Pith. sign in

REVIEW 3 minor 285 cited by

Understanding intermediate layers using linear classifier probes

T0 review · 0 major / 3 minor · reviewed 2026-05-11 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Linear probes show that feature separability increases monotonically along the depth of neural networks.

desk verdict Linear probes give a clean, independent way to measure rising class separability layer by layer in modern CNNs like Inception and ResNet. read the letter →

arxiv 1610.01644 v4 pith:MJDBDKNQ submitted 2016-10-05 stat.ML cs.LG

Guillaume Alain , Yoshua Bengio This is my paper · ORCID
classification stat.MLcs.LG
keywords linearprobesintermediatelayersneuralnetworksfeatureseparabilityInceptionv3ResNet-50modelinterpretabilitylayer-wiseanalysis
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The authors propose training independent linear classifiers, called probes, on the activations from each layer of a neural network to measure how well those features support the classification task. This allows tracking the evolution of useful information as it passes through the model without changing how the network was trained. When applied to Inception v3 and ResNet-50, the probes demonstrate that accuracy rises steadily with layer depth. Such monitoring can help diagnose issues in models and clarify the function of intermediate layers.

What carries the argument

The linear classifier probe: a simple linear model trained separately on a layer's activations to quantify the linear separability of features for the target classes.

What would settle it

A counterexample would be a trained neural network in which the accuracy of linear probes trained on deeper layers is lower than on shallower layers, despite the model achieving high overall performance on the task.

Watch

Extended reading notes

Core claim

Training linear probes independently on each layer's activations in popular models like Inception v3 and ResNet-50 shows that the probes' classification accuracy increases monotonically with depth. This establishes that the features become progressively more linearly separable for the downstream task.

Load-bearing premise

The accuracy achieved by a linear probe trained on a layer's activations is a reliable indicator of how informative and useful those activations are for solving the classification problem.

Editorial extensions

If this is right

  • The deeper layers of the network hold features that are more readily usable by a linear classifier for the task.
  • Breaks in the monotonic increase of probe accuracy can indicate locations where the model may have training problems or suboptimal feature extraction.
  • The method provides a layer-by-layer view that can inform decisions about network architecture and where to focus debugging efforts.
  • Similar probes could be used to study how information is transformed in other deep learning models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the monotonic increase is general, it would support the view that depth allows for successive refinement of representations.
  • The approach could be used to evaluate the quality of individual layers for purposes like model pruning or transfer learning.
  • One might investigate whether the same pattern holds when using non-linear probes or in different domains such as natural language processing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

0 major / 3 minor

Summary. The paper proposes training linear classifiers, termed 'probes,' independently on the frozen activations of each layer in a neural network to measure how linearly separable the features are for the target classification task. Applied to Inception-v3 and ResNet-50, the central experimental result is that probe accuracy increases monotonically with network depth; the method is further shown to provide diagnostic value for understanding layer roles and identifying model issues.

Significance. The probe technique supplies a simple, reproducible diagnostic that requires no changes to the original model and yields direct empirical observations about feature evolution across depth. The monotonic separability finding on two standard architectures offers a concrete, falsifiable insight into how task-relevant information accumulates in deep networks. Strengths include the independent training protocol on held-out activations and the absence of post-hoc fitting or circular definitions, making the approach broadly applicable for interpretability studies.

minor comments (3)
  1. Abstract: the sentence 'the linear separability of features increase monotonically' contains a subject-verb agreement error ('increase' should be 'increases').
  2. The description of probe training (independent linear classifiers on layer activations) would benefit from an explicit statement of the loss function and optimizer used for the probes, even if standard cross-entropy and SGD are implied.
  3. Figure captions and axis labels for the accuracy-vs-depth plots should include the number of probe training runs or error bars to convey variability in the reported monotonic trend.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive review, accurate summary of the work, and recommendation to accept. The referee correctly identifies the core contribution of the linear probe technique and the monotonic separability observation on Inception-v3 and ResNet-50.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; experimental measurements are independent

full rationale

The paper presents an empirical method of training linear probes independently on frozen layer activations from held-out data to measure linear separability. The central observation—that probe accuracy increases monotonically with depth—is a direct experimental result on Inception-v3 and ResNet-50 with no equations, fitted parameters, or self-referential definitions that reduce the reported quantities back to the inputs by construction. No self-citation load-bearing steps, uniqueness theorems, or ansatz smuggling appear in the derivation chain. The method is self-contained and externally verifiable via standard supervised training on activations.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The work is empirical and rests on standard machine-learning assumptions about feature representations rather than introducing new free parameters, axioms, or entities.

assumptions (1)
  • domain assumption Linear classifier accuracy on layer activations measures the linear separability of those features for the classification task.
    This assumption underpins the claim that probe performance indicates feature quality at each layer.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Understanding intermediate layers using linear classifier probes." pith.science (2026). https://pith.science/paper/MJDBDKNQ

@misc{pith2026161001644,
  author       = {Pith},
  title        = {Pith review of: Understanding intermediate layers using linear classifier probes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJDBDKNQ}},
  note         = {Machine review of arXiv:1610.01644}
}
read the original abstract

Neural network models have a reputation for being black boxes. We propose to monitor the features at every layer of a model and measure how suitable they are for classification. We use linear classifiers, which we refer to as "probes", trained entirely independently of the model itself. This helps us better understand the roles and dynamics of the intermediate layers. We demonstrate how this can be used to develop a better intuition about models and to diagnose potential problems. We apply this technique to the popular models Inception v3 and Resnet-50. Among other things, we observe experimentally that the linear separability of features increase monotonically along the depth of the model.

Discussion (0). Continue with ORCID to comment.

Forward citations

Showing 60 of 285 Pith papers that cite this

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. See all 285 Pith citations

  1. LLM Layers Immediately Correct Each Other

    cs.CL 2026-09 accept novelty 8.0 of 10

    Adjacent transformer layers in LLMs consistently oppose each other's contributions, a mechanism called TLCM that emerges during training and selectively corrects subspaces.

  2. Do Activation Monitors Survive Model Updates? Benchmarking, Predicting, and Repairing Activation-Monitor Staleness

    cs.LG 2026-06 unverdicted novelty 8.0 of 10

    Fine-tuning updates frequently stale activation monitors for language model safety while quantization does not, with degradation predictable and repairable via label-free realignment.

  3. The Signs Were Always There: Training-Free Concept Detection and Steering in Raw Transformer Dimensions

    cs.LG 2026-06 unverdicted novelty 8.0 of 10

    Sign patterns in the unrotated standard basis of transformer activations form independent binary feature registers that support training-free detection, prediction, and causal intervention across language, vision, and...

  4. When Does LeJEPA Learn a World Model?

    stat.ML 2026-05 unverdicted novelty 8.0 of 10

    LeJEPA achieves linear identifiability of latent variables uniquely when the latents are Gaussian in worlds with stationary additive-noise transitions.

  5. Dissecting Jet-Tagger Through Mechanistic Interpretability

    hep-ph 2026-05 accept novelty 8.0 of 10

    A Particle Transformer jet tagger contains a sparse six-head circuit whose source-relay-readout structure recovers most performance and whose residual stream preferentially encodes 2-prong energy correlators.

  6. Do Audio-Visual Large Language Models Really See and Hear?

    cs.AI 2026-04 unverdicted novelty 8.0 of 10

    AVLLMs encode audio semantics in middle layers but suppress them in final text outputs when audio conflicts with vision, due to training that largely inherits from vision-language base models.

  7. What learning algorithm is in-context learning? Investigations with linear models

    cs.LG 2022-11 accept novelty 8.0 of 10

    Transformers performing in-context learning implicitly implement gradient descent, ridge regression, and least-squares predictors for linear models, with behavior shifting based on model depth, width, and data noise.

  8. Diagnosing Shape-Prior Shortcuts in Long-Range Single-Shot Fringe Projection Profilometry

    cs.LG 2026-06 conditional novelty 7.5 of 10

    PhiCalNet cuts object MAE from 14.54 mm to 4.46 mm on a 15,600-image synthetic long-range FPP benchmark by architecturally removing the shape-prior shortcut that baseline UNets exploit.

  9. From Parameters to Answers: How LLMs Retrieve and Use Their Internal Knowledge

    cs.AI 2026-09 accept novelty 7.0 of 10

    Through layerwise interventions on three LLMs, the paper causally separates query-routing information from answer-supporting content in hidden states and shows that routing control precedes content formation and then ...

  10. FARM: Reading Failure Signals from the Internal Predictive States of a Frozen Robotic World Model

    cs.RO 2026-09 accept novelty 7.0 of 10

    Frozen VLA-JEPA world model states contain failure information decodable by a lightweight readout, enabling competitive multi-task failure detection and transfer across policies and platforms.

  11. Certified Topological Interaction in Neural Representations: Class Disentanglement Is Mostly Pairwise

    cs.LG 2026-09 accept novelty 7.0 of 10

    Across 111 trained networks, certified topological overlap measurements show class disentanglement is depth-graded and essentially pairwise: a class triple is almost always less entangled than its strongest pair.

  12. Tracing Stereotypes from Representation to Output in Multilingual LLMs

    cs.CL 2026-09 accept novelty 7.0 of 10

    In Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B, stereotype-related information is linearly decodable in middle layers but causally influential only near the output, and most sparse features are language- and category-specific.

  13. Lngram v2: Latent N-Gram Memory with Interpretable Discrete Representations

    cs.CL 2026-09 accept novelty 7.0 of 10

    Lngram v2 decouples route count, memory dimension, and backbone width to scale conditional memory independently, using grouped-query attention readout and counterfactual surrogate gradients for training, achieving con...

  14. Joint Training Is Not Enough: Conditioned Cross-Granularity Training for Multimodal Document Understanding

    cs.CL 2026-09 accept novelty 7.0 of 10

    Conditioned training, but not mixed joint training, achieves mutual reinforcement between span-level and document-level tasks in multimodal document understanding on two of three corpora, with controls separating cont...

  15. Asymmetries in Spontaneous and Instructed Deception

    cs.AI 2026-08 conditional novelty 7.0 of 10

    Spontaneous and instructed deception in Llama-3.1-70B-Instruct share a direction component (cosine ≈ 0.5) but show asymmetric transfer in probes and steering between the two settings.

  16. Representation of syntax in LLMs through the lens of linear distance and similarity-aware entropy

    cs.CL 2026-08 accept novelty 7.0 of 10

    Structural probe accuracy for each syntactic relation (UASL) is predicted by the log-distance distribution and the similarity-aware entropy of the relation's head, revealing limitations of abstract syntactic represent...

  17. When does fusing hand-crafted knowledge with learned representations pay? A cost-normalized benchmark of stacking, substitution, and interference

    cs.CV 2026-08 conditional novelty 7.0 of 10

    A cost-normalized benchmark of 3,077 training configurations finds that a fixed Gabor prior stacks with different-currency sources, substitutes with same-currency sources, interferes with pretrained features, and that...

  18. ChartProbe: A Diagnostic Study on Visual Reasoning through Perception, Grounding, and Simple Reasoning

    cs.CV 2026-08 conditional novelty 7.0 of 10

    Fine-tuning VLMs on perception, grounding, and single-step arithmetic probes, with no complex-reasoning supervision, substantially improves held-out complex chart reasoning.

  19. A Probe Direction Is a Property of Its Prompt

    cs.LG 2026-08 conditional novelty 7.0 of 10

    The contrastive prompt used to build an evaluation-awareness probe determines the reported score and even the sign of its trend with model size, so the statistic is a property of the prompt rather than of the model.

  20. Excess Separability: Nuisance-Controlled Residual-Stream Probing for Benchmark Contamination Detection

    cs.CL 2026-08 conditional novelty 7.0 of 10

    A zero-sum contrast on probe accuracy across layers, recentred on a placebo baseline from the reference set, holds nominal false positive rates in contamination detection where level-based and flat-profile statistics fail.

  21. Task- and dataset-specific information in protein language models

    cs.LG 2026-08 conditional novelty 7.0 of 10

    Across 13 protein language models and 15 downstream tasks, the last layer is rarely the best source of embeddings; intermediate layers work better, and the best layer is set by dataset structure rather than by the task alone.

  22. Scaling Inherently Interpretable Language Models

    cs.CL 2026-08 conditional novelty 7.0 of 10

    Training a language model with a built-in concept bottleneck preserves compute-optimal scaling and yields interpretability metrics that improve with scale, demonstrated on an 8B causal diffusion model.

  23. Recovering Lesion Parameters from Aphasic Picture Naming Error Profiles in Large Language Models

    cs.CL 2026-08 conditional novelty 7.0 of 10

    A multi-task inverse model recovers perturbation intensity but not exact layer from LLM naming-error profiles, and the recovered parameters reproduce target profiles in 81.4% of fresh-instance tests.

  24. SPAE: Spectrally Guided Autoencoder for Pretrained Visual Latents

    cs.CV 2026-08 conditional novelty 7.0 of 10

    SPAE compresses pretrained vision latents into a 64-channel bottleneck with channel-wise masking, improving image generation while keeping reconstruction and semantic understanding strong.

  25. Sky sphere representation in language models

    cs.LG 2026-07 conditional novelty 7.0 of 10

    Most ~100B open LLMs linearly decode celestial positions from residual-stream PCA on proximity prompts, with LOO R² up to 65–85% and median angular error 12–21°, distinct from a flat RA/Dec map.

  26. LAWFUL: Law-Aligned Witness for Faithful Use of Latents

    cs.LG 2026-07 conditional novelty 7.0 of 10

    LAWFUL defines coverage-aware physical-consistency scores and circuit tests, reporting that a MoCap-to-Radar transformer's 9-component temporal circuit carries Doppler-law consistency via attention patterns.

  27. Verbalizable Representations Form a Global Workspace in Language Models

    cs.CL 2026-07 conditional novelty 7.0 of 10

    Language models represent their current reasoning in a small, readable set of verbalizable vectors (the J-space) that functions like a global workspace.

  28. The Equilibrium Is the Initialization: Lazy Identity Collapse in Physics-Structured Deep Equilibrium Reasoning

    cs.LG 2026-07 accept novelty 7.0 of 10

    In a port-Hamiltonian DEQ with learned initialization, the equilibrium equals the start to numerical precision and contributes +0.00 pp accuracy in 18 of 19 runs; a four-test diagnostic exposes the no-op.

  29. Conditional Optimal Bridge for Riemannian Activation Steering

    cs.LG 2026-07 accept novelty 7.0 of 10

    Casting activation steering as a Schrödinger Bridge on the residual hypersphere derives the log-density-ratio objective and yields query-adaptive directions that beat fixed baselines without OOD collapse.

  30. When Does Reward Teach State? A Hidden-Automaton Instrument and a Group-Language Warning Signal

    cs.LG 2026-07 conditional novelty 7.0 of 10

    High reward in sparse RL does not imply latent-state recovery; a hidden-DFA instrument separates perception from planning gaps and flags group-language structure as a pre-training warning.

  31. Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade

    cs.AI 2026-07 conditional novelty 7.0 of 10

    Hidden-state probes can flag doomed LLM-agent episodes from the first round, and a recall-calibrated cascade of abort gates cuts generated tokens by up to 60% while preserving a chosen success-recall target.

  32. Present but Not Remembered: Auditing How Frozen VLAs Encode, Deploy, and Steer Visual History

    cs.CV 2026-07 conditional novelty 7.0 of 10

    VLAs linearly encode past frames as near-redundant copies of the present and deploy them only as architecture-conditional fallback or standing use, so steerability tracks deployment regime not encoding.

  33. Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention

    cs.CV 2026-07 unverdicted novelty 7.0 of 10

    The subspace intervention framework reveals that pre-training objectives shape how ViTs encode geometric information in compressible low-rank subspaces, with peak precision at intermediate layers.

  34. Efficient and Trainable Language Model Test-Time Scaling via Local Branch Routing

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    Local Branch Routing (LBR) is a token-level framework for test-time scaling in language models that uses local branch hidden states for routing and supports end-to-end RL, showing gains in Pass@1 and Pass@32 on math r...

  35. Learning to Place Guards by Reinforcement: A Geo-Free Neural Policy for the Vertex-Guard Art Gallery Problem

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    A reinforcement learning policy for the vertex-guard art gallery problem encodes sufficient geometric information in its encoder to allow a simple classifier to achieve high coverage feasibility out of distribution.

  36. Comparing Linear Probes with Mahalanobis Cosine Similarity

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    For balanced Gaussian class projections, OOD AUROC is a linear function of MCS to the reference probe because both are sigmoid-shaped functions of the probe SNR on test data.

  37. Code Correctness Is Linearly Decodable from LLM Hidden States Before Generation

    cs.LG 2026-06 conditional novelty 7.0 of 10

    On Qwen3-4B and LiveCodeBench, first-attempt code correctness is linearly decodable from the prompt-final hidden state after residualizing prompt length, while a raw repair-success direction collapses under residualiz...

  38. How Linear Is a Transformer Feed-Forward Block? Per-Block Linear Recoverability Is Learned, Not Architectural

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    Linear recoverability of transformer FFN blocks varies widely across depth, is learned during training, and is independent of the activation function.

  39. When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis

    cs.CL 2026-06 unverdicted novelty 7.0 of 10

    Fragility, the activation noise level causing probe accuracy collapse, reveals evolving lexical-to-compositional moral encoding, layer robustness gradients, and fine-tuning differences invisible to saturated probing accuracy.

  40. ProbeAct: Probe-Guided Training-Free Failure Recovery in Vision-Language-Action Models

    cs.RO 2026-06 unverdicted novelty 7.0 of 10

    PROBEACT is a plug-and-play intervention framework that combines hidden-state probing, kinematic failure detection, and CBF-based correction to boost success rates of pre-trained VLA models on the LIBERO-plus benchmar...

  41. TRL-Bench: Standardizing Cross-Paradigm Representation-Level Evaluation of Tabular Encoders

    cs.AI 2026-06 unverdicted novelty 7.0 of 10

    TRL-Bench is a new multi-granular benchmark that releases 50 OpenML tables, linkage tasks, and a 47k-table data lake to show that tabular encoder performance is capability-specific rather than captured by one leaderboard.

  42. Anchored, Not Graded: Vision-Language Models Fail at Slant-from-Texture Perception

    cs.CV 2026-06 unverdicted novelty 7.0 of 10

    VLMs exhibit anchoring to discrete slant angles rather than graded responses across zero-shot, in-context, and fine-tuned settings, unlike human psychophysical patterns.

  43. Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability

    cs.LG 2026-06 conditional novelty 7.0 of 10

    SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimensi...

  44. Probing Spatial Structure in Pretrained Audio Representations

    cs.SD 2026-06 unverdicted novelty 7.0 of 10

    Introduces SARL benchmark showing pretrained audio encoders encode source-level spatial factors more readily than room-level factors, with patterns shaped by input configuration and training paradigm.

  45. Toward Calibrated, Fair, and accurate Deepfake Detection

    cs.LG 2026-06 unverdicted novelty 7.0 of 10

    Face-Feature Tuning is a label-free logit remapping method that reduces FPR/TPR gaps across groups in deepfake detection while preserving overall accuracy.

  46. UWM-JEPA: Predictive World Models That Imagine in Belief Space

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    UWM-JEPA uses a density-matrix latent and unitary predictor in JEPA to preserve joint-state spectrum during blind rollouts, achieving 0.77 accuracy on a five-step hidden-velocity task versus 0.53 for an LSTM baseline.

  47. The Perception-Physics Paradox: Probing Scientific Alignment with TC-Bench

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    Introduces the Perception-Physics Paradox and TC-Bench benchmark demonstrating that vision foundation models rely on visual shortcuts that fail in intense regimes rather than achieving scientific alignment via structu...

  48. LASH: Adaptive Semantic Hybridization for Black-Box Jailbreaking of Large Language Models

    cs.CL 2026-05 unverdicted novelty 7.0 of 10

    LASH adaptively composes multiple jailbreak seed prompts via genetic search over subsets and mixture weights to reach 84.5% keyword ASR and 74.5% two-stage ASR on JailbreakBench while using only 30 queries per prompt.

  49. Markovian Circuit Tracing for Transformer State Dynamic

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    This paper presents Markovian Circuit Tracing (MCT) as a benchmark and pipeline to extract and test state-transition structures in transformer activations using synthetic HMM tasks, demonstrating that state patching i...

  50. MAPS: A Synthetic Dataset for Probing Vision Models in a Controlled 3D Scene Space

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    MAPS provides 2618 validated 3D meshes and a controllable rendering pipeline to attribute vision model recognition failures to specific scene parameters, finding camera distance and elevation as the dominant failure f...

  51. Do Vision Models Truly Forget? New Findings from Representation-Level Certification of Visual Unlearning in Vertical Federated Learning

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    Mirage auditing reveals that VFL unlearning methods passing output-level checks still retain substantial class structure in representations across multiple datasets and baselines.

  52. Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance

    nlin.AO 2026-05 unverdicted novelty 7.0 of 10

    LLM societies in Nomic show non-monotonic collective adaptation peaking at mid-scales, with smaller models rule-inert and larger ones restrictive.

  53. Uncovering the Representation Geometry of Minimal Cores in Overcomplete Reasoning Traces

    cs.AI 2026-05 unverdicted novelty 7.0 of 10

    Language models produce overcomplete reasoning traces where on average 46% of steps can be removed while preserving the answer in 86% of cases, with necessity concentrated in the top three steps.

  54. Diagnosing and Correcting Concept Omission in Multimodal Diffusion Transformers

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    Text embeddings in MM-DiTs encode a detectable omission signal for missing concepts; amplifying it via OSI reduces concept omission in text-to-image outputs on FLUX.1-Dev and SD3.5-Medium.

  55. Controlling Logical Collapse in LLMs via Algebraic Ontology Projection over F2

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    Projecting LLM hidden states onto F2 algebra with 42 pairs yields 93% zero-shot accuracy on logical relations and identifies prompt-preventable late-layer collapse.

  56. A Controlled Counterexample to Strong Proxy-Based Explanations of OOD Performance: in a Fixed Pretraining-and-Probing Setup

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    Proxy rankings of pretraining datasets by learned structure can reverse the actual OOD accuracy rankings in a synthetic sequence modeling task.

  57. Deep Minds and Shallow Probes

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    Symmetry under affine reparameterizations of hidden coordinates selects a unique hierarchy of shallow coordinate-stable probes and a probe-visible quotient for cross-model transfer.

  58. From Mechanistic to Compositional Interpretability

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    The paper introduces compositional interpretability as a category-theoretic framework that casts mechanistic explanations as commuting syntactic-semantic mappings optimized under faithfulness and complexity constraint...

  59. Privacy-Aware Video Anomaly Detection through Orthogonal Subspace Projection

    cs.CV 2026-05 unverdicted novelty 7.0 of 10

    A new orthogonal projection module for video anomaly detection suppresses facial attributes via weak face-presence signals and cosine alignment while preserving anomaly-relevant features like pose and motion.

  60. SeBA: Semi-supervised few-shot learning via Separated-at-Birth Alignment for tabular data

    cs.LG 2026-05 unverdicted novelty 7.0 of 10

    SeBA is a joint-embedding framework that separates tabular data into two complementary views and aligns one view's representations to the nearest-neighbor structure of the other, improving feature-label relationships ...

See all 285 Pith citations

Reference graph

Works this paper leans on

28 extracted references · 28 canonical work pages · cited by 285 Pith papers (see all)

  1. [1]

    Understanding intermediate layers using linear classifier probes

    Alain, G. and Bengio, Y. (2016). Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644\/

    work page Pith review arXiv 2016

  2. [2]

    Arras, L., Montavon, G., M \"u ller, K.-R., and Samek, W. (2017). Explaining recurrent neural network predictions in sentiment analysis. arXiv preprint arXiv:1706.07206\/

    work page Pith review arXiv 2017

  3. [3]

    Bach, S., Binder, A., Montavon, G., Klauschen, F., M \"u ller, K.-R., and Samek, W. (2015). On pixel-wise explanations for non-linear classifier decisions by layer-wise relevance propagation. PloS one\/ , 10 (7), e0130140

    work page 2015

  4. [4]

    Biggio, B., Corona, I., Maiorca, D., Nelson, B., S rndi \'c , N., Laskov, P., Giacinto, G., and Roli, F. (2013). Evasion attacks against machine learning at test time. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases\/ , pages 387--402. Springer

    work page 2013

  5. [5]

    Binder, A., Montavon, G., Lapuschkin, S., M \"u ller, K.-R., and Samek, W. (2016). Layer-wise relevance propagation for neural networks with local renormalization layers. In International Conference on Artificial Neural Networks\/ , pages 63--71. Springer

    work page 2016

  6. [6]

    Chollet, F. et al. (2015). Keras. https://github.com/fchollet/keras

    work page 2015

  7. [7]

    Donahue, J., Jia, Y., Vinyals, O., Hoffman, J., Zhang, N., Tzeng, E., and Darrell, T. (2014). Decaf: A deep convolutional activation feature for generic visual recognition. In International conference on machine learning\/ , pages 647--655

    work page 2014

  8. [8]

    and Brox, T

    Dosovitskiy, A. and Brox, T. (2016). Inverting visual representations with convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition\/ , pages 4829--4837

    work page 2016

Show all 28 references
  1. [9]

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. In Advances in neural information processing systems\/ , pages 2672--2680

    2014

  2. [10]

    He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition\/ , pages 770--778

    2016

  3. [11]

    Jarrett, K., Kavukcuoglu, K., Lecun, Y., et al. (2009). What is the best multi-stage architecture for object recognition? In 2009 IEEE 12th International Conference on Computer Vision\/ , pages 2146--2153. IEEE

    2009

  4. [12]

    Jastrzebski, S., Arpit, D., Ballas, N., Verma, V., Che, T., and Bengio, Y. (2017). Residual connections encourage iterative inference. arXiv preprint arXiv:1710.04773\/

    2017

  5. [13]

    Lapuschkin, S., Binder, A., Montavon, G., M \"u ller, K.-R., and Samek, W. (2016). Analyzing classifiers: Fisher vectors and deep neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition\/ , pages 2912--2920

    2016

  6. [14]

    Larsson, G., Maire, M., and Shakhnarovich, G. (2016). Fractalnet: Ultra-deep neural networks without residuals. arXiv preprint arXiv:1605.07648\/

    2016

  7. [15]

    and Vedaldi, A

    Mahendran, A. and Vedaldi, A. (2015). Understanding deep image representations by inverting them. In Proceedings of the IEEE conference on computer vision and pattern recognition\/ , pages 5188--5196

    2015

  8. [16]

    and Vedaldi, A

    Mahendran, A. and Vedaldi, A. (2016). Visualizing deep convolutional neural networks using natural pre-images. International Journal of Computer Vision\/ , 120 (3), 233--255

    2016

  9. [17]

    L., and M \"u ller, K.-R

    Montavon, G., Braun, M. L., and M \"u ller, K.-R. (2011). Kernel analysis of deep networks. Journal of Machine Learning Research\/ , 12 (Sep), 2563--2581

    2011

  10. [18]

    Raghu, M., Yosinski, J., and Sohl-Dickstein, J. (2017a). Bottom up or top down? dynamics of deep representations via canonical correlation analysis. arxiv\/

  11. [19]

    Raghu, M., Gilmer, J., Yosinski, J., and Sohl-Dickstein, J. (2017b). Svcca: Singular vector canonical correlation analysis for deep understanding and improvement. arXiv preprint arXiv:1706.05806\/

  12. [20]

    C., and Fei-Fei, L

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015). ImageNet Large Scale Visual Recognition Challenge . International Journal of Computer Vision (IJCV)\/ , 115 (3), 211--252

    2015

  13. [21]

    Singh, S., Hoiem, D., and Forsyth, D. (2016). Swapout: Learning an ensemble of deep architectures. In Advances In Neural Information Processing Systems\/ , pages 28--36

    2016

  14. [22]

    Szegedy, C., Zaremba, W., Sutskever, I., Bruna, J., Erhan, D., Goodfellow, I., and Fergus, R. (2013). Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199\/

    2013

  15. [23]

    Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D., Vanhoucke, V., and Rabinovich, A. (2015). Going deeper with convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition\/ , pages 1--9

    2015

  16. [24]

    J., and Belongie, S

    Veit, A., Wilber, M. J., and Belongie, S. (2016). Residual networks behave like ensembles of relatively shallow networks. In Advances in Neural Information Processing Systems\/ , pages 550--558

    2016

  17. [25]

    Xu, K., Ba, J., Kiros, R., Cho, K., Courville, A., Salakhudinov, R., Zemel, R., and Bengio, Y. (2015). Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning\/ , pages 2048--2057

    2015

  18. [26]

    Yosinski, J., Clune, J., Bengio, Y., and Lipson, H. (2014). How transferable are features in deep neural networks? In Advances in neural information processing systems\/ , pages 3320--3328

    2014

  19. [27]

    Zeiler, M. D. and Fergus, R. (2014). Visualizing and understanding convolutional networks. In European conference on computer vision\/ , pages 818--833. Springer

    2014

  20. [28]

    Zhang, C., Bengio, S., Hardt, M., Recht, B., and Vinyals, O. (2016). Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530\/

    2016

Pith tools

Reviewed May 11, 2026 · model on record in the stance chip above.