{"id":"4e47fa38-78df-4bbb-b9cd-316aab1baf5c","arxiv_id":"2501.04816","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Probabilistic skip connections attach a distance-aware probabilistic model to an intermediate layer of a pretrained classifier, selected by neural-collapse metrics, yielding deterministic UQ and OOD detection without retraining.","lead":"This paper shows how to attach uncertainty estimates to an already-trained image classifier without retraining it: find an intermediate layer that still preserves input differences, project its features down, and fit a simple probabilistic model to them. The practical payoff is that existing neural networks can gain out-of-distribution detection and calibrated confidence scores as a retrofit.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Neural-collapse proxy for the bi-Lipschitz condition is validated only on CIFAR-10/ResNet-50 and is not shown to transfer to unseen OOD inputs; if the selected layer is not OOD-sensitive, the central retrofit claim fails.","rationale":"The paper is practical and clearly written, and the Dirty-MNIST/FashionMNIST experiments in Sec. 5.2 provide some genuine support: on two architectures the intermediate-layer density separates iD from OOD at least as well as spectral normalization. The synthetic sign-example in App. C.1 also illustrates that non-collapsed early layers can preserve perturbation sensitivity. However, these demonstrations do not exercise the load-bearing assumption at scale. The layer-selection rule is a proxy, not a derivation: NC1 and NC4 are functions of h^(j) evaluated on the training distribution and its labels, whereas the bi-Lipschitz condition in Eq. 1 that motivates deterministic UQ is a statement about all inputs, including OOD inputs. There is no argument that low within-class variance and high nearest-centroid accuracy on iD points imply that OOD points remain far from iD features in the selected layer; in fact, an over-parameterized network can be arbitrarily non-injective beyond the training manifold while still having good class geometry on it. If that happens, the QDA density can assign high likelihood to OOD points near class centroids, and the primary claimed advantage over retraining-based methods disappears. The reader's conditional verdict already flags the layer-selection rule as the weakest assumption; this stress-test sharpens that concern by identifying the missing piece as empirical evidence of OOD sensitivity of the selected layer, not merely the crispness of the 'especially close' criterion. The proposed check directly supplies such evidence. If it passes on a standard multi-OOD benchmark, the central claim is substantially supported and the paper merits acceptance after minor revisions; if it fails, the claim should be weakened to 'PSC can work for some architectures and OOD shifts,' and the selection criterion would need further validation. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":16127,"tokens_out":6513,"duration_ms":65881,"concrete_test":"Take the same setting used to motivate the rule in Sec. 5.1: a ResNet-50 trained on CIFAR-10 without spectral normalization. Apply the Sec. 4.1 rule to select the candidate layer(s). Then evaluate PSC on a standard multi-OOD benchmark with SVHN, CIFAR-100, LSUN, and Textures as OOD sets, computing the AUROC of the QDA feature density. Compare with (i) PSC at the penultimate layer (DDU baseline without SN) and (ii) DDU+SN. If PSC at the selected layer does not match or exceed DDU+SN on a majority of these OOD sets, or if the penultimate-layer baseline is not clearly worse, then the NC1/NC4 selection rule has not been shown to provide OOD sensitivity. This check reuses already-trained models and requires only forward passes plus density estimation, so it is cheap to run.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that retrofitting a pretrained classifier at a layer chosen with NC1/NC4 yields OOD detection matching DDU+SN. This requires the selected subnetwork h^(j) to satisfy the sensitivity half of Eq. 1 for inputs far from the training distribution. Section 4.1 selects the layer using Tr(Sigma_W)/Tr(Sigma_T) < epsilon and nearest-centroid accuracy, both computed on the training/validation set. These are within-distribution, class-conditioned statistics: they measure how separated class means are and how much within-class variance remains, but they do not constrain the mapping of an OOD input. A layer can have NC1 > 0.2 and high NC4 on CIFAR-10 while mapping SVHN or LSUN images arbitrarily close to in-distribution class means at that layer, destroying the distance-awareness needed by QDA. The only direct evidence for the proxy is the ResNet-50/CIFAR-10 layerwise analysis in Sec. 5.1, which contains no OOD inputs, plus the small Dirty-MNIST/FashionMNIST experiments in Sec. 5.2. Those experiments use one OOD source per model and do not test whether the selection rule generalizes across architectures, datasets, or OOD shifts. Thus the key premise, that neural-collapse health on the training set identifies a layer whose geometry is also healthy for OOD data, is unverified. If it fails, PSC's reported performance is not a property of the method but an accident of the particular layers and OOD benchmarks chosen.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces probabilistic skip connections (PSCs) as a post-hoc method for deterministic uncertainty quantification (UQ) in classification neural networks. Rather than retraining a network with spectral normalization (SN) to satisfy a bi-Lipschitz condition, the authors propose using neural-collapse metrics (NC1 and NC4) to identify an intermediate layer that is both sensitive and smooth. They then project that layer's activations to a low-dimensional space via Tucker decomposition and fit a distance-aware probabilistic model (quadratic discriminant analysis for OOD detection and a KFAC-Laplace linear model for in-distribution UQ). Experiments on CIFAR-10, Dirty-MNIST, and FashionMNIST with ResNet-18, ResNet-50, VGG-16, and Wide-ResNet-28-10 suggest that PSCs can match or exceed DDU with spectral normalization, while also applying to architectures without residual connections.","tokens_in":16442,"tokens_out":4650,"duration_ms":44492,"significance":"If the central claims hold, PSCs would be a practically valuable contribution: they promise a drop-in retrofit for existing pretrained classifiers, require no retraining, and extend deterministic UQ to non-residual architectures. The paper provides pseudo-code for the full pipeline, reports 10-seed error bars in the main in-distribution UQ table (Table 4), and includes a robustness study of the projection step. However, the layer-selection rule that the method's success depends on is under-specified and is validated only on a single architecture/dataset pair (ResNet-50/CIFAR-10) in Section 5.1. The empirical evidence for the headline OOD claim is also thinner than the text suggests, with no error bars in Table 2 and narrow baseline coverage. These gaps currently limit the strength of the conclusions.","major_comments":[{"comment":"The layer-selection rule in Section 4.1 is not precisely specified. The criterion \"taking both the layer before and after collapse if NC1 is especially close to epsilon\" admits subjective judgment, and the only validation that NC1/NC4 serve as proxies for the bi-Lipschitz condition is a single architecture/dataset pair (ResNet-50 on CIFAR-10 in Section 5.1) with no out-of-distribution inputs. Since the paper's central claim is that PSCs can be dropped into any pretrained classifier, the authors need a fully specified selection rule or a systematic study across multiple architectures, datasets, and OOD shifts to support generality.","section":"Section 4.1, Section 5.1"},{"comment":"Table 2 reports AUROC values for OOD detection and aleatoric/epistemic separation without any error bars or repeated-seed variability, in contrast to the 10-seed results in Table 4. These point estimates are the primary evidence that PSCs \"match or exceed\" DDU with spectral normalization. Without variability estimates it is impossible to judge whether the reported differences are meaningful; the authors should report means and standard errors over multiple seeds or otherwise justify the single-run comparison.","section":"Section 5.2, Table 2"},{"comment":"The evaluation that projection preserves sensitivity and smoothness is circular: the projection dimensions c_proj and d_proj are chosen to keep NC1 and NC4 nearly unchanged, and the same metrics are then reported in Table 3 as evidence of preservation. The robustness study in Figure 5 also uses NC1/NC4 as the sole outcome. To break this loop, the authors should validate the projection using external metrics, such as OOD detection AUROC or in-distribution NLL, as a function of the projection dimensions.","section":"Section 5.3, Table 3"},{"comment":"The headline claim that PSCs \"match or exceed existing single-pass methods requiring training modifications\" is only tested against a single baseline, DDU with spectral normalization (Table 2), and each model is evaluated on only one OOD source (FashionMNIST for Dirty-MNIST). No comparisons to SNGP, DUQ, or other deterministic UQ methods are reported in the experiments, and there is no evaluation on more diverse OOD benchmarks (e.g., SVHN, LSUN, CIFAR-100 vs CIFAR-10). The stated breadth of the claim is therefore not supported by the presented experimental evidence.","section":"Section 5, overall baselines"}],"minor_comments":[{"comment":"The text says both that a multinomial logistic regression model is used to \"replace the original network predictions\" and that \"in-distribution (iD) predictions use network outputs.\" This is contradictory or at least unclear; please clarify which outputs are used for iD predictions and for predictive entropy.","section":"Section 5.2"},{"comment":"The value of epsilon = 0.2 is presented as a fixed cutoff, but the phrase \"especially close to epsilon\" leaves the rule ambiguous; a precise operationalization (e.g., a tolerance on |NC1 - epsilon|) would make the method reproducible.","section":"Section 4.1"},{"comment":"The caption and text refer to \"boxes\" highlighting layers that meet the criteria, but no boxes are actually visible in the figure; please add the annotation or remove the reference.","section":"Figure 3"},{"comment":"The abbreviation \"Jacob.\" in the \"F.C. Soln.\" row for DUQ is not defined in the caption or the body; please spell out \"Jacobian regularization\" or provide a reference.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely and practically relevant problem, and the core idea of using neural-collapse metrics to select an intermediate layer for deterministic UQ is interesting. The main weakness is the empirical support for the layer-selection heuristic: it is only validated on one architecture/dataset pair, and the OOD comparison lacks error bars and sufficient baselines. I believe these issues are addressable through additional experiments and a sharper specification of the rule, so I recommend major revision rather than rejection. I also encourage the authors to consult the DDU and SNGP literature more broadly when framing their novelty claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a sensible, well-motivated retrofit for deterministic UQ. The central idea—use neural-collapse metrics to pick an intermediate layer that still separates classes, then fit a distance-aware head there—works in the experiments they actually ran. The paper deserves a serious referee, but it needs more evidence that the layer-selection rule transfers beyond CIFAR-10/ResNet-50.\n\nWhat's new: prior intermediate-feature work (DVE, same authors) was regression-only. Classifying with a neural-collapse-based layer choice, plus the Tucker projection, is a legitimate extension. The empirical observation that spectral normalization delays collapse and that non-SN networks have usable intermediate layers is cleanly made in Fig. 3. The VGG-16 result—no residual connections, no SN, PSC still separates iD/OOD—is the strongest point in the paper and supports the 'broader family of architectures' claim.\n\nThe main soft spot is the layer-selection rule in Sec. 4.1. It relies on NC1 > 0.2 and NCC accuracy, both computed on the training/validation set. That is within-distribution evidence, and the paper does not show that a layer chosen this way stays sensitive for OOD inputs. The stress-test worry is fair: a layer can have nice class-separated geometry for CIFAR-10 and still map SVHN close to a class mean. The only direct validation is one architecture/dataset pair plus small Dirty-MNIST runs. That is thin for a method whose whole pitch is 'works for any pretrained classifier.' I would want a few more architectures and a harder OOD shift before trusting the rule. The 'especially close to epsilon' criterion is vague and needs operationalization. Table 2 has no error bars; the projection dimensions are tuned against the same metrics used to evaluate; and there is no code. None of these are fatal, but they add up. The self-referential use of NC1/NC4 is real but partially broken by the external UQ metrics, so I would call it minor. The paper itself admits the projection dimensions are not standardized and parameter sweeping remains; that is an honest limitation statement, though it does soften the 'drop-in' framing.\n\nBottom line: give it to a referee. The empirical claims for the tested setups are credible, the method is clearly described, and the extension to non-residual networks is valuable. But the layer-selection heuristic needs stronger validation and a sharper specification before the broad retrofit claim can be taken at face value. I would expect major revisions, not a desk reject.","headline":"A practical retrofit for deterministic UQ that mostly delivers on its claims, but the layer-selection heuristic needs stronger evidence before the no-retraining promise generalizes.","tokens_in":17002,"tokens_out":2298,"would_cite":true,"duration_ms":22157,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Any pretrained classifier can be retrofitted with uncertainty by attaching a probabilistic head to an intermediate layer selected with neural-collapse metrics, avoiding retraining.","keywords":["deterministic uncertainty quantification","out-of-distribution detection","neural collapse","probabilistic skip connections","spectral normalization","intermediate layers","feature collapse","single-pass inference"],"falsifier":"Measure the empirical sensitivity constant $L_1$ of Eq. (1) at the selected candidate layer on an out-of-distribution benchmark; if $L_1$ is effectively zero for OOD inputs even though $\\mathrm{NC}_1$ stays above 0.2 on the training set, the neural-collapse proxy has failed and PSC uncertainty would be overconfident.","tokens_in":15878,"feed_emoji":"🧠","tokens_out":15308,"duration_ms":123355,"temperature":0.7,"pith_summary":"Deterministic uncertainty quantification (UQ) normally requires a feature extractor that is both sensitive and smooth, and existing single-pass methods enforce this by retraining with spectral normalization. This paper proposes probabilistic skip connections (PSCs): instead of retraining, use two neural-collapse statistics to find an intermediate layer of an already-trained network that already has both properties. A distance-aware probabilistic model is fit to a projected version of that layer's features, giving epistemic and aleatoric uncertainty in one forward pass. On CIFAR-10 and Dirty-MNIST experiments, PSC-equipped networks match or exceed the in-distribution calibration and out-of-distribution detection of spectral-normalized networks, and the approach also works on networks without residual connections. The practical payoff is that uncertainty estimates can be added to existing classifiers with a lightweight post-hoc fit on training data.","feed_headline":"Neural-collapse metrics add uncertainty to pretrained classifiers","feed_subtitle":"A post-hoc fit to a chosen intermediate layer matches spectral-normalization baselines on OOD detection.","key_machinery":"The probabilistic skip connection is a distance-aware probabilistic head attached to the output of a chosen intermediate layer instead of the network's penultimate layer. The layer is selected by a collapse-accuracy trade-off rule: compute $\\mathrm{NC}_1 = \\operatorname{Tr}(\\Sigma_W)/\\operatorname{Tr}(\\Sigma_T)$ and $\\mathrm{NC}_4$ (nearest-centroid accuracy) at every layer, then take the deepest layer with $\\mathrm{NC}_1 > 0.2$ while $\\mathrm{NC}_4$ remains high, optionally combining the layer just before and after collapse. $\\mathrm{NC}_1$ and $\\mathrm{NC}_4$ act as proxies for the sensitivity and smoothness bounds of Eq. (1). The intermediate feature map is reshaped, channel-wise whitened, and projected with a tensor factorisation of its channel-wise covariance to a low-dimensional vector, on which the probabilistic head is fit.","core_discovery":"The paper's central claim is that the bi-Lipschitz condition underlying deterministic UQ does not need to be enforced during training: a normally trained network contains an intermediate layer that is both sensitive and smooth, because sensitivity and smoothness trade off across depth. The neural-collapse metric $\\mathrm{NC}_1$ (the ratio of within-class to total feature variance) is used as a proxy for sensitivity, and $\\mathrm{NC}_4$ (nearest-centroid accuracy) as a proxy for smoothness; the candidate layer is the deepest one before collapse sets in that still has high $\\mathrm{NC}_4$. Placing a distance-aware head—quadratic discriminant analysis for out-of-distribution detection and a linear model with a Gaussian posterior approximation for predictions—on projected features from this layer yields uncertainty quality comparable to spectral-normalized training. Experiments on ResNet-50, ResNet-18, Wide-ResNet-28-10, and VGG-16 support the claim, including a VGG-16 network without residual connections where the penultimate-layer feature space is poorly structured.","pith_inferences":["A natural extension would be to apply the same collapse-accuracy rule to language models or other non-vision architectures, where spectral normalization is rarely used and retraining is expensive; the projection step would need to be adapted to token-level representations.","The fixed $\\mathrm{NC}_1 > 0.2$ cutoff is a plausible weak point; a data-driven variant that picks the layer maximizing validation out-of-distribution separation could make the rule more robust across architectures.","If the proxy holds broadly, PSCs effectively give a cheap drop-in UQ retrofit for deployed models, relevant for production systems that cannot afford ensembles or Bayesian inference at inference time."],"forward_implications":["Any pretrained classification network can be given epistemic and aleatoric uncertainty estimates via a post-hoc fit on training data, with no change to the network architecture or training procedure.","Deterministic UQ methods such as DDU and SNGP-style distance-aware heads become usable on networks that lack residual connections, where the penultimate-layer feature space is typically too collapsed for reliable uncertainty.","Spectral normalization is demoted from a requirement to one possible way of satisfying the sensitivity-smoothness condition, because the same effect is achievable by layer selection.","The collapse-accuracy trade-off can also serve as a measure of a network's effective depth, identifying how early the classifier has already committed to its decision boundary."],"supporting_citations":[{"why":"Defines distance-aware output layers and the spectral-normalization training that PSCs aim to replace for deterministic UQ.","marker":"[14]"},{"why":"Provides the DDU method (Gaussian discriminant analysis plus predictive entropy) and the experimental setup reused in the case studies.","marker":"[18]"},{"why":"Introduces the neural-collapse criteria NC1-NC4 that the layer-selection metrics approximate.","marker":"[20]"},{"why":"Supplies the intermediate neural-collapse metrics and the implementation used to measure NC1 and NC4 at every layer.","marker":"[21]"},{"why":"Demonstrates that intermediate-layer representations can support UQ in regression, the line of work PSCs extend to classification.","marker":"[10]"},{"why":"Introduces spectral normalization, the training modification whose necessity the paper argues against.","marker":"[17]"},{"why":"Presents DUQ, a deterministic UQ method requiring RBF networks and modified training, used as a baseline for comparison.","marker":"[25]"},{"why":"Provides a scalable approximation to the posterior covariance used for the uncertainty head.","marker":"[22]"},{"why":"Supplies the tensor-decomposition background and matricization convention needed for the projection of intermediate representations.","marker":"[11]"}],"fun_headline_variants":["Skip retraining: pick a layer, fit a model, get UQ","Neural collapse picks the right layer for uncertainty","Add uncertainty to any deep net without retraining","Find the sensitive-smooth layer and get UQ for free","Pick a layer by neural collapse, add a head, estimate uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that an intermediate layer picked by neural-collapse statistics on training and validation data will also be both sensitive and smooth for unseen out-of-distribution inputs, so that distances in that feature space remain meaningful for inputs the network has never seen.","fun_headline_variants_meta":{"raw":{"variants":["Skip retraining: pick a layer, fit a model, get UQ","Neural collapse picks the right layer for uncertainty","Add uncertainty to any deep net without retraining","Find the sensitive-smooth layer and get UQ for free","Pick a layer by neural collapse, add a head, estimate uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000349,"raw_usage":{"total_tokens":1916,"prompt_tokens":963,"completion_tokens":953,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":579,"completion_tokens_details":{"reasoning_tokens":883}},"tokens_in":579,"tokens_out":953,"duration_ms":7275,"temperature":1.0,"reasoning_tokens":883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:24:50.541424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the empirical sensitivity constant $L_1$ of Eq. (1) at the selected candidate layer on an out-of-distribution benchmark; if $L_1$ is effectively zero for OOD inputs even though $\\mathrm{NC}_1$ stays above 0.2 on the training set, the neural-collapse proxy has failed and PSC uncertainty would be overconfident.","supporting_citations":[{"cited_title":"Simple and principled uncertainty estimation with deterministic deep learning via distance awareness","cited_arxiv_id":null,"evidence_quote":"Defines distance-aware output layers and the spectral-normalization training that PSCs aim to replace for deterministic UQ."},{"cited_title":"Deep deterministic un- certainty: A new simple baseline","cited_arxiv_id":null,"evidence_quote":"Provides the DDU method (Gaussian discriminant analysis plus predictive entropy) and the experimental setup reused in the case studies."},{"cited_title":"Prevalence of neural collapse during the terminal phase of deep learning training","cited_arxiv_id":null,"evidence_quote":"Introduces the neural-collapse criteria NC1-NC4 that the layer-selection metrics approximate."},{"cited_title":"Feature learning in deep classifiers through intermediate neural collapse","cited_arxiv_id":null,"evidence_quote":"Supplies the intermediate neural-collapse metrics and the implementation used to measure NC1 and NC4 at every layer."},{"cited_title":"Vecchia Gaussian Process Ensembles on Internal Representations of Deep Neural Networks","cited_arxiv_id":"2305.17063","evidence_quote":"Demonstrates that intermediate-layer representations can support UQ in regression, the line of work PSCs extend to classification."},{"cited_title":"tucker\" ) # fit PSC. net = fit_psc( net=net, projection=projection, psc=","cited_arxiv_id":null,"evidence_quote":"Presents DUQ, a deterministic UQ method requiring RBF networks and modified training, used as a baseline for comparison."},{"cited_title":"Tensor decomposi- tions and applications","cited_arxiv_id":null,"evidence_quote":"Supplies the tensor-decomposition background and matricization convention needed for the projection of intermediate representations."}],"review_version":1}