{"id":"987f0db4-192b-48c5-910b-b73aec54d6ec","arxiv_id":"2608.02140","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"HiResNets use foveated log-polar warps directly inside a residual stream, giving convolution cost that grows logarithmically with resolution and better accuracy on egocentric video benchmarks.","lead":"This paper introduces HiResNets, a video recognition architecture that applies log-polar warps inside every residual block, so convolutions run on a small foveated view while a separate buffer keeps the whole frame. The goal is to remove the quadratic memory and compute cost of full-HD video, and the paper reports gains on egocentric video tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sec. 6.3's memory-free construction omits the center predictor: with x never instantiated, Eq. 6 has no input, so the claimed quadratic-elimination does not hold for the adaptive method.","rationale":"The strongest claim, as identified by the reader, is the log-square scaling plus the theoretical construction eliminating the full-resolution residual stream. The convolution-block arithmetic leading to O(k² log W log H) is sound for the warped view, and the linearity identity in Sec. 6.3 is valid if the centers are fixed. The empirical section, however, is under-supported (no error bars, unmatched baselines, missing appendix for the 5% claim), which independently justifies the reader's conditional verdict. My stress-test focuses on an internal gap that is more fundamental than the empirical weaknesses: the adaptive center predictor is an essential part of the method, and it is defined on the very tensor the Sec. 6.3 construction removes. Without a specification of how to predict c_l from the stored log-polar tensors, the quadratic-elimination theorem applies to a different, non-adaptive algorithm. This does not invalidate the practical idea of keeping a full-res residual stream with log-polar blocks, nor the log-square scaling of convolutions; those remain plausible. But it means the paper's strongest theoretical claim (sub-quadratic memory via eliminating x) is not established for the proposed method. Since the reader already judged the paper CONDITIONAL based on missing evidence and loose claims, my concern reinforces that verdict rather than moving it to reject: the core architecture may still work, and the Sec. 6.3 construction could be repaired by specifying an attention computation that avoids x. Thus verdict_should_be UNCHANGED (CONDITIONAL).","tokens_in":12487,"tokens_out":6868,"duration_ms":69238,"concrete_test":"Write the full forward graph of the Sec. 6.3 memory-free variant including the center predictor: substitute x_l=Σ_{k<l}ψ^{-1}_{c_k}(y_k) into the attention map a(x_l) used in Eq. 6 and derive the cost of evaluating c_l. If evaluating softargmax over a∈R^{H×W} requires materializing x_l or costs Ω(HW) per block, the quadratic-elimination claim fails for adaptive centers. Alternatively, run the memory-free training loop with centers predicted from log-polar tensors only and compare accuracy to the full-res implementation; any degradation or inability to train confirms the missing dependency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is that the log-polar update can be rearranged as u_l = Σ_{k<l} W_{k→l}(y_k) with W_{k→l}=ψ_{c_l}∘ψ^{-1}_{c_k}, so the full-resolution residual stream x can be eliminated (Sec. 6.3). This is arithmetically correct for fixed centers, but it silently drops the center predictor. Sec. 3.4 defines the focus point via softargmax over an attention map a∈R^{H×W} derived from the current residual state x (Eq. 6). The whole point of c is that it is adaptive per block/stage. If x is never instantiated, the attention map cannot be formed from x as specified. The paper gives no alternative way to compute c_l from the stored log-polar tensors {y_k}. Thus the only version of HiResNets that realizes adaptive foveation is the one that keeps the O(HW) residual stream, whose memory is quadratic and whose per-block warp/unwarp/copy-add cost is linear in resolution. The paper acknowledges this in Fig. 6's caption ('linearity dominates'), so the asymptotic win is limited to the convolution blocks, not the pipeline. The theoretical construction in Sec. 6.3 therefore does not establish the headline 'eliminates the quadratic dependency of the residual stream resolution' for the actual adaptive method; at best it is a memory optimization for fixed, precomputed centers, which is not the proposed method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HiResNets, a residual architecture that applies convolutional blocks in a log-polar warped view around a dynamically predicted focus point, while a full-resolution residual stream is preserved and updated by inverse warps. The authors derive logarithmic-square scaling for the convolutional blocks, present a theoretical construction intended to eliminate the full-resolution residual stream by composing log-polar warps, and report experiments on egocentric gaze estimation, object detection, and fine-grained part detection. The central idea is to embed foveation inside the backbone rather than wrapping a standard network with external glimpse or zoom modules.","tokens_in":12718,"tokens_out":6073,"duration_ms":85123,"significance":"If the log-squared scaling for the convolutional blocks and the warp-composition construction could be made to cover the full adaptive method, the paper would be a useful step toward efficient high-resolution video recognition. The core conceptual contribution is clear: treating the residual stream as a persistent high-resolution buffer that is only read and written through warped views is an elegant way to localize the resolution bottleneck. The complexity derivation is self-contained and does not rely on fitted parameters, which is a strength. However, the paper's empirical support is thin, the appendix that is referenced for key results is absent, and the theoretical construction as written does not apply to the adaptive center predictor. The value of the paper currently lies more in the architecture concept than in the demonstrated results.","major_comments":[{"comment":"The memory-free construction in Sec. 6.3 eliminates the residual stream only when the centers c_k are fixed or precomputed. In the proposed method, the center predictor receives the current residual state x (Sec. 3.4) and computes a full-resolution attention map a in R^{H×W}; with x never instantiated, the adaptive centers cannot be computed as specified. The paper should either show how c_l is obtained from the stored log-polar tensors {y_k}, or explicitly restrict the construction to fixed centers, which would not represent the proposed adaptive method.","section":"Sec. 6.3, Eq. (6)"},{"comment":"The claim that the residual stream 'maintains full fidelity' and that 'no fine detail is lost' is not supported. Every update to x is the inverse warp of a log-polar tensor whose peripheral sampling is coarser than the foveal region, and bilinear interpolation cannot reconstruct high-frequency peripheral content that was not sampled at warp time. The paper should provide an error analysis of the warp/unwarp composition, such as a bound on the reconstruction error in the periphery, or soften the fidelity claim to note that the buffer is full-resolution but the updates are lossy outside the fovea.","section":"Sec. 3.6 and Abstract"},{"comment":"The complexity analysis counts only the convolution in the warped view and omits the O(HW) center predictor, which runs two 1x1 convolutions over a full-resolution attention map, and the O(HW) residual-stream update. Figure 6's caption already concedes that 'linearity dominates,' but the paper still attributes nearly-linear scaling to the whole model. The full pipeline complexity should be reported, and Fig. 5b should separately measure the center predictor, warp, and copy/add contributions to latency.","section":"Sec. 3.6 and Sec. 5.3"},{"comment":"The claimed 5 percent accuracy improvement over a ResNet baseline in image classification is not backed by any experiment in the manuscript, and the appendix that is said to contain it is absent. Additionally, Tables 1-3 report no error bars, number of seeds, or training details, and the comparison baselines are not matched in backbone or compute. These omissions make the empirical claim of 'superior performance' hard to verify; add matched-backbone ablations with multiple seeds and error bars.","section":"Sec. 5.2 and Table 3"}],"minor_comments":[{"comment":"There is a typo: 'For the the Ego4D dataset' contains a duplicated article.","section":"Sec. 5.1"},{"comment":"The column header 'Max Res.' is not defined; clarify whether it refers to training resolution, evaluation resolution, or both.","section":"Table 1"},{"comment":"The text says latency is measured from 224px to 1K, but Fig. 5b lists resolutions 700, 1.3K, 1.7K, and 2K; the two ranges should be made consistent.","section":"Sec. 5.3 and Fig. 5b"},{"comment":"The inverse warp is defined in continuous coordinates but the output grid size is not specified; state explicitly that it maps to the full H by W grid.","section":"Sec. 3.2, Eq. (7)"},{"comment":"The abstract and title claim 'native Full-HD' processing, but the largest input resolution reported in Tables 1-3 is 1400 pixels; no experiment at 1920 by 1080 is shown.","section":"Abstract and Title"}],"recommendation":"major_revision","confidential_remarks":"The conditional verdict from the first reader is reasonable. The Sec. 6.3 omission is decisive for the headline claim about eliminating the quadratic residual stream, and the missing appendix undermines the image-classification claim. If the authors cannot address these points, the claims should be substantially narrowed. I would not reject outright because the core architecture idea is clear and the convolutional complexity derivation is a useful contribution, but the manuscript needs a major revision before it can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"HiResNets is a genuinely interesting architecture idea, but the paper's headline claims outrun the evidence. The log-squared convolution complexity is real; the claim that you can eliminate the quadratic residual stream without breaking adaptive foveation is not, as written.\n\nWhat's new: embedding log-polar warps directly into each residual block, with the residual stream acting as a persistent high-resolution buffer, is a fresh twist on foveated vision. The direct log-polar-to-log-polar warp identity in Sec. 6.3 (W_{k→l} = ψ_{c_l}∘ψ^{-1}_{c_k}) is a neat observation that could be useful beyond this paper. The complexity calculation for the convolution blocks—O(k^2 log W log H)—is straightforward and correct. The idea that convolutions dominate the cost and can be made sub-quadratic while the stream remains full-res is sensible, and the reported gains on egocentric detection and gaze are plausible.\n\nBut there are real soft spots. The empirical support is thin: no error bars anywhere, baselines are not clearly matched for compute/parameters, and the claimed 5% image-classification improvement points to an appendix that isn't in the manuscript. The 'full fidelity' of the residual stream is overstated—bilinear log-polar resampling is lossy in the periphery, and the paper gives no error analysis for composing many warps.\n\nMore seriously, the theoretical construction in Sec. 6.3 doesn't hold up for the adaptive method. The center predictor (Eq. 6) needs an attention map derived from the full-resolution residual stream x. If x is never instantiated, there is no input to that predictor. The memory-free formulation only works for fixed, precomputed centers, which is not the proposed method. The paper's own Fig. 6 caption admits that log-square scaling applies to conv only and that other operators (warp, unwarp, copy/add) are still linear, so the asymptotic win is more limited than the abstract suggests.\n\nThis is a solid workshop-caliber idea that needs more work before it supports the strong claims. A good referee could push the authors to fix the Sec. 6.3 gap, supply the missing experiments and error bars, and clarify what exactly is matched in the baselines. I'd send it to review rather than desk-reject, but I would not bet on the current version.","headline":"Interesting architecture with a real log-squared conv scaling, but the paper overclaims the memory-free construction and under-delivers on empirical support.","tokens_in":13340,"tokens_out":3559,"would_cite":false,"duration_ms":34738,"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":"HiResNets embed foveation directly into residual backbones, making convolutional block cost grow logarithmically-squared rather than quadratically with resolution, so full-HD video can be processed natively.","keywords":["foveated vision","log-polar warp","residual networks","video recognition","egocentric vision","high-resolution video","efficient inference"],"falsifier":"Measure the reconstruction fidelity of composing the inverse log-polar warp from one focus center with the forward warp from a different center: if $\\|\\psi_{c_2}(\\psi^{-1}_{c_1}(y)) - y\\|$ is large for high-frequency peripheral content, the accumulated residual stream has already lost the fine detail the method claims to preserve. Concretely, run the model on a test frame containing small high-contrast targets far from the focus point and compare detection accuracy against the same model with the focus forced onto those targets; a large accuracy drop when targets stay peripheral would indicate the peripheral-sampling assumption fails.","tokens_in":12197,"feed_emoji":"👁️","tokens_out":6430,"duration_ms":51513,"temperature":0.7,"pith_summary":"This paper tries to show that the quadratic cost of high-resolution video recognition is not inevitable: a residual network whose convolutional blocks interact with a full-resolution buffer only through log-polar warps can keep block compute logarithmic-squared in resolution. The payoff, if true, is that full-HD and higher video can be processed natively, with fine detail preserved in the residual stream while convolutions run on small foveated views. The paper reports that this design scales nearly linearly in latency, learns human-like fixation points, and beats standard baselines on egocentric gaze, object detection, and fine-grained part recognition. A theoretical construction is also given that removes the quadratic memory of the residual stream itself by composing log-polar warps directly.","feed_headline":"Foveated residual nets process full-HD video at log-squared cost","feed_subtitle":"Convolutions run on small warped views while a high-res buffer keeps full fidelity, shifting scaling from quadratic to log-squared.","key_machinery":"The central object is the log-polar warp pair $\\psi_c$ and $\\psi^{-1}_c$ (Eqs. 3-4 and 7), which maps a Cartesian feature map to a warped view with $\\Theta = O(\\log W)$ angular and $P = O(\\log H)$ radial samples around a predicted focus center $c$. A residual block computes $u = \\psi_c(x)$, $y = f(u)$, and $x \\leftarrow x + \\psi^{-1}_c(y)$, so convolutions run only on the compact log-polar tensor while the full-resolution residual stream acts as a persistent buffer. The paper's complexity argument shows the convolutional cost drops to $O(k^2 \\log W \\log H)$, and its theoretical construction composes log-polar-to-log-polar warps $W_{k\\to l} = \\psi_{c_l} \\circ \\psi^{-1}_{c_k}$ so the stream $x$ is never instantiated.","core_discovery":"The paper claims that residual networks can process full-HD video natively by treating the residual stream as a high-resolution buffer that convolutional blocks read and write only through log-polar warps. Each block's convolutions operate on a small warped view centered on a predicted focus point, so block cost scales as $O(\\log W \\log H)$ instead of $O(WH)$, and repeated blocks integrate multiple focused updates into the buffer, like saccades building a complete percept. A theoretical construction eliminates the quadratic residual-stream memory by rewriting the accumulated update as sums of direct log-polar-to-log-polar warps $W_{k\\to l} = \\psi_{c_l} \\circ \\psi^{-1}_{c_k}$, storing only log-polar tensors. Experiments on egocentric benchmarks report accuracy gains over baselines, especially for small objects and fine-grained part detection, with latency growing nearly linearly rather than quadratically as input resolution increases.","pith_inferences":["Editorial inference: if the warp-composition construction holds as stated, the same design should extend beyond full HD to 4K and 8K inputs, where the memory savings are larger; the paper does not report such an extrapolation.","Editorial inference: the center predictor is effectively a differentiable spatial-attention mechanism, so HiResNets could plausibly be combined with token-pruning or sparse-attention transformers to extend log-squared savings beyond convolutional components.","Editorial inference: a direct test of the peripheral-sampling assumption would force the focus point away from a small target and measure the accuracy drop; the paper's 'no prediction' ablation already hints that focus location matters, but it does not isolate peripheral reconstruction error.","Editorial inference: the theoretical construction shifts cost from resolution to depth, so very deep networks may accumulate warp-composition error; quantifying that drift would show whether log-squared scaling remains practical at scale."],"forward_implications":["Residual blocks in HiResNets scale as $O(k^2 \\log W \\log H)$, so increasing input resolution costs far less than in standard convnets or vision transformers.","The residual stream preserves full fidelity, and the theoretical construction removes its quadratic memory by storing only log-polar tensors and composing warps directly, at the price of a quadratic dependency on depth.","On egocentric benchmarks (Ego4D, EgoObjects, PACO, EGTEA), the method improves small-object and fine-grained accuracies over YOLOv11, GLC, and SqueezeTime baselines at matched resolutions.","Latency grows nearly linearly with input resolution up to 1K, whereas comparable baselines grow quadratically.","The learned foveation is interpretable: predicted focus points align with gaze trajectories and alternate between near and far regions in panoramic scenes."],"supporting_citations":[{"why":"Supplies the residual-block formulation and the ResNet baseline that HiResNets modifies.","marker":"[17]"},{"why":"Provides the log-polar reparameterization that defines the warp used to compress the periphery.","marker":"[12]"},{"why":"Supplies the differentiable spatial transformer mechanism behind the center predictor and warp modules.","marker":"[19]"},{"why":"Provides the warped-convolution machinery used to motivate input-dependent geometric transforms of feature maps, including closed-form grids.","marker":"[18]"},{"why":"The strong one-stage detection baseline that HiResNets is compared against on Ego4D, EgoObjects, and PACO.","marker":"[20]"},{"why":"Provides the PACO object-part dataset used to test fine-grained and small-object recognition.","marker":"[39]"},{"why":"Provides the Ego4D dataset and gaze trajectories used in egocentric evaluation and saccade analysis.","marker":"[16]"},{"why":"Supplies the SqueezeTime baseline for latency scaling and egocentric gaze comparison.","marker":"[51]"}],"fun_headline_variants":["HiResNets foveate video, cutting cost from quadratic to log-squared","Foveal residual streams enable native full-HD video recognition","Log-polar warps let residual nets see full-HD video cheaply","Saccade-like video nets process full-HD at log-squared cost","Foveated residual nets: full-HD video without quadratic blowup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the bilinear log-polar warp and its inverse are accurate enough that, after many warped updates are accumulated into the residual stream, no task-relevant fine detail is lost; the finite log-polar grid samples the periphery far more coarsely than the fovea, and the paper does not analyze the error of composing the forward and inverse warps.","fun_headline_variants_meta":{"raw":{"variants":["HiResNets foveate video, cutting cost from quadratic to log-squared","Foveal residual streams enable native full-HD video recognition","Log-polar warps let residual nets see full-HD video cheaply","Saccade-like video nets process full-HD at log-squared cost","Foveated residual nets: full-HD video without quadratic blowup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1401,"prompt_tokens":946,"completion_tokens":455,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":359}},"tokens_in":562,"tokens_out":455,"duration_ms":5133,"temperature":1.0,"reasoning_tokens":359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:09:46.977187+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the reconstruction fidelity of composing the inverse log-polar warp from one focus center with the forward warp from a different center: if $\\|\\psi_{c_2}(\\psi^{-1}_{c_1}(y)) - y\\|$ is large for high-frequency peripheral content, the accumulated residual stream has already lost the fine detail the method claims to preserve. Concretely, run the model on a test frame containing small high-contrast targets far from the focus point and compare detection accuracy against the same model with the focus forced onto those targets; a large accuracy drop when targets stay peripheral would indicate the peripheral-sampling assumption fails.","supporting_citations":[{"cited_title":"Deep residual learning for image recognition","cited_arxiv_id":null,"evidence_quote":"Supplies the residual-block formulation and the ResNet baseline that HiResNets modifies."},{"cited_title":"Polar transformer networks","cited_arxiv_id":null,"evidence_quote":"Provides the log-polar reparameterization that defines the warp used to compress the periphery."},{"cited_title":"Warped convolutions: Efficient invariance to spatial transformations","cited_arxiv_id":null,"evidence_quote":"Provides the warped-convolution machinery used to motivate input-dependent geometric transforms of feature maps, including closed-form grids."},{"cited_title":"Ultralytics yolo11, 2024","cited_arxiv_id":null,"evidence_quote":"The strong one-stage detection baseline that HiResNets is compared against on Ego4D, EgoObjects, and PACO."}],"review_version":2}