{"id":"5433043e-bed2-42fb-9eee-8f6112e268ba","arxiv_id":"1908.10737","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Incorporating residual learning into a deep neural decision forest yields age-estimation accuracy comparable to the prior deep regression forest with about 4x less compute, plus gradient-based routing saliency maps.","lead":"A neural decision forest for facial age estimation is upgraded with residual network blocks, matching prior accuracy on some benchmarks while using about a quarter of the memory and compute, and the paper adds saliency maps that show which face regions drive the routing decisions.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Residual learning's contribution is untested: RNDF differs from DRFs in both shortcut connections and backbone, so the central mechanism is not isolated.","rationale":"The reader's weakest assumption pinpoints the same issue I find most load-bearing: the paper never isolates residual learning from the simultaneous backbone change. The central claim is not merely that RNDF is accurate and efficient, but that residual learning is what makes it so; without a shortcut-free ablation, the efficiency and accuracy results are consistent with the alternative explanation that a ResNet50-like backbone alone outperforms VGG-16. I also note that the 'state-of-the-art accuracy' wording is overstated relative to Table 1, since RNDF does not beat DRFs on FG-NET or MORPH, but that is secondary to the attribution problem. The proposed ablation directly settles the causal role of the residual mechanism, so the reader's conditional verdict is appropriate and no change to it is needed.","tokens_in":8720,"tokens_out":2769,"duration_ms":29904,"concrete_test":"Train the exact RNDF architecture with each residual block replaced by a plain bottleneck block (same conv layers and channels, no shortcut addition) on CACD and MORPH using identical hyperparameters, preprocessing, and forest settings; compare MAE, CS, model size, and FLOPs to RNDF. If the plain variant matches RNDF within expected run-to-run variation, residual learning is not the active ingredient; if the residual version is clearly better, the claim is supported. A secondary check is to train DRFs with the ResNet50-like backbone without residual connections, to separate backbone effects from forest-specific residual effects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3 interprets the comparison with DRFs in Tables 1 and 2 as evidence that residual learning improves NDF. But RNDF and DRFs differ in two variables at once: RNDF uses a ResNet50-like feature extractor with shortcut connections, while DRFs uses VGG-16, and no ablation removes only the shortcuts from RNDF. The accuracy differences (RNDF trails DRFs by 0.02 MAE on FG-NET and 0.06 MAE on MORPH, wins by 0.042 MAE on CACD) and the 4.8x memory and 4x FLOPs savings in Table 2 could therefore be caused entirely by the backbone change, not by residual learning. The central contribution, incorporating residual learning into NDF, is exactly the factor left uncontrolled. The absence of an ablation is the load-bearing gap; a plain-backbone variant must be shown to be worse before the title's mechanism can be credited.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Residual Neural Decision Forest (RNDF), which incorporates residual connections into the feature extractor of a deep neural decision forest, and applies it to facial age estimation on FG-NET, MORPH, and CACD. The authors report accuracy comparable to or slightly better than the prior DRFs model on CACD, with 4.8x memory savings and 4x FLOPs savings, and also introduce decision saliency maps (DSMs) to visualize the routing decisions. The abstract and introduction claim state-of-the-art accuracy and a first application of residual learning to NDF, with code and pre-trained models to be released.","tokens_in":8922,"tokens_out":3361,"duration_ms":35086,"significance":"If the central claim is validated, the contribution is an efficient age estimator with accuracy competitive with a strong prior model, plus a novel visualization method for NDF inference. The paper is clearly written, the training procedure is fully specified, and the promised code release is a concrete asset. However, the key attribution of the gains to residual learning is not supported by the experiments as presented: the comparison to DRFs changes both the residual connections and the backbone architecture at once, so the title's mechanism is not isolated. The accuracy differences from DRFs on two of three benchmarks are small and go in opposite directions, and no error bars are reported. The visualization results are qualitative and would benefit from a more rigorous assessment. With an added ablation and toned-down claims, the paper could be a solid empirical contribution.","major_comments":[{"comment":"The comparison to DRFs cannot support the claim that residual learning improves NDF, because RNDF and DRFs differ in two variables at once: RNDF uses a ResNet50-like feature extractor with shortcut connections, while DRFs uses VGG-16. No ablation removes only the shortcut connections from RNDF while keeping the remaining architecture identical. The reported accuracy differences (RNDF trails DRFs by 0.06 MAE on FG-NET and by 0.02 MAE and 4.5 CS points on MORPH, and leads by 0.042 MAE on CACD) and the 4.8x memory and 4x FLOPs savings in Table 2 could be caused entirely by the backbone change. The central contribution of the paper, incorporating residual learning into NDF, is exactly the factor left uncontrolled; please add an ablation with the same architecture but with the shortcut connections removed (or an equivalent control) before the title's mechanism can be credited.","section":"Section 4.3, Tables 1 and 2"},{"comment":"The abstract states the model 'achieves state-of-the-art level accuracy' and the introduction claims it 'achieves state-of-the-art accuracy', but Table 1 shows RNDF does not beat DRFs on FG-NET (2.97 vs. 2.91 MAE) or on MORPH (3.87 vs. 3.85 MAE; 76.1% vs. 80.6% CS). Only on CACD does RNDF improve over DRFs (4.595 vs. 4.637 MAE), and even there the difference is small. The claim should be restricted to 'comparable accuracy on some benchmarks with a large efficiency gain', or the experiments should be extended to support a true state-of-the-art claim.","section":"Abstract and Section 1"},{"comment":"No error bars or statistical significance tests are reported. The MORPH results are averages over 5 random splits, but the variance across splits is not given, and the differences with DRFs are small (0.06, 0.02, and 0.042 MAE). Without variance information or a paired test, it is impossible to tell whether RNDF is actually better, worse, or equivalent to DRFs on any of the benchmarks. Please report standard deviations (and, if available, the variance of the DRFs results under the same protocol) or otherwise justify that the differences are meaningful.","section":"Section 4.3 and Table 1"}],"minor_comments":[{"comment":"Typo: 'Redisual learning' should be 'Residual learning'.","section":"Figure 1 caption"},{"comment":"The definition Mk(x) = x + Hk(x) is informal: it should be stated that x is the input to the block and Hk is the residual mapping (e.g., a stack of convolution layers), and that the addition is performed element-wise on the feature maps; otherwise the notation is ambiguous.","section":"Section 3.1, Eq. (4)"},{"comment":"The decision saliency map is defined as the gradient of the routing probability si with respect to the input, but it is not specified which si is used when multiple splitting nodes are on the traced path; clarify whether the gradient is computed for each node individually and how the maps in Figures 3 and 4 are combined.","section":"Section 3.2, Eq. (11)"},{"comment":"The description 'Resnet50-like architecture' is too vague for reproducibility; please provide the exact block layout (number of layers per stage, filter sizes, strides) or refer the reader to a specific configuration in the released code.","section":"Section 4.2, Model Architecture"},{"comment":"The statement that the model 'is not sensitive to pre-processing' based on the qualitative DSMs (larger face regions in CACD, smaller in MORPH) is not supported by any quantitative comparison; please either add a controlled experiment or tone down the claim.","section":"Section 4.3, discussion of Figures 3 and 4"},{"comment":"Minor typos and inconsistent wording appear, e.g., 'paramters' in Algorithm 1 line 2, 'which is a another type' in Section 2, and inconsistent use of 'state-of-the-art level accuracy' vs. 'state-of-the-art accuracy'; please proofread carefully.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical study of an NDF variant for age estimation, but the main claimed mechanism (residual learning) is not isolated from a backbone change. The authors should be encouraged to add the missing ablation; if they do, the paper could be acceptable. The self-citation [18] is benign as it is the authors' own earlier workshop paper and the current work extends it to regression, but the relationship should be made clearer in the text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a competent, honest engineering paper that swaps the feature extractor in an existing deep regression forest (DRFs) for a ResNet50-style block and reports meaningful efficiency gains at roughly equal accuracy. The main flaw is that the central mechanism — residual learning — is never isolated, so the title's promise is not actually supported. The stress-test note gets this exactly right.\n\nWhat's new and what it does well. As far as I can tell, this is the first time residual connections have been used inside an NDF for age estimation, and the first time gradient-based saliency maps have been computed on routing probabilities. Those are modest firsts, but they are real. The experiments cover three standard benchmarks (FG-NET, MORPH, CACD), the implementation looks clean, and the efficiency numbers (4.8x smaller model, 4x fewer FLOPs) are credible. Accuracy is roughly on par with DRFs: slightly worse on FG-NET, slightly better on MORPH cumulative score, and better on CACD by 0.04 MAE. That is a reasonable trade-off for a 4-5x cheaper model.\n\nWhere the soft spots are. The big one, as noted, is the missing ablation. RNDF differs from DRFs in two ways at once: residual shortcuts and a ResNet50-like backbone instead of VGG-16. The accuracy differences and the efficiency gains could entirely come from the backbone change. Running a plain ResNet50-style NDF without the shortcuts would settle it. The introduction overclaims by saying \"state-of-the-art accuracy\" when Table 1 shows the model trails DRFs on FG-NET MAE and CS and on MORPH MAE; the abstract's \"state-of-the-art level\" is the fairer phrasing. There are also no error bars or significance tests, so the small metric differences are likely within noise. The saliency analysis is a standard gradient technique applied to routing scores; the observations are plausible but not deep.\n\nBottom line: this is a useful data point for people working on efficient age estimators or on NDFs in general. It is not a breakthrough, but it is not a waste of time either. The missing ablation is addressable, not load-bearing in the sense of invalidating the efficiency claim — it just means the residual mechanism remains unproven.\n\nMy recommendation: send it to peer review, but with the clear expectation that the authors either run the shortcut-removal ablation or reframe the contribution as \"an efficient NDF-based estimator\" rather than a validation of residual learning for NDFs. A serious referee would catch the gap, so the paper deserves the chance to fix it rather than a desk reject.\n\nBest,\n[You]","headline":"Solid engineering extension of deep regression forests with real efficiency gains, but the paper's central residual-learning claim is untested because the comparison changes both the shortcuts and the backbone.","tokens_in":9388,"tokens_out":2116,"would_cite":false,"duration_ms":25217,"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":"By adding residual shortcuts to neural decision forests, the model matches state-of-the-art age estimation accuracy while consuming 4.8x less memory and 4x less compute.","keywords":["facial age estimation","deep neural decision forest","residual learning","neural decision forest","decision saliency map","age regression","model efficiency","soft decision trees"],"falsifier":"Train the same residual-block network with the shortcut connection removed, so $M_k(x)=H_k(x)$ instead of $M_k(x)=x+H_k(x)$, keeping all other architecture, training, and data settings identical; if MAE, model size, and FLOPs stay the same, the residual mapping is not the cause of the reported gains.","tokens_in":8543,"feed_emoji":"","tokens_out":7143,"duration_ms":66208,"temperature":0.7,"pith_summary":"This paper brings residual learning—the skip-connection trick that helps train very deep neural networks—into deep neural decision forests (NDFs), where trees made of differentiable splitting functions learn to route an image to a leaf prediction. On three facial age estimation benchmarks, the resulting model, called RNDF, lands at state-of-the-art-level accuracy: it trails the prior deep regression forest (DRFs) slightly on FG-NET and MORPH, beats it on CACD, and does so with about 4.8x smaller model size and 4x fewer FLOPs. The paper also introduces decision saliency maps, gradients of routing probabilities with respect to the input, to expose which facial regions drive each decision. The central claim is that residual learning helps optimize the soft decision functions themselves, not just the feature extractor, and that this makes accurate age estimation cheaper to run.","feed_headline":"Residual shortcuts shrink age-estimation forests 4.8x","feed_subtitle":"Adding skip connections matches age-estimation accuracy while using a quarter of the compute.","key_machinery":"The central mechanism is the residual mapping inserted into each feature extraction stage of the neural decision forest: $M_k(x)=x+H_k(x)$, where $H_k$ is the learned transformation inside one residual bottleneck block. A deep neural decision forest is an ensemble of binary decision trees whose splitting nodes compute differentiable routing probabilities; the final prediction is the convex combination of leaf predictions weighted by path probabilities, and the paper sets each routing score as $s_i=\\sigma(f_i(M_n(\\dots M_1(x))))$. The identity shortcut is the load-bearing component claimed to ease optimization, allowing a ResNet50-like feature extractor to feed five depth-6 trees, with Gaussian covariance updates for leaf nodes alternating with gradient descent on the network parameters.","core_discovery":"The paper's central claim is that replacing plain feature mappings in a deep neural decision forest with residual mappings $M_k(x)=x+H_k(x)$ makes the soft routing functions easier to optimize. The resulting residual neural decision forest (RNDF) reports mean absolute errors of 2.97 on FG-NET (against 2.91 for DRFs), 3.87 on MORPH (against 3.85), and 4.595 on CACD (against 4.637), while model size drops from 539.4 MB to 112.4 MB and FLOPs drop from 16 G to 4 G. The paper further derives decision saliency maps, $\\partial s_i/\\partial x$, to visualize which image regions influence each routing decision along the most probable computation path.","pith_inferences":["Our inference: the reported savings may come largely from switching the feature-extraction network to a ResNet-style backbone, since the non-residual baseline uses a VGG-style network; an ablation with only the shortcuts removed would isolate the residual effect.","Our inference: aggregating decision saliency maps across all trees and leaves, rather than only the most probable path, could produce a more complete explanation heatmap and is a natural extension.","Our inference: if residual shortcuts are the cause, the same shortcut should improve other neural decision forest baselines, which is a testable prediction beyond the present benchmarks."],"forward_implications":["Facial age estimation can run with 4.8x less memory and 4x fewer FLOPs than the previous best neural-forest model at nearly unchanged accuracy, easing deployment on embedded or mobile devices.","Decision saliency maps make the forest's routing choices inspectable, so a user can check whether the model keys on facial skin rather than hair or background.","Residual learning transfers from plain CNNs to neural decision forests, suggesting the same shortcut can benefit other NDF regression tasks.","Because routing in this design uses a single shared fully-connected layer, the model can be extended by assigning splitting nodes to different feature layers or by learning non-binary tree topologies."],"supporting_citations":[{"why":"Supplies the previous deep regression forest whose accuracy, model size, and FLOPs are the main comparison, and the leaf-node Gaussian update rule adopted here.","marker":"[29]"},{"why":"Defines the residual learning hypothesis and the identity shortcut that the paper transfers into the forest's feature mappings.","marker":"[11]"},{"why":"Introduces the deep neural decision forest framework with soft splitting functions and path-weighted leaf predictions.","marker":"[16]"},{"why":"Provides the gradient-based saliency computation adapted here into decision saliency maps for routing probabilities.","marker":"[32]"},{"why":"Supplies the FG-NET dataset and the leave-one-subject-out cross-validation protocol.","marker":"[23]"},{"why":"Supplies the MORPH dataset and the selected-image training/testing split used in the experiments.","marker":"[25]"},{"why":"Supplies the CACD dataset with its fixed training, validation, and testing subsets.","marker":"[5]"}],"fun_headline_variants":["Residual skip connections slim age-estimation forests 4.8x","Residual decision forests: 4x less compute, same age accuracy","Residual learning makes decision forests 4x cheaper for age estimation","Skip connections give age-estimation forests a 4.8x shrink","Residual forests: same age accuracy, a quarter of the compute"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper credits residual learning for its gains, but the comparison model also uses a different feature-extraction network, so the residual shortcut itself is never tested on its own.","fun_headline_variants_meta":{"raw":{"variants":["Residual skip connections slim age-estimation forests 4.8x","Residual decision forests: 4x less compute, same age accuracy","Residual learning makes decision forests 4x cheaper for age estimation","Skip connections give age-estimation forests a 4.8x shrink","Residual forests: same age accuracy, a quarter of the compute"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000975,"raw_usage":{"total_tokens":4065,"prompt_tokens":792,"completion_tokens":3273,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":408,"completion_tokens_details":{"reasoning_tokens":3179}},"tokens_in":408,"tokens_out":3273,"duration_ms":24528,"temperature":1.0,"reasoning_tokens":3179,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:34:45.628224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same residual-block network with the shortcut connection removed, so $M_k(x)=H_k(x)$ instead of $M_k(x)=x+H_k(x)$, keeping all other architecture, training, and data settings identical; if MAE, model size, and FLOPs stay the same, the residual mapping is not the cause of the reported gains.","supporting_citations":[{"cited_title":"Overview of research on facial ageing using the fg-net ageing database","cited_arxiv_id":null,"evidence_quote":"Supplies the FG-NET dataset and the leave-one-subject-out cross-validation protocol."}],"review_version":1}