{"id":"f6fa3596-f0d4-4b70-9a15-5d9269f0cfbe","arxiv_id":"2506.15559","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"LogNet replaces neural network layers with fixed logic gates on binarized Wi-Fi signals, reporting lower localization error and improved interpretability over time.","lead":"This paper proposes LogNet, a Wi-Fi indoor localization model built from simple logic gates that turns received signal strengths into binary yes/no patterns. The authors claim it is easier to interpret, more robust to noise over two years, and smaller than existing deep learning models.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"LogNet's interpretability claim is unsupported: its binary latent space is never compared against the DNN's own attribution or decision logic, so the central claim that LogNet interprets DNN behavior does not follow from the experiments.","rationale":"The reader's weakest_assumption focuses on the fixed threshold phi and the test-data-driven choice of gate, and on whether the binary pattern is a faithful interpretation of DNN reasoning. I agree with the broader concern, but I would center it more sharply: the interpretability claim fails even before threshold sensitivity is examined, because the paper never tests whether LogNet's logic-derived AP importance matches the DNN's actual learned importance. The comparison in Figure 6 is purely visual and side-by-side; it shows that LogNet's latent space is binary and traceable, not that it explains DNN-DownSample's continuous latent encodings. The paper does provide real evidence for some claims: a two-year real-world dataset is collected, code and data appear to be open-sourced, and LogNet's deterministic architecture is reproducible and does not require training of hidden layers. Those support a revised claim that LogNet is a lightweight, transparent alternative to DNNs for fingerprinting localization. But the abstract and conclusions assert a stronger claim: that LogNet interprets the internal behavior of DL models and can be used to diagnose DNN failures. That claim would require a quantitative comparison between LogNet's critical bits and the DNN's per-RP input attributions or decision-boundary geometry. The proposed test would settle the question directly. If the agreement is low, the paper's central novelty is unsupported and the strongest defensible contribution is a compact rule-based localizer, not an explanation method for DNNs.","tokens_in":10134,"tokens_out":2615,"duration_ms":32153,"concrete_test":"After training DNN-DownSample on CI:0, compute per-RP AP influence using input-gradient, integrated gradients, or SHAP values from the trained DNN, and extract the top-k APs per RP. Independently, extract LogNet-NOR's critical bit positions and map them back to AP pairs for the same RPs. Report top-k agreement (e.g., mean Jaccard or rank correlation) over all RPs in Building 1. If agreement is near random, LogNet does not interpret DNN behavior. As an additional control, re-run LogNet with phi in {0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8} and select the gate on the training CI only; if the 1.1x-2.8x gain or the bit assignments change materially, the headline comparisons depend on test-set selection.","verdict_should_be":"REJECT","load_bearing_attack":"The paper's central claim is that LogNet 'interprets the internal behavior of DL models' (Abstract, Section VI). For this to hold, LogNet's per-RP critical bits must correspond to the AP-importance structure that the DNN actually learned. The paper never establishes that correspondence. LogNet is constructed by binarizing RSS with a fixed threshold phi=0.5 and applying NOR gates; its latent space is a function of the input alone, not of DNN weights, activations, or gradients. Figure 6 only shows the two latent spaces side-by-side; it does not show that LogNet's bit positions 12/46 (RPs 5/6) or 6/41/42/49/61/81 (RPs 59/60) coincide with the APs that drive DNN-DownSample's predictions. Since LogNet's gates are chosen by evaluating all six gates on the test data (Section V.B) and the threshold is fixed without sensitivity analysis, even the performance claim is not robustly separated from selection-on-test-data. But the more load-bearing gap is interpretability: a fixed binary encoder may be a useful lightweight model, yet it provides no evidence about how DNNs reason. Without a mapping from LogNet's logic rules to DNN decision boundaries, the abstract's claim that LogNet explains internal DL behavior does not follow.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LogNet, a logic-gate-based architecture for Wi-Fi fingerprint indoor localization. LogNet binarizes RSS values with a fixed threshold, applies pairwise logic gates (e.g., NOR) over adjacent AP pairs across several layers, and trains only a final softmax classifier. The authors claim that LogNet interprets the internal behavior of DNN-based localization models, identifies the most influential APs per reference point, and diagnoses failures under temporal variation, while also improving localization error, model size, and latency relative to prior DNN models. Experiments use a two-building, ten-collection-instance dataset collected over two years, comparing LogNet variants against DNN-DownSample and prior DNN architectures.","tokens_in":10548,"tokens_out":4049,"duration_ms":52612,"significance":"If the central interpretability claim were established, LogNet would be a useful contribution to explainable indoor localization, and the open-sourced dataset [22] is a valuable asset for reproducibility. The paper also demonstrates a lightweight, low-latency alternative that may be attractive for deployment. However, as written, the paper does not show that LogNet explains the decisions of a trained DNN; it presents a standalone deterministic binary encoder whose latent space is never quantitatively connected to DNN weights, activations, or attributions. The performance comparison also contains a selection-on-test-data issue in choosing the NOR gate. These are load-bearing gaps rather than presentation issues, so the manuscript requires substantial revision before its claims can be accepted.","major_comments":[{"comment":"The central claim that LogNet “interprets the internal behavior of DL models” is not supported by the experiments. LogNet is constructed by binarizing RSS with a fixed threshold and applying deterministic logic gates; its latent space is a function of the input alone, not of the DNN's weights, activations, or gradients. Section V.C compares the two latent spaces side by side, but it never establishes that LogNet's critical bits correspond to the APs that actually drive DNN-DownSample's predictions. A quantitative test, such as verifying that flipping LogNet-identified critical APs changes the DNN's output in the predicted direction, or comparing LogNet's per-RP AP importance against DNN attribution methods, is needed. Without such evidence, the abstract's and conclusions' interpretability claims should be revised to describe LogNet as a transparent alternative model rather than an interpretation of DNN behavior.","section":"Abstract and Section VI"},{"comment":"The selection of the NOR gate is performed on the test data, which biases the reported performance. Section V.B evaluates all six LogNet variants on the CI:0 holdout fingerprint and then selects the best-performing gate (NOR) for subsequent comparisons. This is a form of test-set selection and inflates the reported gains. The gate choice should be made on a validation split or via nested cross-validation, and the paper should report performance for all gates without selecting on test data. Relatedly, the binarization threshold ϕ=0.5 in Section IV is fixed without sensitivity analysis; because the entire architecture and interpretability claims depend on this threshold, the paper should show results across a range of thresholds and justify the chosen value.","section":"Section V.B and Figure 4"},{"comment":"The DNN-DownSample baseline appears undertuned, making the reported performance gains less convincing. At CI:0, DNN-DownSample achieves a mean localization error of 5.81 m on a 61-RP building, which is high relative to typical Wi-Fi fingerprinting results, and the paper reports no hyperparameter search, regularization, early stopping, or ensemble averaging for the DNN. Training for 500 epochs at a fixed learning rate of 0.01 may not converge to a competitive baseline. The paper should tune the DNN baselines fairly (including prior architectures compared in Section V.D) and report standard errors, confidence intervals, or significance tests across the six devices and two buildings. Without this, the claimed 1.1x–2.8x improvements are not robustly established.","section":"Section V.A and Figure 4"},{"comment":"The interpretability analysis in Section V.C is based on manual inspection of two RP pairs (RPs 5/6 and 59/60) and does not provide a validated method for tracing latent bits to AP-level influence. The paper states that DNN-DownSample's varied bits are “not truly discriminative” and that LogNet exposes “the correct structure of the decision boundary,” but no ground-truth or quantitative measure of discriminative APs is given, and no stability analysis across collection instances is provided. The claim that LogNet filters non-Euclidean noise should be tested by measuring bit-flip rates under temporal variation and by showing that the identified critical bits are stable and indeed correspond to APs whose RSS values are informative for the RP pair.","section":"Section V.C"},{"comment":"The comparison against prior DNN-based models (AUTLOC, SANGRIA, DNNLOC) is not controlled: the prior models are reimplemented with different architectures and training settings, and no details are given for how their hyperparameters were selected. The latency measurements in Table 2 are reported without specifying the hardware, framework, or batching conditions, making the 1.5x–3.6x latency claims difficult to verify. A fair head-to-head comparison should use the same dataset, training pipeline, and inference environment, and should report variance across runs.","section":"Section V.D and Table 2"}],"minor_comments":[{"comment":"Algorithm 1 trains the softmax layer using backpropagation, so LogNet is not fully rule-based; the final layer is learned. The text should state this clearly to avoid overstating the absence of training.","section":"Section IV, Algorithm 1"},{"comment":"The pairing of APs into adjacent pairs is presented without justification. The paper should discuss whether the ordering of AP indices has any physical or semantic meaning and whether results depend on the pairing scheme.","section":"Section III.B"},{"comment":"The phrase “all models containing 1 hidden layer” is ambiguous for LogNet because the logic layers are not trained; specifying the exact layer counts and parameter counts would improve reproducibility.","section":"Section V.B"},{"comment":"The heading contains a typo: “Comparision” should be “Comparison.”","section":"Section V.D heading"},{"comment":"The paper says the data is open-sourced [22], but no link to code for LogNet is provided. Releasing the implementation would strengthen reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's core claim is currently overbroad: LogNet is a transparent model, not an explanation of a trained DNN. The selection-on-test-data issue in Section V.B and the undertuned DNN baseline undermine the performance claims. These problems are substantial but addressable by re-scoping the claims and redoing the evaluation with proper validation and fair baselines. The open dataset is a positive feature, and the lightweight architecture may have value even without the DNN-interpretation framing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is a mixed bag. The useful parts are the new two-year Wi-Fi fingerprint dataset and the demonstration that a fixed logic-gate network (binarize RSS, apply pairwise NOR gates, train a softmax) can beat a DNN baseline under temporal drift. The dataset is open-sourced, which is real value. The central claim in the title and abstract—that LogNet interprets what a DNN has learned—is not supported.\n\nWhat the paper does well: it identifies a real problem (temporal drift breaks DNN fingerprint localization), collects a longitudinal dataset with six phones and two buildings, and shows that a trivial, near-training-free architecture gets lower error and much smaller model size than the DNN baselines. If that holds up, it is a useful lightweight alternative for mobile deployment. Releasing the data is a plus.\n\nThe soft spots are load-bearing. LogNet is a standalone binary feature extractor; it never examines the DNN's weights, gradients, or attributions. Figure 6 just shows two latent spaces side-by-side; it does not demonstrate that LogNet's critical bit positions correspond to the APs that drive the DNN's decisions. So the phrase \"interprets the internal behavior of DL models\" overstates what the experiments show. The DNN baseline error of 5.81 m at CI:0 seems high for a single-fingerprint test; if the baseline is undertuned, the performance comparisons are not conclusive. There is also no sensitivity analysis for the binarization threshold (fixed at 0.5), and the gate type is chosen after evaluating all six gates on the same test set, which is a mild selection-on-test problem. None of this kills the paper as a lightweight-model contribution, but it does kill the interpretability claim.\n\nWho is this for? Researchers working on indoor localization with temporal drift, and anyone interested in when a simple baseline can beat a DNN. It is not a paper about explaining DNNs.\n\nMy recommendation: I would not accept the interpretability claim as stated. But the dataset and the method deserve a serious referee—send it to review with a request to reframe it as a lightweight alternative model and add a quantitative comparison to eXLoc and XGBoost, which are cited but never compared. A revision could be solid.","headline":"LogNet is a simple binarized logic-gate classifier that beats a DNN on a new two-year Wi-Fi dataset, but the claim that it interprets DNN behavior is not supported by the experiments.","tokens_in":10941,"tokens_out":2001,"would_cite":true,"duration_ms":20433,"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":"LogNet claims a fixed logic-gate architecture can interpret and outperform DNNs on Wi-Fi indoor localization.","keywords":["Explainable AI","Indoor Localization","Temporal Variations","Logic Gate Networks","Wi-Fi fingerprinting","received signal strength","latent space interpretability","non-Euclidean noise"],"falsifier":"Sweep the binarization threshold from 0 to 1 on the same two-building dataset and retrain the softmax layer for each value; if LogNet-NOR's mean localization error changes substantially with the threshold, the fixed 0.5 cutoff is load-bearing and the reported robustness is not intrinsic to the logic-gate design.","tokens_in":9924,"feed_emoji":"📡","tokens_out":8887,"duration_ms":91015,"temperature":0.7,"pith_summary":"LogNet is an attempt to open the black box of deep-learning-based Wi-Fi indoor localization without giving up accuracy. It binarizes received signal strength fingerprints and passes adjacent pairs through fixed logic gates, layer by layer, producing a compact binary latent code before a final trained softmax layer. The paper argues this code is interpretable: each bit corresponds to a rule about a pair of access points, so a prediction can be traced back to specific APs. On fingerprints collected over two years in two buildings, the authors report that the NOR-gate variant of LogNet localizes with 1.1–2.8× lower mean error, needs 3.4–43.3× fewer parameters, and runs 1.5–3.6× faster than prior DNN-based models while staying stable under temporal drift.","feed_headline":"Logic-gate net explains Wi-Fi localization and beats deeper nets","feed_subtitle":"Fixed NOR-gate rules trace decisions to specific access points, stay accurate over 2 years, and shrink models by up to 43x.","key_machinery":"The load-bearing object is the logic-gate layer built on binarized RSS values. Each fingerprint is normalized, thresholded at $\\phi=0.5$ to turn APs into active or inactive bits, and adjacent pairs are fed through one fixed gate ($Z=1-(X+Y)$ for the best-performing NOR variant). The outputs are paired again in each successive hidden layer, halving dimensionality each time, until a compact binary latent vector is formed; only the final softmax layer is trained. Because every bit in that latent vector is the deterministic output of a gate applied to a specific pair of upstream bits, each decision can be traced back to a rule about two access points. The contrasting assumption, that DNNs treat fingerprint noise as uniform Euclidean noise while real temporal noise is per-AP and non-Euclidean, is what the gates exploit.","core_discovery":"The central claim is that a fixed, untrained stack of logic gates can reproduce, and improve on, what DNNs learn for Wi-Fi fingerprint localization, and that the resulting binary latent space is interpretable. LogNet's design assumes that the key information in an RSS fingerprint is which access points are jointly strong or jointly weak, not the exact signal values. Under that view, non-Euclidean temporal noise, which perturbs different APs by different amounts, is filtered out by the discrete gate operations. The paper shows latent-space comparisons where a DNN separates two reference points using continuous-valued bits that are actually noise-susceptible, while LogNet separates the same reference points with stable binary bits that trace back to specific AP-pair rules. The claimed payoff is that the model becomes diagnosable: a failure can be attributed to particular access points, and the same rule-based encoding is what keeps mean localization error low over two years of environmental change.","pith_inferences":["A natural test the paper does not run is to sweep the binarization threshold and gate choice per layer; if performance is highly sensitive to the 0.5 cutoff, the fixed threshold is doing more work than the logic-gate concept itself.","The same pairwise binarization scheme transfers in principle to other RF fingerprint modalities such as Bluetooth or UWB; if it transfers, the interpretability mechanism is not specific to Wi-Fi.","Because each latent bit names an AP-pair rule, LogNet could be inverted to recommend which access points to keep or reposition in a deployment, a use the authors gesture at but do not develop.","The interpretability claim implicitly asserts that DNNs would have learned those same AP-pair rules; an explicit comparison of LogNet's bit-level attribution against gradient-based attribution on the DNN would test that equivalence."],"forward_implications":["If LogNet's claims hold, Wi-Fi indoor localization can be made auditable: each prediction carries a rule trace showing which access points were decisive.","The reported parameter counts and latency suggest the approach can run on mobile or embedded devices, where the DNN baselines were substantially larger and slower.","Because the latent codes are binary and deterministic, model failure under temporal drift can be diagnosed by comparing which bits flip between training and test time, pointing to the access points responsible.","The method's stability across one to four hidden layers implies that depth can be chosen to trade size against accuracy without retraining the gate stack.","The reported gains over DNNLOC, SANGRIA, and AUTLOC would make logic-gate networks a competitive baseline for future fingerprint-based indoor localization research."],"supporting_citations":[{"why":"DNNLOC, the four-hidden-layer DNN baseline whose error, size, and latency LogNet is compared against.","marker":"[15]"},{"why":"SANGRIA, a three-hidden-layer stacked-autoencoder baseline in the hidden-depth comparison.","marker":"[20]"},{"why":"AUTLOC, a two-hidden-layer autoencoder baseline used in the depth and efficiency comparison.","marker":"[21]"},{"why":"eXLoc, the explainable-localization work that motivates the need for interpreting DNN decisions in fingerprinting.","marker":"[9]"},{"why":"RADAR, the classic Wi-Fi fingerprinting method whose assumptions LogNet builds on conceptually.","marker":"[14]"},{"why":"The open-sourced two-building, two-year RSS dataset on which all models are trained and evaluated.","marker":"[22]"},{"why":"STELLAR, which documents temporal variation and device heterogeneity in indoor localization, framing the long-term robustness problem.","marker":"[8]"}],"fun_headline_variants":["Logic gates decode Wi-Fi maps, outperform deep nets","Explainable Wi-Fi locator: fixed gates, 43x smaller, 2-year stable","No deep net needed: logic gates localize Wi-Fi better","Wi-Fi localization that explains itself: logic gates win","Binary logic beats black box for Wi-Fi indoor location"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the information a DNN uses for Wi-Fi localization can be faithfully captured by fixed pairwise logic rules on binarized signal strengths, so the specific choice of threshold (0.5) and gate (NOR) is what makes both the interpretability and performance claims true.","fun_headline_variants_meta":{"raw":{"variants":["Logic gates decode Wi-Fi maps, outperform deep nets","Explainable Wi-Fi locator: fixed gates, 43x smaller, 2-year stable","No deep net needed: logic gates localize Wi-Fi better","Wi-Fi localization that explains itself: logic gates win","Binary logic beats black box for Wi-Fi indoor location"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000576,"raw_usage":{"total_tokens":2733,"prompt_tokens":977,"completion_tokens":1756,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":593,"completion_tokens_details":{"reasoning_tokens":1679}},"tokens_in":593,"tokens_out":1756,"duration_ms":12963,"temperature":1.0,"reasoning_tokens":1679,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:52:10.851648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Sweep the binarization threshold from 0 to 1 on the same two-building dataset and retrain the softmax layer for each value; if LogNet-NOR's mean localization error changes substantially with the threshold, the fixed 0.5 cutoff is load-bearing and the reported robustness is not intrinsic to the logic-gate design.","supporting_citations":[{"cited_title":"A DNN -based WiFi-RSSI indoor localization method in IoT","cited_arxiv_id":null,"evidence_quote":"DNNLOC, the four-hidden-layer DNN baseline whose error, size, and latency LogNet is compared against."},{"cited_title":"SANGRIA: Stacked Autoencoder Neural Networks with Gradient Boosting for Indoor Localization,","cited_arxiv_id":null,"evidence_quote":"SANGRIA, a three-hidden-layer stacked-autoencoder baseline in the hidden-depth comparison."},{"cited_title":"AutLoc: Deep Autoencoder for Indoor Localization with RSS Fingerprinting,","cited_arxiv_id":null,"evidence_quote":"AUTLOC, a two-hidden-layer autoencoder baseline used in the depth and efficiency comparison."},{"cited_title":"eXLoc: Understanding Deep Learning -driven Indoor Localization with eXplainable AI","cited_arxiv_id":null,"evidence_quote":"eXLoc, the explainable-localization work that motivates the need for interpreting DNN decisions in fingerprinting."},{"cited_title":"RADAR: An in -building RF -based user location and tracking system","cited_arxiv_id":null,"evidence_quote":"RADAR, the classic Wi-Fi fingerprinting method whose assumptions LogNet builds on conceptually."},{"cited_title":"https://github.com/EPIC-CSU/heterogeneous -rssi- indoor- nav","cited_arxiv_id":null,"evidence_quote":"The open-sourced two-building, two-year RSS dataset on which all models are trained and evaluated."},{"cited_title":"STELLAR: Siamese multiheaded attention neural networks for overcoming temporal variations and device heterogeneity with indoor localization","cited_arxiv_id":null,"evidence_quote":"STELLAR, which documents temporal variation and device heterogeneity in indoor localization, framing the long-term robustness problem."}],"review_version":1}