriptions include label-free policy training for hPancreas cell-type annotation, sequential NSCLC marker acquisition with a 10-marker budget, and soil-pH regression across geospatial modality budgets.Source: §G.1 · p. 28, §G.1.3 · p. 29, §G.1.4 · p. 29

G.1.6 Synthetic Dataset

The appendix describes synthetic data and the representation backbones used across acquisition settings. The backbones include masked encoders and frozen foundation models that support partial observations.

G.1.6 Synthetic Dataset

  • The synthetic gradient-correlation data use a shared latent state and Gumbel noise, with early features more correlated and later features increasingly dominated by independent noise.Figure 5 depicts log-feature distributions across three equiprobable latent states and feature correlations under this family.Source: §G.1.6 · pp. 29–30, Figure 5 · p. 29

G.2.1 Masked Autoencoder (MNIST, Fashion-MNIST, MiniBooNE)

  • For image and tabular data, a masked autoencoder is pretrained to reconstruct full inputs from randomly masked features, then frozen as the representation backbone.The encoder receives the mask as well as the masked input, and downstream probes are trained under random masking.Source: §G.2.1 · p. 30

G.2.2 TABULA Foundation Model

  • The frozen TABULA model embeds partially observed gene-expression profiles, with a linear probe evaluating cell-type annotation.TABULA is used without fine-tuning during policy learning or evaluation.Source: §G.2.2 · p. 30

G.2.3 VirTues Foundation Model

  • VirTues produces tissue-level representations from cell-level protein profiles, and the same ECHO-k policy is evaluated on both NSCLC tasks.Unobserved markers are treated as missing in the cell-level input.Source: §G.2.3 · p. 30

G.2.4 GeoSpatial Encoder

  • The geospatial encoder learns representations under random modality masking, explicitly distinguishing missing modalities from zero-valued features.ECHO-k interfaces with backbones through partial embeddings and the full-information target, without architecture-specific modifications.Source: §G.2.4 · p. 31

G.3 Evaluation Protocol

The protocol evaluates downstream performance across acquisition budgets while keeping acquisition policies task-agnostic. Baselines include fixed rankings, an oracle and random selection, and adaptive label-free methods.

G.3 Evaluation Protocol

  • Experiments use dataset-specific train, validation, and test splits, with three seeds for policy and downstream evaluation and five additional random-baseline runs per seed.Image and tabular tasks use a 70:15:15 split; other splits follow the cited dataset releases or prior work.Source: §G.3 · p. 31
  • Acquisition curves plot downstream MSE or accuracy against the number of acquired features, with the early 5–25% budget range highlighted as especially relevant for cost savings.Real-world evaluation trains one task-specific linear probe on partial embeddings and holds it fixed across budgets; policies remain task-agnostic.Source: §G.3 · p. 31

G.4.1 Oracle and Random Baselines; G.4.2 Model-Driven Heuristic Baselines

  • The baseline suite includes random selection and a one-step oracle, plus fixed model-driven rankings such as Prior, GIR, and ESM-based marker orders.The oracle accesses unobserved features and is not deployable; Prior and GIR use population-level importance rather than instance-specific adaptation.Source: §G.4.1 · p. 31, §G.4.2 · p. 32

G.4.3 Unsupervised Feature Selection Baselines

  • Unsupervised ranking baselines include variance, graph-Laplacian methods, MCFS, UDFS, and SVD-Entropy, adapted from feature-selection procedures into fixed test-time orders.These methods are computed on validation data and do not adapt their ordering to each instance.Source: §G.4.3 · p. 32

G.4.4 Adaptive Label-Free Baselines; G.4.5 Summary Table

  • Adaptive label-free comparisons include GSMRL-AIR, ALG, and CIG; Table 6 summarizes the implemented baselines and marks methods with full-feature ranking access or adaptive deployment.ALG predicts conditional one-step representation gain, whereas ECHO-k optimizes a long-horizon objective.Source: §G.4.4 · p. 33, Table 6 · p. 33

G.5 Implementation Details of ECHO-k

ECHO-k frames sequential feature acquisition as a finite-horizon decision problem that rewards recovering a frozen self-supervised model’s full-view representation. It learns a label-free policy and evaluates how conditioning on partial observations helps in a controlled synthetic setting.

G.5 Implementation Details of ECHO-k

  • ECHO-k seeks to minimize the discrepancy between the latent representation from acquired features and the frozen model’s full-information representation.The acquisition episode selects T features without replacement from d available features.Source: §G.5 · p. 33, §G.5.1 · p. 34

G.5.1 MDP Formulation

  • The MDP state records the acquired set, partial embedding, normalized time, embedding norm, and masked input; actions select unobserved features and transitions update the set and embedding.Previously acquired features are masked from the action space, and transitions are deterministic.Source: §G.5.1 · p. 34

G.5.2 Reward Design; G.5.4 Hyperparameters

  • The reward is a shaped, normalized reduction in representation error, using marginal gain as a proxy for conditional expected feature contribution.A step-dependent amplification encourages refinement before threshold t0, set to half the episode length.Source: §G.5.2 · p. 34, §G.5.4 · p. 34

G.5.3 Policy Learning; G.5.4 Hyperparameters

  • PPO learns a stochastic policy using only the frozen encoder and latent representations; downstream labels and probes are reserved for post-hoc evaluation.Training budgets range from 5% to 25% of available features, depending on the dataset.Source: §G.5.3 · p. 34, §G.5.4 · p. 34

H.1 Controlled Synthetic Experiment

  • The controlled synthetic experiment isolates whether conditioning acquisition on partial observations helps when the data-generating process is known.The provided setup describes a randomized isotropic linear regression task.Source: §H.1 · p. 36

H.2 Extended Baseline Comparisons

ECHO-k performs strongly across image and tabular comparisons, including label-free adaptive controls, and approaches supervised methods at low acquisition budgets. Its long-horizon planning is highlighted as a distinction from one-step or fixed-ranking strategies.

H.2 Extended Baseline Comparisons

  • At 10–20% of pixels, ECHO-k gains up to 12% over all non-oracle baselines on MNIST and Fashion-MNIST; no baseline is consistently second-best.Variance leads on MNIST, while Random is competitive on Fashion-MNIST.Source: §H.2 · p. 36

H.3 Comparison with Adaptive Acquisition Policies

  • ECHO-k outperforms all adaptive controls at every budget on MiniBooNE and hPancreas, including GSMRL-AIR under the same label-free constraints.The comparison with ALG, which shares ECHO-k’s representation-recovery signal but uses one-step gains, is presented as evidence for the value of long-horizon planning.Source: §H.3 · p. 37, Table 8 · p. 37

H.4 Comparison with Supervised Acquisition

  • At 10% acquisition on MiniBooNE, ECHO-k is within 0.2 points of the best supervised method and exceeds four of five methods, despite using no downstream labels.The supervised lead over ECHO-k is largest at 20% (3.0 points), narrowing to 0.9 and 0.5 points at 40% and 60%.Source: §H.4 · p. 37

H.5 Cost of Task Agnosticism

The section compares task-agnostic ECHO-k with supervised policies trained for NSCLC subtyping, grading, or both. ECHO-k is close to multitask performance while task-specific policies transfer less well across endpoints.

H.5 Cost of Task Agnosticism

  • At 20% acquisition, ECHO-k trails the matching task-specific policy by 0.4 points on subtyping and 2.0 points on grading.The policies use the same architecture and training budget, but supervised PPO replaces the representation reward with downstream-loss improvement.Source: §H.5 · p. 38, Table 10 · p. 38
  • Task-specific policies lose approximately 3–5 points when evaluated on the other endpoint, while ECHO-k remains within 0.4, 0.1, and 0.0 mean-accuracy points of PPO-Multitask at 20%, 40%, and 60%.ECHO-k uses neither task label during policy learning.Source: §H.5 · p. 38, Table 10 · p. 38

H.6 Robustness to Partially Observed Data

These experiments test ECHO-k under incomplete reference data, heterogeneous acquisition costs, and adaptive stopping, then examine the near-isometry assumption in pretrained and trained models. Results show gradual robustness to missing views, cost-aware gains at low budgets, and stopping without a preset budget.

H.6 Robustness to Partially Observed Data

  • On MNIST-14, partially observed-reference policies stay near or slightly above complete-reference performance; on FMNIST-14, ECHO-k degrades gradually and remains competitive with 50% missing reference views.The reference representation is computed after randomly masking a fraction of views.Source: §H.6 · p. 38, Table 11 · p. 39

H.7 Heterogeneous Acquisition Costs

  • Mean acquisition cost falls from 0.89 to 0.56 with cost-aware training, which improves low-budget accuracy; both policies converge as the budget grows.The cost-aware policy selects cheaper pixels early and reaches high accuracy with fewer cumulative cost units.Source: Figure 8 · p. 39, §H.7 · p. 39

H.8 Adaptive Stopping

  • 86.06% accuracy at a mean of 9.86 acquired features with stopping, matching 86.11% at t = 10 without requiring a preset budget.The learned stopping time trades representation recovery against total acquisition cost.Source: §H.8 · pp. 39–40, Table 12 · p. 39

H.9 Empirical Validation of the Near Isometry Assumption

  • Near Isometry requires normalized embedding-weight columns to be approximately orthogonal and is a key ingredient in the theoretical analysis.The authors validate this assumption empirically on two classes of models.Source: §H.9 · p. 40

H.9.1 DinoV2 Foundation Models

  • Across four DINOv2 scales, off-diagonal values of B^⊤B − Id concentrate near zero, with tighter concentration in larger models.The models are ViT-S/14, ViT-B/14, ViT-L/14, and ViT-G/14 with register tokens.Source: Figure 9 · p. 40, §H.9.1 · p. 40

H.9.2 Trained Autoencoders

  • Across all tested dimensions, C ∈ [0.93, 1.73]; as d increases, the standard deviation of off-diagonal elements decreases, indicating trained neural network weights form near-orthonormal systems.Source: §H.9.2 · p. 41

H.10 Analysis of Markers Acquired by ECHO-k on Spatial Proteomics

This section examines markers acquired by ECHO-k on NSCLC spatial-proteomics tasks and reports checklist disclosures on the paper’s limitations, theory, experiments, research practices, human-subject research, and LLM use.

H.10 Analysis of Markers Acquired by ECHO-k on Spatial Proteomics

  • ECHO-k consistently selects panCK first, followed by clinically relevant markers including Ki-67, CD4, and LYVE-1 for NSCLC subtyping and grading.The authors interpret this unsupervised selection as alignment with clinical domain knowledge and evidence that the acquisition objective captures diagnostically meaningful information.Source: §H.10 · p. 42

NeurIPS Paper Checklist

  • The checklist says the paper discusses limitations and provides assumptions and detailed proofs for its theoretical result.It also states that the abstract and introduction reflect the paper’s contributions and scope.Source: NeurIPS Paper Checklist · pp. 43–44
  • The authors report providing experimental details, protocols, hyperparameters, hardware information, and Python libraries, with code and publicly available data and models in the supplement.They also state that training and test details are provided in the main paper and appendix.Source: NeurIPS Paper Checklist · pp. 44–46
  • The checklist reports that experiments were run multiple times and that error bars are provided for all experiments.It also says hardware details are provided in the appendix.Source: NeurIPS Paper Checklist · p. 46
  • The checklist states that the research conforms to the NeurIPS Code of Ethics, discusses societal impact, credits asset creators, and introduces no new assets.It also marks crowdsourcing and human-subject research as not applicable because the paper contains no crowdsourcing experiments.Source: NeurIPS Paper Checklist · pp. 47–49
  • The paper reports no research with human subjects, so the IRB question is marked N/A.The checklist justification states that the paper does not contain research with human subjects.Source: NeurIPS Paper Checklist · p. 49
  • LLM usage is marked N/A because LLMs were used only to edit grammar and were not part of the core method.The checklist says LLMs were used only for grammar editing.Source: NeurIPS Paper Checklist · p. 49

Related papers

Loading 2610.03454v1…