{"id":"f4d780e3-4bbf-4f1b-a05a-71db44429ddd","arxiv_id":"2501.09531","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tiny-memory CNN generates part of its weights from a cellular automaton and uses multiplexer-based quantized residual links, reporting up to 1% higher accuracy than two compressed baselines at sub-2Mb sizes.","lead":"MOGNET is a compact neural network that shrinks memory by generating some of its weights on the fly with a cellular automaton instead of storing them, and by using low-precision numbers and a multiplexer-based shortcut. In tests on CIFAR image datasets, it reports up to about one percent higher accuracy than two prior compression methods at similar or smaller model sizes under 2 megabits.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CA-generated pointwise weights are the linchpin of the memory savings, yet the paper provides no ablation, no seed-sensitivity analysis, and no multi-seed variance; until that evidence exists, the reported sub-2Mb accuracy edge is not established.","rationale":"The reader's weakest_assumption identifies the CA-generated pointwise weights, and I agree that this is the correct load-bearing concern. The memory comparison in Table II counts those weights as zero-cost, and the paper's novelty and headline claim rest on that. I do not see an internal contradiction: the architecture is described precisely enough to reproduce, and the reported numbers are consistent with the selected configuration. The weakness is the missing evidence around the fixed projection. A CA-seed sweep and a learned-ternary replacement would directly test whether the reported accuracy is a genuine property of the design or an accident of one pseudo-random realization. The paper also lacks multi-seed training variance, which matters because the claimed gap is under 1%. I also note that Eq. (4) uses a '32' factor that is not clearly justified by the stated 1-bit/2-bit weight precisions, although the Table II sizes are empirical and would not be affected by this notation issue. These gaps are exactly why the paper should remain conditional rather than be accepted outright. My read does not change the reader's verdict; it reinforces it. The concrete ablation proposed above would settle the concern: if accuracy is stable across CA seeds and not substantially worse than a learned pointwise kernel, the central claim stands; if not, the claim needs substantial qualification.","tokens_in":6427,"tokens_out":11371,"duration_ms":128235,"concrete_test":"On CIFAR-100 with the exact reported configuration (n=128, g=4, k=3), retrain MOGNET for at least 3 random training seeds across 5 different CA seeds and report mean and standard deviation of test accuracy. In the same training pipeline, replace the CA-generated final pointwise kernel with (a) a learned ternary kernel of identical shape, reporting the added storage, and (b) a fixed random binary matrix from a different generator or seed. If mean accuracy varies by more than about 0.5% across CA seeds, or if the learned projection beats the CA projection by a significant margin while the total model size stays under 2Mb, then the claim that CA-generated weights are an adequate zero-cost projection is not supported. If accuracy is stable across CA seeds and comparable to the learned projection, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MOGNET beats SLBF by roughly 1% at sub-2Mb model size. The memory advantage comes almost entirely from treating the final 1x1 pointwise weights in CFLOG (Section III-B, Fig. 2, Eq. 4) as zero-cost: they are fixed Rule-30 CA states, not trained or stored. The paper does not test whether this fixed projection is actually expressive: no replacement with a learned ternary projection of the same shape, no replacement with other random projections, no CA-seed sensitivity, and no theoretical bound on rank or conditioning. Because the layer is followed by Batch Normalization, the effective weights can be rescaled, but the binary 0/1 pattern itself is never adapted to the task. If a different CA seed, or a trivial change to the generator, changes accuracy materially, then the reported gap is an artifact of a particular pseudo-random choice rather than a property of the architecture. Additionally, Table II reports a single run per configuration, so the 0.91% CIFAR-100 gap is within plausible run-to-run noise for tiny quantized CNNs. This is an evidence gap rather than a demonstrated contradiction, but it is exactly the load-bearing element of the claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MOGNET, a compact CNN for resource-limited hardware, combining MUX residual blocks with bitshift rescaling, a convolution factorization (CFLOG) whose final pointwise weights are generated online by a Rule-30 cellular automaton (so they are not stored), and a balanced ternary quantization scheme (BTQ). The authors report CIFAR-10 and CIFAR-100 accuracy and model-size numbers in Table II, claiming sub-2Mb models with a roughly 1% accuracy gap over SLBF/ETD baselines at similar or lower model size. The derivation of the CFLOG compression rate is internally consistent if the displayed formulas are read as containing squared quantities, but the empirical support for the headline claim is incomplete: there are no multi-run statistics, no ablation of the CA-generated weights, and no full architecture specification.","tokens_in":6757,"tokens_out":7924,"duration_ms":77050,"significance":"If the results are reproducible, the idea of replacing a stored weight matrix with online-generated CA states is a genuine and creative contribution to ultra-low-memory inference, and the hardware-oriented MRB design with multiplexer-based residual connections is interesting. The paper provides a clean parameter-count formula, reports comparisons against published methods on two standard benchmarks, and explicitly targets a practical sub-2Mb regime that is relevant to edge-AI hardware. However, the absence of seed-sensitivity studies, ablations, and multi-run statistics currently leaves the central accuracy claim unproven; the method is of moderate significance and its empirical evidence must be hardened before the claim can be accepted.","major_comments":[{"comment":"The memory savings and the accuracy claim depend on treating the CA-generated pointwise weights as zero-cost parameters, yet the paper provides no evidence that these fixed Rule-30 projections are as expressive as learned or random projections of the same shape. Please add (i) an ablation replacing the CA kernel with a learned ternary kernel of identical dimensions, (ii) at least one alternative random projection, and (iii) a scan over CA seeds and initial conditions, reporting accuracy variation. Without these experiments, the reported gap may be an artifact of a particular CA seed rather than a property of the architecture.","section":"Section III-B, Eq. (4), Fig. 2"},{"comment":"All accuracy numbers in Table II appear to come from a single run per configuration, and the headline comparison on CIFAR-100 (68.80% versus 67.89%) is a 0.91-percentage-point difference. For small quantized CNNs, run-to-run variance is often of this order, so the claim of a \"clear gap\" is not statistically supported. Please report the mean and standard deviation over at least three independent training runs for each configuration, or otherwise demonstrate that the gap exceeds the noise floor.","section":"Section IV-B, Table II"},{"comment":"The manuscript does not specify the complete network configuration: the number of MRB/CFLOG blocks, the channel schedule, the placement of max pooling, the exact first and last layers, and the CA seed/initial state are not given. This makes the experiments irreproducible and prevents independent verification of the model-size calculations in Eq. (4) and Table II. A layer-by-layer specification and the exact CA setup should be included in the revised manuscript.","section":"Section III, Fig. 1"},{"comment":"The term \"model size\" appears to count only weight memory, while the abstract and introduction claim reductions in \"on-chip required memory.\" Because activations also occupy memory in a hardware implementation, and MOGNET uses 3-bit activations while the baselines use 32-bit activations, the total-memory comparison may differ materially from the weight-only comparison. Please state explicitly whether Table II excludes activation memory, and if the claim is about total on-chip memory, provide an estimate that includes activation buffers.","section":"Section IV, Table II"}],"minor_comments":[{"comment":"The displayed formula appears to have lost superscripts: \"32m^2/g\" and \"32CiCo\" should likely read \"3^2 m^2/g\" and \"3^2 C_i C_o\" (i.e., 9) for a 3x3 convolution, and the simplified expression on the right is consistent with this corrected reading; the typesetting should be fixed.","section":"Section III-B, Eq. (4)"},{"comment":"The symmetry assumption q1 ≈ -q2 is central to the BTQ step-size update, but no empirical validation (e.g., plots of proxy-weight histograms during training) is provided; adding such evidence would strengthen the novelty claim.","section":"Section III-C, Eq. (6)"},{"comment":"The symbol m is used both as the latent dimension in CFLOG and as the maximum value in the TGAP threshold; this overloaded notation is confusing and should be disambiguated.","section":"Section III-A"},{"comment":"The learning-rate schedules are underspecified (e.g., whether the exponential decay is applied per epoch or per iteration); please clarify the schedule details.","section":"Table I"},{"comment":"The arXiv header includes the IEEE AICAS 2022 copyright notice and DOI; the authors should state explicitly what, if anything, has been updated in this arXiv version relative to the published conference paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This manuscript is a reprint of an IEEE AICAS 2022 paper (DOI 10.1109/AICAS54282.2022.9869933), and the editor should consider whether the journal's policy permits submission of an already-published conference paper without substantial new content. The central technical weaknesses are the lack of CA-seed sensitivity analysis and the absence of multi-run statistics; both are fixable by additional experiments. If the authors can provide those experiments and a full architecture specification, the contribution could be viable, but the current evidence is insufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about MOGNET. First, the core idea is concrete: the second pointwise convolution of a ResNext-style factorized block is replaced by a fixed, on-the-fly Rule-30 CA-generated kernel, so the parameter count drops to zero for that layer. Second, the paper never tests whether that fixed kernel is actually good. There is no ablation replacing it with a learned ternary kernel, no seed sensitivity, no multi-seed accuracy. The memory savings are real, but the claimed accuracy edge at sub-2Mb rests on that untested component.\n\nWhat's genuinely good: the MUX residual with Bitshift rescaling is a clean hardware trick—keeps activations at k-bit without a full adder, and the TGAP attention is parameter-free. The balanced ternary step-size update from weight tertiles is simple and plausible; it's the kind of recipe that hardware folks can actually implement. Eq. (4) for the compression rate is consistent with the architecture description, and Table II does show a 1% gap on CIFAR-100 on paper. The authors also stay honest: they note that performance degrades at 1-bit/2-bit and that MOGNET is only competitive below 2Mb.\n\nThe soft spots are exactly where the reader's stress-test lands. The CA weights are 0/1 binary patterns from Rule 30; after BatchNorm the layer can rescale them, but the projection never adapts to the task. A different CA seed could change the effective rank or conditioning of that projection, and we have no evidence it doesn't. Also, all numbers are single runs. For tiny quantized CNNs, 0.9% differences are within run-to-run noise; the CIFAR-100 gap between 67.89 and 68.80 is not established. The model-size metric looks like weights only—activations are not counted, which matters for on-chip memory. The comparison is only against SLBF and ETD, both from 2021; no newer tiny-model baselines are included. These are evidence gaps, not contradictions. The paper does not overclaim: it says 'up to 1%' and bounds its regime.\n\nWho should read this: hardware-oriented researchers working on sub-2Mb inference. They'll find the MUX-bitshift and CA-generated weight sections useful. I'd take it to a reading group as an example of how to present a hardware-motivated architecture with clear reporting of limitations—and as a case study in what ablation evidence is missing.\n\nMy recommendation: send to peer review, but with a major-revision bar. The authors need to release code, provide multi-seed statistics, and ablate the CA generator (learned ternary, other random projections, seed sweep). Without that, the central accuracy claim is plausible but unproven.","headline":"The architecture is real and the paper is honest, but the fixed CA-generated weights are never tested, and single-run numbers can't carry a 1% claim.","tokens_in":7222,"tokens_out":2343,"would_cite":true,"duration_ms":22311,"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":"MOGNET claims a sub-2-Mb quantized network beats larger compressed models on CIFAR.","keywords":["quantized neural networks","cellular automaton weights","ternary quantization","model compression","hardware-efficient deep learning","CIFAR-10","CIFAR-100","edge inference"],"falsifier":"Train the same MOGNET configuration with the CA kernel replaced by (1) a learned 1x1 kernel of equal size, (2) a fixed random matrix initialized from a seeded RNG, and (3) a different CA rule or seed; if (2) or (3) match Rule 30's accuracy at 1.72 Mb, the specific CA mechanism is not load-bearing, and if (1) clearly exceeds it, the fixed projection is costing accuracy that a stored kernel would recover.","tokens_in":6238,"feed_emoji":"🧠","tokens_out":5555,"duration_ms":49690,"temperature":0.7,"pith_summary":"This paper proposes MOGNET, a convolutional architecture designed to run image classification on chips with very little memory. The central claim is that a sub-2-Mb model can match or beat recent compressed networks: the full configuration reports 92.12% on CIFAR-10 and 68.80% on CIFAR-100 at 1.72 Mb, roughly one point above the SLBF-ResNet-18 baseline at a similar size. The savings come from generating part of the weights on the fly with a cellular automaton, so they are never stored, and from a multiplexer-based residual path that keeps activations at 3 bits without extra hardware. The paper also introduces a balanced ternary quantization scheme that sets its step size from the tertiles of the weight distribution during training.","feed_headline":"Sub-2Mb quantized network beats compressed rivals on CIFAR","feed_subtitle":"CA-generated weights cut storage to 1.72 Mb while lifting accuracy roughly a point over SLBF.","key_machinery":"The CFLOG factorization is the backbone: two pointwise convolutions around a grouped convolution, with the second pointwise kernel fixed and generated online by a one-dimensional cellular automaton (rule 30), so the compression rate in Eq. (4) counts those weights as free. The MUX Residual Block (MRB) is an attention-like skip: a channel-wise binary signal from thresholded global average pooling chooses, per channel, between the residual input and the block output, and a BitShift halves the sum so the activation stays in k-bit range. Balanced Ternary Quantization (BTQ) sets the ternary step size each epoch to |q1|+|q2|, the sum of the absolute tertiles of the proxy weights, pushing the quantized levels toward equal occupancy.","core_discovery":"MOGNET replaces every internal convolution except the first and last with a CFLOG block: a 1x1 projection to a low dimension, a grouped convolution, then a 1x1 projection back to full width whose kernel is not trained but produced in real time by Wolfram's rule-30 cellular automaton from a seed. Because this kernel is fixed and generated, its parameters cost no storage, which is what brings the model under 2 Mb. Residual connections are integrated through a multiplexer steered by thresholded global average pooling and a bit-shift rescale, so activations stay at k bits and the operation maps to simple hardware. With 3-bit activations, binarized pointwise and ternarized grouped weights, the paper reports 92.12% on CIFAR-10 and 68.80% on CIFAR-100 at 1.72 Mb, compared with 91.70% and 67.89% for SLBF-ResNet-18 at a similar reported size. It also finds that shrinking activations to 1 or 2 bits costs several points of accuracy.","pith_inferences":["The paper does not ablate the CA kernel choice: a fixed random matrix or another CA rule might give the same accuracy, which would mean the specific cellular-automaton mechanism is not the source of the gain; a direct comparison would separate the storage trick from the projection's representational quality.","Because BTQ's tertile rule depends on the proxy weights staying roughly symmetric around zero, it may transfer to other ternary networks as a drop-in quantizer, independent of the CA-generated weights.","The MUX-plus-bitshift residual path is a candidate building block for any quantized architecture targeting integer-only hardware, since it replaces an addition with a 2-input mux and a shift."],"forward_implications":["At 1.72 Mb with 3-bit activations, the n=128, g=4 MOGNET reaches 92.12% on CIFAR-10 and 68.80% on CIFAR-100, beating SLBF-ResNet-18 by about a point at similar size.","The CA-generated final pointwise kernel removes an entire class of stored parameters from each block, so the memory saving scales with the number of blocks.","Reducing activation precision to 2 bits or 1 bit degrades accuracy substantially (91.16% and 88.99% on CIFAR-10), so 3-bit activations appear necessary for the claimed accuracy.","Beyond n=128, MOGNET's accuracy-versus-size curve falls below SLBF-RN18, so the reported advantage is specific to the sub-2-Mb regime."],"supporting_citations":[{"why":"Provides the pointwise-grouped-pointwise block shape that CFLOG simplifies.","marker":"[16]"},{"why":"Defines the rule-30 cellular automaton used to generate the stored-free weights.","marker":"[17]"},{"why":"Supplies the residual idea that MRB adapts into a multiplexer skip.","marker":"[6]"},{"why":"Source of the channel attention concept behind thresholded global average pooling.","marker":"[8]"},{"why":"Straight-through estimator used to train through QReLU and ternary quantizers.","marker":"[20]"},{"why":"SLBF is the main baseline MOGNET claims to beat by about one point at similar size.","marker":"[26]"},{"why":"ETD provides a second tensor-decomposition compression baseline on CIFAR.","marker":"[28]"}],"fun_headline_variants":["Sub-2Mb MOGNET with online CA weights tops CIFAR accuracy","CA-generated weights cut model to 1.72 Mb, lift CIFAR scores","MOGNET: 1.72 Mb, CA-generated kernels, higher CIFAR accuracy","Online CA weights enable sub-2Mb quantized net that beats SLBF","MOGNET uses cellular automaton to fit 1.72 Mb and win on CIFAR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that fixed, untrained Rule-30 CA-generated projection weights are expressive enough to replace a learned final pointwise convolution; the paper reports no experiment that varies the seed, rule, or a learned substitute.","fun_headline_variants_meta":{"raw":{"variants":["Sub-2Mb MOGNET with online CA weights tops CIFAR accuracy","CA-generated weights cut model to 1.72 Mb, lift CIFAR scores","MOGNET: 1.72 Mb, CA-generated kernels, higher CIFAR accuracy","Online CA weights enable sub-2Mb quantized net that beats SLBF","MOGNET uses cellular automaton to fit 1.72 Mb and win on CIFAR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1563,"prompt_tokens":948,"completion_tokens":615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":502}},"tokens_in":564,"tokens_out":615,"duration_ms":6909,"temperature":1.0,"reasoning_tokens":502,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:55:42.754810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same MOGNET configuration with the CA kernel replaced by (1) a learned 1x1 kernel of equal size, (2) a fixed random matrix initialized from a seeded RNG, and (3) a different CA rule or seed; if (2) or (3) match Rule 30's accuracy at 1.72 Mb, the specific CA mechanism is not load-bearing, and if (1) clearly exceeds it, the fixed projection is costing accuracy that a stored kernel would recover.","supporting_citations":[{"cited_title":"Aggregated residual transformations for deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Provides the pointwise-grouped-pointwise block shape that CFLOG simplifies."},{"cited_title":"Wolfram, A New Kind of Science","cited_arxiv_id":null,"evidence_quote":"Defines the rule-30 cellular automaton used to generate the stored-free weights."},{"cited_title":"Compressing deep convolutional neural networks by stacking low-dimensional binary convolution filters,","cited_arxiv_id":null,"evidence_quote":"SLBF is the main baseline MOGNET claims to beat by about one point at similar size."},{"cited_title":"Towards efficient tensor decomposition-based DNN model compression with optimization frame- work,","cited_arxiv_id":null,"evidence_quote":"ETD provides a second tensor-decomposition compression baseline on CIFAR."}],"review_version":1}