{"id":"093d1d58-e51f-4ca9-a1c5-8ffba0c8a44b","arxiv_id":"1908.05254","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Tree regularization, which penalizes the decision path length of a tree fitted to a deep network's predictions, produces deep models with higher accuracy at low complexity than L1 or L2 penalties.","lead":"This paper trains deep neural networks to be easy for people to step through by adding a penalty that pushes the network's decisions toward a small yes/no decision tree. On sepsis, HIV, and speech tasks, the resulting models stay accurate at low tree complexity, beating L1 and L2 penalties.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The low-APL advantage over L1/L2 may be an artifact: trade-off plots plot the deep network's AUC against distilled-tree APL, but fidelity or distilled-tree AUC for L1/L2 baselines is never reported; a human simulates the tree, not the network.","rationale":"I focused on the comparison that makes the central claim true. The paper's stated contribution is training-time regularization for simulability, and its main evidence is the low-APL/high-AUC curves. Those curves are produced by fitting a tree to each model for the APL axis, but evaluating the network for the AUC axis. The fidelity tables show the tree is a good surrogate for tree-regularized models, but without the same fidelity numbers for L1/L2, the comparison is asymmetric. This is not a purely aesthetic issue: simulability means a human can step through the tree, and if an L1/L2 model's tree has low fidelity, a human simulating that tree is not simulating the model, so the low APL is misleading. I considered the surrogate-accuracy concern the reader identified, but the paper provides some direct evidence for surrogate tracking in Figure 2 and Section 4, and even a perfect surrogate would not fix the missing fidelity baseline. The single concrete test that would settle the concern is to evaluate the distilled tree, not the network, for L1/L2 baselines. This is an added-experiment fix, so I keep the conditional verdict rather than rejecting outright.","tokens_in":36585,"tokens_out":6218,"duration_ms":66564,"concrete_test":"Use the same TrainTree/pruning procedure (Algorithm 2 with the same hyperparameters) to fit a decision tree to every L1/L2 model whose (APL, AUC) point appears in Figures 7, 8, 18, 19, and 20. For each fitted tree, compute (i) fidelity to the L1/L2 network and (ii) the held-out AUC of the tree itself. Redraw the trade-off curves with the y-axis set to distilled-tree AUC at matched APL, and separately report fidelity at APL values below 20. If L1/L2 distilled-tree AUCs drop below the tree-regularized curve, or if their fidelity is substantially below the tree-regularized fidelity, the claimed low-APL advantage holds; if not, the headline result is an artifact of comparing the network for baselines but the tree for the proposed method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 6.2 and Figures 5c-d, 7a-g, 8a-c, 18, 19, and 20 define the headline comparison: the x-axis is the APL of a decision tree fit to the model (Algorithms 1/2), while the y-axis is the held-out AUC of the deep network. For tree-regularized models, Table 1 and Table 4 report fidelity between 0.80 and 0.94, so the distilled tree is a reasonably faithful stand-in for the network. For L1/L2 baselines, no fidelity or distilled-tree AUC is reported. If an L1/L2 network's decision boundary is not axis-aligned, a low-APL tree fit to it can have low fidelity; the plotted point then pairs a low APL with a network AUC that a human simulating the tree would not actually achieve. The paper itself argues in Section 1 that post-hoc distillation of unregularized networks is ill-posed for exactly this reason. Thus the central claim, that tree regularization produces models that are easier to simulate than L1/L2 without sacrificing predictive power, is not established by the reported comparisons: the simulable artifact, the tree, is evaluated for tree-regularized models via fidelity, but never for the baselines. A proper comparison requires either reporting fidelity of L1/L2 distilled trees or plotting the held-out AUC of the distilled tree rather than the network.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes tree regularization as a way to train deep networks whose decision boundaries are well approximated by compact, axis-aligned decision trees. The regularizer measures the average path length (APL) of a decision tree fit to the network's predictions, and a separately trained surrogate MLP is used to make the APL penalty differentiable. The method is demonstrated on synthetic tasks, UCI data, and medical time series (sepsis/MIMIC, HIV/EuResist, TIMIT), and is extended to regional tree regularization in which different, expert-defined regions of the input space are each matched to a compact tree, with sparsemax used to select the most complex region. The central empirical claim is that tree-regularized models achieve higher held-out AUC at low APL than models trained with L1 or L2 regularization.","tokens_in":36929,"tokens_out":4860,"duration_ms":49858,"significance":"If the central comparison is established, this is a valuable contribution: it moves interpretability from post-hoc explanation to direct optimization, introduces a non-trivial differentiable proxy for a discrete tree-complexity measure, and demonstrates the idea across several realistic tasks. The regional extension, with sparsemax over per-region APLs, is a reasonable and potentially useful generalization, and the paper reports fidelity of the distilled trees to the deep models, which is the right kind of evidence for simulability. The main empirical claim, however, is not fully supported as reported because the baseline comparison is asymmetric: the x-axis is the APL of a distilled tree, but the y-axis is the network's held-out AUC, and the fidelity of the distilled tree to L1/L2-regularized networks is never reported. The central result therefore needs additional evidence before the claimed 'easier to simulate without sacrificing predictive power' can be accepted.","major_comments":[{"comment":"The headline comparison is asymmetric. In the trade-off plots, the x-axis is the APL of a decision tree fitted to the model (Algorithm 1 or Algorithm 2), while the y-axis is the held-out AUC of the deep network itself. Fidelity of the distilled tree to the deep model is reported only for tree-regularized models (Tables 1 and 4, values 0.80-0.94). For L1/L2 baselines, no fidelity or distilled-tree AUC is reported. Because a human simulating the model would simulate the tree rather than the network, the claim that tree regularization yields models that are easier to simulate than L1/L2 without sacrificing accuracy is not established: a low-APL tree fitted to an L1/L2 network could be unfaithful, in which case the plotted low APL would not be attainable by a human simulating the actual model. This is exactly the ill-posedness the paper itself attributes to post-hoc distillation in Section 1. Please report fidelity or held-out AUC of the distilled trees for L1/L2 baselines, or re-plot the comparison using the distilled tree's own AUC on the y-axis.","section":"Section 6.2, Figures 5c-d, 7a-g, 8a-c, 18-20, Tables 1 and 4"},{"comment":"The validity of the entire optimization depends on the surrogate function \\hat{\\Omega}(\\theta) providing accurate and smooth estimates of the true APL along the optimization trajectory. The paper shows that heuristics (data augmentation, random restarts, a sliding window) are needed even on a 2D toy problem, and Section 7.4 documents instability in the regional setting. However, no real-data diagnostic of surrogate accuracy is reported, such as MSE or correlation between \\hat{\\Omega}(\\theta) and the true APL on held-out parameter vectors. If the surrogate's gradient is not a reliable descent direction, the regularization signal is noise and the empirical results may not reflect the proposed objective. Please include surrogate-versus-true-APL diagnostics for at least one real task, or an ablation showing that the downstream AUC/APL conclusions require an accurate surrogate.","section":"Section 4, Eq. (10), Figure 2; Section 7.4"},{"comment":"No confidence intervals or significance tests are reported for the central trade-off curves. Figure 18 states that the plotted points are minima from 3 independent runs but does not show error bars or variance; Figures 5, 7, and 8 appear to show single runs. The abstract and Section 6.2 claim that tree regularization 'consistently' achieves higher accuracy in the low-APL regime. Given that the reported AUC differences are sometimes modest (0.03-0.15), the absence of repeated-run variability makes it difficult to judge whether the advantage is reliable. Please report the number of runs and the variance (e.g., confidence intervals or standard errors) for each curve, or state clearly which curves are single runs.","section":"Section 6.2 and Figures 5, 7, 8, 18"},{"comment":"The regional evaluation metric is subtly different from the global objective. For regional experiments, the reported test APL is always computed as \\Omega^{regional}(\\theta), even for globally regularized models, which are trained to minimize the global APL. The paper explains that this is to make models comparable, but this means the global tree models are evaluated on a metric that is not their training objective. Please make this mismatch explicit in the main text and discuss whether it could systematically disadvantage global tree regularization, since the global objective may allow one region to be complex as long as the overall APL is low, while the regional evaluation penalizes that complexity.","section":"Section 9.1 and Figures 18-20"}],"minor_comments":[{"comment":"The text says 'Figure 4(b) shows each trained model as a single point in a 2D fitness space,' but Figure 4 contains decision boundaries; the fitness-space plot appears to be Figure 3(b). Please correct the cross-reference.","section":"Section 5, discussion of 2D Parabola"},{"comment":"The body text refers to 'Table 15' ('Table 15 shows this to reduce noise'), while the cited entity appears to be the MSE table in Figure 15. Please unify the numbering and caption.","section":"Section 7.4, Figure 15"},{"comment":"Algorithm 1 is described as an APL cost function but returns only the mean path length; please define explicitly that the tree is trained on the thresholded predictions of the network and that the path length is computed on that tree, and state what happens when the tree's class differs from the network's probability output.","section":"Algorithm 1 and Equations (9)-(10)"},{"comment":"The HMM objective in Equation (8) is written as a maximum over a joint probability, but the earlier objectives (1) and (6) include a regularization term; it is unclear whether tree regularization is applied to the HMM component alone, the GRU component alone, or both in the GRU-HMM experiments. Please clarify the exact objective optimized for GRU-HMM.","section":"Equation (8)"},{"comment":"The term 'L0 regional tree regularization' is used for the sparsemax-based method, while Equation (13) is a max over regions. Since sparsemax is a smooth approximation and not an exact L0 norm, the terminology could be misleading; please clarify the relationship or choose a different name.","section":"Section 7.2 and Equation (13)"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a genuinely useful idea and an impressive volume of experiments, but the central empirical claim needs one additional piece of evidence: fidelity or distilled-tree AUC for the L1/L2 baselines. Without that, the main comparison is not apples-to-apples. The surrogate-accuracy diagnostics and variance reporting are also important for the paper to be persuasive. I would be willing to look at a revised version. Releasing code for the surrogate training would also strengthen reproducibility, though I did not treat this as a blocking issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the core idea is new and worth taking seriously: instead of distilling a trained network post hoc, they add a differentiable penalty that encourages the network to be well-approximated by a compact axis-aligned decision tree. The surrogate-network trick for estimating average path length is a real technical contribution, and the regional extension using sparsemax is a sensible way to keep all regions simple rather than letting one collapse. Second, the empirical story is broader and more careful than most interpretability papers: MIMIC, HIV, TIMIT, UCI benchmarks, fidelity numbers, and a stability analysis over random restarts. That is real work.\n\nThe soft spot is the central comparison. The trade-off plots put the deep network's held-out AUC on the y-axis and the distilled tree's APL on the x-axis. For tree-regularized models they report fidelity (0.80–0.94), so the tree is a reasonably faithful stand-in for the network. For L1 and L2 baselines they report neither fidelity nor the distilled tree's AUC. If an L1/L2 network's boundary is not axis-aligned, the tree fit to it can have low fidelity, and the plotted point then pairs a low APL with an AUC that a human simulating the tree would not actually get. That is not a minor quibble: the paper itself argues in the introduction that post-hoc distillation of unregularized networks is ill-posed for exactly this reason. So the headline claim—tree regularization gives models that are easier to simulate than L1/L2 without sacrificing predictive power—is not fully established by the reported comparisons. The fix is straightforward: report fidelity for the baselines' distilled trees, or plot the distilled tree's AUC instead of the network's.\n\nThere is also a mild circularity issue: APL is both the objective and the primary evaluation metric, so \"tree regularization produces simpler models\" is partly true by construction. The accuracy-at-matched-APL comparison is external and held-out, so the main claim survives, but confidence intervals or significance tests would strengthen it. The surrogate training involves several heuristics (augmentation, restarts, sliding windows), and those are free parameters that could be hard to transfer, though the paper is honest about them.\n\nBottom line: this deserves a serious referee. The idea is sound, the experiments are extensive, and the missing baseline fidelity is a fixable gap rather than a fatal flaw. I would send it out, and I would ask for that comparison before accepting.","headline":"A genuinely new training-time regularizer for simulability, with broad experiments, but the headline comparison is weakened by an asymmetry: network AUC is plotted against distilled-tree path length, and fidelity is only reported for tree-regularized models, never for L1/L2 baselines.","tokens_in":37428,"tokens_out":1418,"would_cite":true,"duration_ms":14941,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that training deep networks to be well approximated by compact decision trees makes them simulable by humans while retaining accuracy.","keywords":["tree regularization","interpretability","human-simulability","average path length","decision tree","neural network regularization","regional explanations","sparsemax"],"falsifier":"Refit a decision tree to the tree-regularized network at every checkpoint, compute the true average path length, and compare it with the surrogate's estimate; if the two diverge as training progresses, or if the final network's average path length at matched accuracy is not below that of L1- or L2-regularized networks, the central claim fails.","tokens_in":36395,"feed_emoji":"🌳","tokens_out":10010,"duration_ms":84996,"temperature":0.7,"pith_summary":"The paper tries to establish that a deep network can be trained to be interpretable, rather than interpreted afterward. Its proposal, tree regularization, adds a penalty to the training objective that rewards the network whenever its predictions can be reproduced by a compact axis-aligned decision tree, measured by the average number of decision nodes a human would have to step through. Because such trees can be followed by hand, a network that is well approximated by one becomes simulable by a clinician, regulator, or other user. Across synthetic problems, speech data, ICU records, and HIV treatment histories, the authors report that tree regularization reaches higher accuracy at low decision-tree complexity than L1 or L2 penalties, and that a regional version, fitting a separate compact tree for each expert-defined subpopulation, does even better.","feed_headline":"Tree penalty makes neural nets simulable without losing accuracy","feed_subtitle":"Adding a tree-fit penalty yields deep models humans can step through, with better accuracy at low complexity.","key_machinery":"The central object is the average-path-length (APL) regularizer: fit an axis-aligned decision tree to the network's thresholded predictions on a reference set, then measure the average number of nodes traversed from root to leaf. Since tree fitting is not differentiable, the paper replaces APL with a small surrogate neural network that estimates APL from the target network's parameter vector; the surrogate is trained on parameter/APL pairs collected during optimization, using data augmentation, random restarts, and a sliding window of recent parameters to keep the estimate accurate. For regional tree regularization, one surrogate is maintained per human-defined region, and the sparsemax transform over regional APL estimates makes the max operator differentiable while penalizing only the most complex regions, so simple regions are not over-regularized.","core_discovery":"The central claim is that the usual accuracy-interpretability trade-off can be shifted by optimizing for a proxy of simulability during training. The paper's regularizer penalizes the average path length of a binary decision tree fitted to the target network's thresholded predictions; by making this penalty differentiable through a surrogate estimator of path length, gradient descent can find parameter minima whose decision boundaries are nearly tree-shaped. The authors assert that this produces networks that stay accurate while their distilled decision trees are short enough for a person to simulate, with fidelity between the network and its tree around 0.80–0.94 on real-world tasks. For settings where no single global tree can be both small and faithful, the paper extends the penalty region by region, using sparsemax to concentrate regularization on the most complex regions, and reports that this regional version finds better accuracy-versus-complexity trade-offs than any other regularizer tested.","pith_inferences":["Editorial inference: average path length measures tree size, not actual human effort, so a fair test of the interpretability claim would be a user study where people simulate predictions from the distilled trees and compare error rates and response times against L1- or L2-regularized models at matched tree size.","Editorial inference: if the surrogate can learn the parameter-to-APL mapping, the same trick should apply to other discrete complexity measures, such as the number of distinct activation patterns or the description length of a rule list, suggesting a general recipe for procedure regularization beyond trees.","Editorial inference: the regional framework treats any user-defined partition as a place to inject prior knowledge; a natural extension would be to partition by protected attributes so the resulting model has simple, inspectable behavior within each demographic group.","Editorial inference: because fidelity is high but not perfect, the residual disagreement between network and tree defines a small set of hard cases; inspecting those cases could reveal where the tree explanation should not be trusted, which would make the method more usable in practice."],"forward_implications":["Deep networks trained this way could be deployed in decision-critical settings where a human must audit each prediction, because the distilled tree gives the step-by-step logic and the network retains accuracy on cases the tree handles poorly.","If regional tree regularization works as claimed, expert knowledge about meaningful patient or input subgroups can be turned directly into a training constraint, yielding per-group explanations without requiring global simplicity.","Because the penalty is written as a regularizer on any differentiable model, the same procedure could make recurrent, convolutional, or hybrid sequence models simulable, not just plain feed-forward networks.","The reported accuracy gains over L1 and L2 at low average path length imply that norm-based penalties are not the right tool if the goal is human simulation; the tree regularizer targets the complexity measure that actually matters for step-through reasoning."],"supporting_citations":[{"why":"Defines the fidelity measure and the technique of extracting tree-structured representations from trained networks that the regularizer turns into a training objective.","marker":"Craven & Shavlik, 1996"},{"why":"Supplies the premise that deep networks have multiple optima of similar accuracy, so regularization can select simulable minima without a large accuracy cost.","marker":"Goodfellow et al., 2016"},{"why":"Provides the definition of human-simulability that motivates measuring interpretability by decision-tree path length.","marker":"Lipton, 2016"},{"why":"Introduces the sparsemax transformation used to make the regional max over path lengths differentiable and sparse.","marker":"Martins & Astudillo, 2016"},{"why":"Represents prior work on optimizing models for interpretability by constraining input sensitivity, which the paper distinguishes from the stronger goal of simulation.","marker":"Ross et al., 2017"},{"why":"Supports the psychological premise that human understanding is context-dependent, motivating regional explanations.","marker":"Miller, 2018"},{"why":"Provides the efficient Euclidean projection onto the simplex used in the sparsemax computation for regional regularization.","marker":"Duchi et al., 2008"}],"fun_headline_variants":["Tree penalty makes deep nets simulable, accuracy intact","Neural nets that mimic short trees, keep predictive power","Step-through-able AI models via tree regularization","Tree-regularized training: accurate and interpretable","Penalty for tree-like decisions, no accuracy sacrifice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the assumption that a small helper network can accurately and smoothly learn the mapping from the main network's parameters to the average path length of a decision tree fitted to that network; if that mapping is not learnable along the training trajectory, the regularization gradient is noise.","fun_headline_variants_meta":{"raw":{"variants":["Tree penalty makes deep nets simulable, accuracy intact","Neural nets that mimic short trees, keep predictive power","Step-through-able AI models via tree regularization","Tree-regularized training: accurate and interpretable","Penalty for tree-like decisions, no accuracy sacrifice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":1967,"prompt_tokens":947,"completion_tokens":1020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":944}},"tokens_in":563,"tokens_out":1020,"duration_ms":10549,"temperature":1.0,"reasoning_tokens":944,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:19:51.543112+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Refit a decision tree to the tree-regularized network at every checkpoint, compute the true average path length, and compare it with the surrogate's estimate; if the two diverge as training progresses, or if the final network's average path length at matched accuracy is not below that of L1- or L2-regularized networks, the central claim fails.","supporting_citations":[{"cited_title":"\\ \\ Shavlik, J","cited_arxiv_id":null,"evidence_quote":"Defines the fidelity measure and the technique of extracting tree-structured representations from trained networks that the regularizer turns into a training objective."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the premise that deep networks have multiple optima of similar accuracy, so regularization can select simulable minima without a large accuracy cost."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the definition of human-simulability that motivates measuring interpretability by decision-tree path length."},{"cited_title":"\\ \\ Astudillo, R","cited_arxiv_id":null,"evidence_quote":"Introduces the sparsemax transformation used to make the regional max over path lengths differentiable and sparse."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supports the psychological premise that human understanding is context-dependent, motivating regional explanations."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the efficient Euclidean projection onto the simplex used in the sparsemax computation for regional regularization."}],"review_version":1}