{"id":"4668ad32-8abd-484c-ad41-015914f32066","arxiv_id":"1909.05622","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"An Inception-inspired LSTM with 1x1, 3x3, and 5x5 kernels inside each gate yields modestly lower MAE and MSE than a convolutional LSTM for next-frame video prediction on KITTI and KTH.","lead":"This paper replaces the standard convolution in a convolutional LSTM with an Inception-style multi-kernel block for next-frame video prediction. It reports small accuracy gains over a convolutional LSTM baseline on KITTI and KTH, but the comparison is confounded by a large parameter-count gap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Parameter-count confound dominates: Inception LSTM's gains over ConvLSTM could be capacity, not multi-kernel design; without a matched-width baseline and significance testing the central claim is unsupported.","rationale":"The reader's weakest-assumption identification is exactly the load-bearing concern: the paper does not separate architectural benefit from increased parameter count. The central claim requires that multi-kernel design causes the improvement, and the only reported comparison is against a much smaller ConvLSTM. The paper does provide public code, which is genuine and useful for testing, but it does not include the capacity-matched control. Additional weaknesses--overlapping confidence intervals, no significance tests, and inconsistency in the v1-vs-v2 comparison--reinforce the same conclusion: the evidence is suggestive but not conclusive. Since the concern is about missing controls rather than an internal contradiction or a likely false result, conditional acceptance remains the appropriate verdict. The proposed concrete test would settle the issue in one controlled run using the existing repository.","tokens_in":7174,"tokens_out":2295,"duration_ms":23315,"concrete_test":"Using the public code, train a ConvLSTM baseline inside PredNet whose per-layer channel count is widened so its parameter count matches Inception LSTM v1 (about 6,595 parameters per layer vs 1,081), on both KITTI and KTH, with the same training budget, optimization settings, and evaluation protocol. Also compute per-video paired differences with 95% confidence intervals. If the matched-capacity ConvLSTM matches or exceeds Inception LSTM on MAE/MSE/SSIM, the central claim fails; if Inception still wins with non-overlapping intervals, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that replacing the single convolutional kernel in each LSTM gate with multiple kernels of sizes 1, 3, and 5 improves next-frame prediction inside PredNet. For that claim to be supported, the observed performance gap must be attributable to the multi-kernel architecture. The paper reports that Inception LSTM v1 uses 6,595 parameters per layer versus 1,081 for ConvLSTM (Section 3), roughly a sixfold increase, yet no experiment controls for capacity. A wider ConvLSTM with matched parameter count might achieve the same or better MAE/MSE/SSIM; if so, the Inception module's multi-scale receptive fields are not the operative cause. Additional support for this concern is that the only confidence-interval plot (Figure 4/6) shows overlapping intervals for MSE, and Table 1 presents single point estimates without error bars or significance tests. The abstract's secondary claim that Inception v1 beats v2 is also not consistently supported: on KTH, Inception v2 at 4 layers has lower MSE (0.010637) than Inception v1 at 4 layers (0.010959). Thus the load-bearing assumption--that the architecture, not capacity, drives the improvement--is not secured by the reported evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Inception-inspired convolutional LSTM for next-frame video prediction inside the PredNet predictive-coding framework. Two variants are defined: Inception v1 uses parallel 1x1, 3x3, and 5x5 convolutions in each LSTM gate, while Inception v2 replaces the 5x5 convolution with two stacked 3x3 convolutions. The method is compared with the standard convolutional LSTM on the KITTI and KTH data sets at two, three, and four layers, using MAE, MSE, and SSIM, with additional plots of metrics as a function of the number of previous frames. The authors report that the Inception-based LSTM outperforms the convolutional LSTM, that Inception v1 outperforms v2, and that v2 has lower computational cost. The source code is publicly released.","tokens_in":7295,"tokens_out":7758,"duration_ms":71969,"significance":"Taken at face value, the result would provide evidence that multi-scale convolutional kernels inside LSTM gates improve next-frame prediction, which is a modest architectural variation on PredNet with potential practical value. Strengths of the paper include the public code release, the use of standard datasets and metrics, and evaluation on held-out test data; there is no circularity in the sense of fitting the model on the test set. However, the central comparison is not yet established: the capacity difference between the Inception LSTM and the convolutional LSTM is unaddressed, the only confidence-interval evidence explicitly shows overlap, and the reported tables contain internal inconsistencies. The contribution is potentially useful but needs substantially stronger empirical support.","major_comments":[{"comment":"The central architectural claim is confounded by a large difference in model capacity: the parameter-count paragraph reports 6,595 parameters per layer for Inception LSTM version 1 versus 1,081 parameters for the convolutional LSTM, roughly a sixfold increase. The experiments do not include a wider ConvLSTM baseline with a matched parameter count, so the reported improvements over ConvLSTM cannot be attributed to the multi-kernel design as opposed to additional capacity. Please add a capacity-matched baseline, for example a wider ConvLSTM with comparable parameter count, before claiming that the Inception module is responsible for the gain.","section":"Section 3, parameter-count paragraph"},{"comment":"The only quantified uncertainty in the paper is the 95% confidence interval plot, and the accompanying text states that although the MSE of the Inception LSTM is lower than that of the convolutional LSTM, the confidence intervals overlap. This directly weakens the abstract's claim of outperformance. Moreover, the text identifies this figure as KITTI while the caption labels it KTH, so the reader cannot tell which dataset the statement refers to. Table 1 provides no error bars or significance tests; please report them and reconcile the figure label with the text.","section":"Section 4, Figure 6"},{"comment":"Table 1 contradicts the abstract and contains a likely data error. On KTH, Inception v2 (4L) has MSE 0.000463, which is lower than Inception v1 (4L) at 0.000524, contradicting the claim that Inception LSTM has better prediction performance than Inception v2 LSTM and the statement that Inception Version I shows the best performance. In addition, the KTH ConvLSTM (2L) row (MAE 0.044115, MSE 0.007191, SSIM 0.867645) is identical to the KITTI Inception v2 (4L) row, which suggests a copying or bookkeeping error and undermines confidence in the reported numbers as a whole. Please correct the table and rerun or verify the affected experiments.","section":"Table 1"},{"comment":"The paper does not specify the number of training and test sequences used for the KITTI and KTH evaluations, the data splits, the number of runs, or the random seeds. Many of the metric differences in Table 1 are in the fourth decimal place, so run-to-run variance could alter several of the conclusions. Please provide the missing protocol details and repeat measurements so that the ranking of models can be assessed.","section":"Section 4, experimental setup"}],"minor_comments":[{"comment":"The abstract uses \"unsupervised\" while the full-text abstract says \"self-supervised\"; please use one term consistently.","section":"Abstract"},{"comment":"The notation in Eq. (3a) is inconsistent with the stated design: a two-3x3 replacement of the 5x5 kernel should apply the second 3x3 convolution to the output of the first, but the equation introduces three 3x3 weight tensors (W1, W2, W3); please correct the equation.","section":"Section 3, Eq. (3a)"},{"comment":"The claim that using Inception removes the need to choose a kernel size is overstated, since the method still fixes the kernel set {1x1, 3x3, 5x5}.","section":"Section 4"},{"comment":"Reference [9] is listed as a \"git book\" with no author or title; it should be replaced by a citable source or removed.","section":"Related work, reference [9]"},{"comment":"The parameter count (1+9+25)*nc appears to count weights for only one side of the convolution, ignoring that both the input and the recurrent hidden state contribute weight tensors; please show the full count explicitly.","section":"Section 3, parameter-count derivation"}],"recommendation":"major_revision","confidential_remarks":"Given the duplicate row in Table 1 and the conflict between Figure 6's caption and the text, I would ask the authors for the raw evaluation logs before any further review cycle. The editor may also wish to consider whether the incremental nature of the contribution and the current level of empirical support fit the journal's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The Inception-LSTM idea is reasonable and the code is public, but the evaluation as written doesn't pin the improvement to the multi-kernel design. The central comparison gives the Inception LSTM about six times more parameters per layer than the convLSTM baseline (6,595 vs 1,081), so the gain could just be capacity. I'd want a wider ConvLSTM with matched parameters before believing the architectural claim.\n\nWhat's actually new: replacing each gate's single convolution with a stack of 1x1, 3x3, and 5x5 convolutions (v1) or two stacked 3x3s (v2) inside the PredNet framework. That's a straightforward but legitimate extension of both convLSTM and Inception; I don't know prior work doing exactly this. The paper is clearly written, the equations are correct, and the code is on GitHub, which supports reproducibility.\n\nWhere it's soft: no parameter-matched baseline, no significance tests, and Table 1 has no error bars. The one confidence-interval plot (KITTI MSE) shows overlap between Inception and convLSTM, which undercuts the main claim. Also, the abstract says v1 beats v2, but on KTH with four layers v2 has lower MSE (0.010637) than v1 (0.010959), so that claim doesn't hold across the table. These are moderate flaws; they don't make the paper useless, but they mean the current contribution is essentially “here's a variant that seems to help, and we don't know why.”\n\nWho it's for: researchers working on video prediction or predictive coding who want a quick architectural alternative to convLSTM. It's a workshop-level contribution; with a capacity-controlled baseline and proper error analysis it could become a solid short paper. I'd send it to peer review, but with a strong suggestion to the authors to address the confound.","headline":"A plausible incremental architecture with a load-bearing parameter-count confound; the gains are real but not yet attributable to the multi-kernel design.","tokens_in":7904,"tokens_out":1997,"would_cite":false,"duration_ms":20565,"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":"An Inception-inspired LSTM, replacing each convolutional LSTM gate with parallel 1x1, 3x3, and 5x5 convolutions, predicts next video frames more accurately than a standard convolutional LSTM.","keywords":["Inception LSTM","convolutional LSTM","predictive coding","next-frame video prediction","KITTI dataset","KTH dataset","multi-scale kernels","self-supervised learning"],"falsifier":"Train a convolutional LSTM matched in per-layer parameter count to the Inception LSTM's 6,595, for instance by widening its feature channels in the reported 3-48-96-192 configuration, and compare MSE and SSIM on the same KITTI and KTH splits under the same predictive-coding protocol. If the wider ConvLSTM matches or exceeds the Inception LSTM's scores, the attribution of the gain to multi-scale kernels is not supported.","tokens_in":6887,"feed_emoji":"🎬","tokens_out":13693,"duration_ms":114761,"temperature":0.7,"pith_summary":"The paper proposes a self-supervised Inception-inspired convolutional LSTM in which each gate and the candidate cell update run three parallel convolutions of different kernel sizes ($1\\times1$, $3\\times3$, $5\\times5$; a second version replaces the $5\\times5$ branch with two stacked $3\\times3$ convolutions) and then concatenate the results before the gate nonlinearity. Inside the paper's predictive-coding architecture, this multi-scale gate reduces next-frame prediction error and raises structural similarity on both KITTI traffic footage and KTH walking videos when compared with the standard single-kernel convolutional LSTM. Version 1 predicts best overall, while version 2 cuts parameters by $7n_c$ per gate per layer and has lower computational cost. The broader claim is that the fixed kernel size of a convolutional LSTM is a hyperparameter that can be replaced by a small set of parallel scales without deepening the network.","feed_headline":"Multi-kernel LSTM beats single-kernel ConvLSTM at video prediction","feed_subtitle":"Parallel 1x1, 3x3, and 5x5 gate filters cut next-frame error on KITTI and KTH.","key_machinery":"The key object is the Inception-inspired LSTM cell: for each gate (input $i_t$, forget $f_t$, output $o_t$) and for the candidate update $g_t$, the pair $[x_t,h_{t-1}]$ is convolved in parallel by kernels of sizes $1\\times1$, $3\\times3$, and $5\\times5$, and the three output stacks are concatenated and passed through the gate's activation. In version 2 the $5\\times5$ convolution is replaced by two sequential $3\\times3$ convolutions, matching the Inception-v2 factoring idea. The rest of the LSTM update ($c_t = f_t \\odot c_{t-1} + i_t \\odot g_t$, $h_t = o_t \\odot \\tanh(c_t)$) is unchanged, so the mechanism is a drop-in change to the gate internals that widens each gate's receptive-field range.","core_discovery":"The central empirical claim is that an Inception-style replacement for the convolutional LSTM's gates yields better next-frame predictions. With four layers and the same predictive-coding protocol, Inception version 1 reports KITTI MSE of 0.0436 and SSIM of 0.868, against 0.0458 and 0.858 for the convolutional LSTM; on KTH the best version-1 MSE is 0.0103 with SSIM 0.964. Version 2 stays close to version 1 in accuracy while using fewer parameters because its $5\\times5$ path is factored into two $3\\times3$ convolutions. The paper interprets these results as evidence that a richer set of receptive-field scales, not a deeper network, improves spatiotemporal prediction.","pith_inferences":["Because the Inception LSTM's per-layer parameter count (6,595) is roughly six times the ConvLSTM's (1,081) and no wider ConvLSTM baseline was run, the cleanest way to see whether the multi-scale kernels themselves matter is a parameter-matched comparison; the reported gap could partly be a capacity effect.","A natural extension the paper does not pursue is to apply the same gate-level multi-scale idea to the recurrent weight path $h_{t-1}$ alone, or to a GRU cell, to see whether the benefit transfers beyond the convolutional-LSTM setting studied here.","The version-1 versus version-2 comparison suggests that the effective receptive-field size contributes more than raw parameter count on these datasets; systematically varying the kernel set (e.g., $3\\times3$, $5\\times5$, $7\\times7$) would map that trade-off directly."],"forward_implications":["Kernel size no longer needs to be selected as a single hyperparameter for the recurrent gates; the Inception LSTM learns to combine $1\\times1$, $3\\times3$, and $5\\times5$ views of the same input.","Under the tested protocol, the Inception LSTM gives the best MSE and SSIM on both datasets among the three compared models, and version 1 is the overall best predictor.","Version 2 offers a parameter and compute reduction of $7n_c$ per gate per layer relative to version 1 while retaining most of version 1's accuracy.","Prediction quality in these models appears to plateau after about five frames of history, so the gate redesign does not shift the required memory horizon.","Because the surrounding predictive-coding architecture and error-propagation rules are unchanged, the Inception gate can replace a convolutional LSTM without re-engineering the rest of the model."],"supporting_citations":[{"why":"Supplies the predictive-coding recurrent architecture and the convolutional-LSTM baseline whose gates the paper replaces.","marker":"[11]"},{"why":"Provides the convolutional LSTM update equations and the observation that different kernel sizes capture different motion magnitudes.","marker":"[22]"},{"why":"Introduces the Inception design of parallel kernels of different sizes that the paper adapts inside each LSTM gate.","marker":"[18]"},{"why":"Supplies the Inception-v2 factorization idea of replacing a 5x5 convolution with two stacked 3x3 convolutions, used for the paper's version 2.","marker":"[19]"},{"why":"Provides the KITTI driving dataset used for the first evaluation.","marker":"[4]"},{"why":"Provides the KTH human-action walking dataset used for the second evaluation.","marker":"[15]"}],"fun_headline_variants":["Inception LSTM outperforms ConvLSTM on video prediction","Wider LSTM gates beat deeper networks for video frames","Multi-scale kernels in LSTM improve next-frame accuracy","Inception v1 LSTM beats v2 and ConvLSTM on KITTI, KTH","Parallel 1x1, 3x3, 5x5 filters sharpen LSTM video prediction"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper attributes the performance gain to the multi-kernel design, even though the Inception LSTM has roughly six times more weights per layer than the ConvLSTM it is compared against and no equal-capacity ConvLSTM control is tested.","fun_headline_variants_meta":{"raw":{"variants":["Inception LSTM outperforms ConvLSTM on video prediction","Wider LSTM gates beat deeper networks for video frames","Multi-scale kernels in LSTM improve next-frame accuracy","Inception v1 LSTM beats v2 and ConvLSTM on KITTI, KTH","Parallel 1x1, 3x3, 5x5 filters sharpen LSTM video prediction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000609,"raw_usage":{"total_tokens":2805,"prompt_tokens":882,"completion_tokens":1923,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":1823}},"tokens_in":498,"tokens_out":1923,"duration_ms":13342,"temperature":1.0,"reasoning_tokens":1823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:39:31.124613+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a convolutional LSTM matched in per-layer parameter count to the Inception LSTM's 6,595, for instance by widening its feature channels in the reported 3-48-96-192 configuration, and compare MSE and SSIM on the same KITTI and KTH splits under the same predictive-coding protocol. If the wider ConvLSTM matches or exceeds the Inception LSTM's scores, the attribution of the gain to multi-scale kernels is not supported.","supporting_citations":[{"cited_title":"Vision meets robotics: The kitti dataset","cited_arxiv_id":null,"evidence_quote":"Provides the KITTI driving dataset used for the first evaluation."},{"cited_title":"Recog- nizing human actions: a local SVM approach","cited_arxiv_id":null,"evidence_quote":"Provides the KTH human-action walking dataset used for the second evaluation."},{"cited_title":"Convolutional LSTM network: A machine learning approach for pre- cipitation nowcasting","cited_arxiv_id":null,"evidence_quote":"Provides the convolutional LSTM update equations and the observation that different kernel sizes capture different motion magnitudes."},{"cited_title":"Going deeper with convolutions","cited_arxiv_id":null,"evidence_quote":"Introduces the Inception design of parallel kernels of different sizes that the paper adapts inside each LSTM gate."},{"cited_title":"Rethinking the inception architec- ture for computer vision","cited_arxiv_id":null,"evidence_quote":"Supplies the Inception-v2 factorization idea of replacing a 5x5 convolution with two stacked 3x3 convolutions, used for the paper's version 2."}],"review_version":1}