{"id":"bfaecbe2-2b89-4bb7-b452-6d2bcf501655","arxiv_id":"1908.07942","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A design-space study finds that FeFET conductance nonlinearity causes significant accuracy loss on EMNIST, which hardware-aware training and network over-parametrization can partially recover.","lead":"This paper simulates a ferroelectric-FET in-memory DNN accelerator and measures how device conductance non-linearity, ADC resolution, and bit depth affect classification accuracy on the EMNIST dataset. It shows that hardware-aware training and deeper networks can partially offset device limitations, giving designers a map of usable operating points.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The HATI training loop is never specified; the claimed recovery from FeFET nonlinearity is therefore not reproducible and may be an artifact of an assumed training scheme.","rationale":"The reader's CONDITIONAL verdict is appropriate and I do not recommend changing it. My concern overlaps with the reader's rationale about missing training details, but I would place it even more centrally than the device-model uncertainty: even if the sigmoidal conductance model in Eq. (1) is exactly right, the paper's proposed remedy (HATI) is not described well enough to know whether the reported accuracy is a genuine property of FeFET-based training or a consequence of an idealized digital write assumed during training. The device-model concern is real, but it mostly shifts the absolute numbers; the HATI concern threatens the validity of the main practical conclusion. The paper does have independent support: the sigmoidal model is anchored to measured FeFET data at alpha=0.4, and the differential-cell architecture is clearly explained. Those elements should be credited. The requested check is a single reimplementation experiment that would settle whether HATI, as a defined procedure, can actually deliver the claimed recovery. Because the manuscript lacks enough detail to define HATI uniquely, the appropriate status remains conditional rather than outright acceptance or rejection.","tokens_in":5096,"tokens_out":9439,"duration_ms":101055,"concrete_test":"Re-run the alpha=0.4, 8-bit/cell, 10-bit ADC, one-hidden-layer 100-neuron configuration under two explicit HATI implementations: (i) forward pass uses the Eq. (1) nonlinear map plus quantization with straight-through gradients, and no write-pulse dynamics; (ii) each training update advances conductance along Eq. (1) by a finite pulse step, with gradients taken with respect to the underlying continuous variable. If neither implementation reproduces the reported 67.03% within about two percentage points, the HATI claim is not well defined by the manuscript. At minimum, the authors should state which implementation was used and release the training code.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central remedy of the paper, Hardware-Aware Training and Inference (HATI), is defined in only one sentence (Sec. II-C): 'both training and inference are done on the FeFET PIM.' No training algorithm, loss function, backward pass, or weight-programming rule is given. This matters because Eq. (1) describes conductance as a sigmoidal function of the number of write pulses, so a physically faithful HATI must specify whether weights are (a) optimized in a continuous digital space and mapped to conductance levels only at inference, or (b) updated by incremental pulses whose nonlinearity enters the learning dynamics themselves. These two choices will produce different accuracies, and the paper's main quantitative result—HATI at alpha=0.4 reaching 67.03% (or 67.38% in the text; the discrepancy is itself a symptom) versus 58.47% for HAI—depends entirely on this unspecified choice. If HATI is just quantization-aware training with a straight-through estimator, then the reported advantage is not a property of FeFET write dynamics; if it uses pulse-based updates, the backward pass through a saturating sigmoid is nontrivial and must be described. As written, the claimed recovery via hardware-aware training cannot be verified or reproduced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a simulation-based design-space exploration of a Ferroelectric FET (FeFET) processing-in-memory DNN accelerator for the EMNIST Balanced dataset. The authors model FeFET conductance as a sigmoidal function of write-pulse count with a single nonlinearity parameter alpha, compare three schemes (floating-point baseline, hardware-aware inference HAI, and hardware-aware training and inference HATI), and vary the ADC resolution, number of bits per FeFET cell, and network depth. The main claims are that conductance nonlinearity degrades accuracy, that HATI partially recovers the loss relative to HAI, that 4-bit/cell split-cell arrays outperform 8-bit/cell arrays, and that over-parametrization via additional hidden layers helps HATI models.","tokens_in":5302,"tokens_out":3225,"duration_ms":34874,"significance":"If the results hold, the paper provides a useful first-order quantitative mapping between FeFET device nonlinearity and achievable DNN accuracy, and it makes concrete, falsifiable design suggestions (e.g., 10-bit ADC is sufficient, HATI is preferable at high nonlinearity, deeper HATI networks are more robust). The strengths are the use of a physically motivated sigmoidal conductance model calibrated to the authors' prior experimental device, the explicit comparison of HAI versus HATI, and the exploration of a multi-dimensional design space. However, the central HATI mechanism is not specified, the accuracy differences supporting the 4-bit/cell conclusion are small, and no statistical repeats are provided, so the quantitative claims should be treated as preliminary until the training procedure is fully described.","major_comments":[{"comment":"The HATI model is defined in a single sentence: \"both training and inference are done on the FeFET PIM.\" No loss function, optimizer, gradient computation method, weight-programming rule, or quantization-aware training procedure is given. This matters because Eq. (1) describes conductance as a sigmoidal function of write-pulse count; HATI could mean continuous weight optimization with pulse mapping only at inference, or pulse-based updates whose nonlinearity enters the learning dynamics, and these choices yield different accuracy results. The paper's central claim that HATI outperforms HAI at alpha=0.4 cannot be verified or reproduced until this training algorithm is fully specified.","section":"Sec. II-C and Sec. III, Fig. 6"},{"comment":"The reported HATI accuracy for alpha=0.4 is inconsistent: the main text states 67.38%, while the table in Fig. 6(b) reports 67.03%. Since this specific configuration anchors the paper's headline comparison with the 74.92% floating-point baseline, the discrepancy must be resolved and all numerical values cross-checked.","section":"Sec. III, text and Fig. 6"},{"comment":"The conclusion that 4-bit/cell FeFET arrays outperform 8-bit/cell arrays relies on splitting 8-bit weights into two 4-bit cells, which doubles the array area. Some of the reported differences are small (e.g., 66.02% vs. 65.54% at alpha=0.25, ADC=8 bits), and no statistical repeats or error bars are reported. The claim needs either a significance analysis or an explicit statement of whether the comparison is for a fixed total area/energy budget rather than a per-cell precision comparison.","section":"Sec. III, Fig. 5"},{"comment":"The sigmoidal conductance model is fitted to one measured device with alpha=0.4, and alpha is then varied from 0.25 to 2.0 without experimental validation or a physical mapping for those values. Because the design-space conclusions depend on the functional form of Eq. (1), the authors should either validate the model across the alpha range or provide a sensitivity analysis showing that the main conclusions are robust to alternative nonlinearity models (e.g., convex/concave power-law or exponential forms).","section":"Sec. II-A, Eq. (1), Fig. 2"}],"minor_comments":[{"comment":"Equations (1) and (2) are garbled in the submitted text and should be typeset clearly with their symbols and bounds defined in the caption.","section":"Equations"},{"comment":"There is a typo on \"int eh\" that should read \"in the\", and the text contains other minor spelling issues such as \"Kirchoff's\" for \"Kirchhoff's.\"","section":"Sec. III"},{"comment":"Reference [8] appears to duplicate reference [4]; the bibliography should be deduplicated and checked for completeness.","section":"References"},{"comment":"The HAI description states that weights are scaled to [-1,1] during training, but it is not stated how the trained weights are quantized to the number of cell states or how the ADC conversion and sigmoid activation are implemented in the simulation. These details should be added for reproducibility.","section":"Sec. II-C"},{"comment":"Figure 7 lacks a legend or explicit caption labels for the curves; the figure should be annotated so the reader can identify the configurations being compared.","section":"Sec. III, Fig. 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a short design-space study with a plausible simulation approach, but the HATI training scheme is underspecified to the point that the main quantitative results are not reproducible. I would advise the editor that the manuscript requires a substantive revision, including a full description of the training algorithm and clarification of the reported accuracy inconsistencies, before it can be considered for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read of Yoon et al. The useful part is the design-space mapping: they model FeFET conductance with a sigmoidal curve, fit alpha to their own measured device, and show how ADC resolution, bits/cell, and network depth shift EMNIST accuracy. That's a real addition over the earlier MNIST work, and the overparameterization result — HATI accuracy goes up with more layers while HAI degrades — is a genuine insight worth reporting.\n\nThe soft spot is exactly where the stress-test lands. HATI is defined in one sentence: 'both training and inference are done on the FeFET PIM.' No training algorithm, loss, backward pass, or weight-programming rule is given. The reported recovery from nonlinearity depends entirely on that unspecified choice. If weights are optimized in continuous space and only mapped to conductance at inference, that's quantize-aware training; if updates are pulse-based, the nonlinearity enters the learning dynamics. The paper doesn't tell us which, and the two give different accuracy numbers. So the headline quantitative claim — HATI at alpha=0.4 reaching roughly 67% vs 58% for HAI — cannot be reproduced or fully trusted as written. There's also a small internal inconsistency: the text says 67.38% for that case while Table VI says 67.03%; not a big deal by itself, but it signals sloppy final checks.\n\nOther nits: no error bars or repeated runs, training hyperparameters are not reported, and 'target performance' is mentioned but never defined. All of that is fixable in a revision, and for a four-page letter some brevity is expected. The conductance model itself is physically motivated and calibrated to real data, and the comparison of 4-bit vs 8-bit cells is at least internally consistent, even if I'd want more analysis before taking it as general.\n\nBottom line: the paper is a useful engineering study with a plausible central message — device nonlinearity hurts, hardware-aware training and overparameterization help. But the missing HATI details and the numeric inconsistency make the exact numbers unverifiable. I'd send it to peer review, not desk-reject it, and ask for a major revision that specifies the training loop, fixes the discrepancy, and ideally releases code or data. A reader working on FeFET PIM accelerators would get value from this; a general ML audience would not.","headline":"A useful FeFET-PIM design-space study with a genuine overparameterization insight, but the central HATI comparison is underspecified and the numbers conflict internally, so exact claims need a revision.","tokens_in":5851,"tokens_out":2636,"would_cite":false,"duration_ms":25936,"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":"FeFET nonlinearity cuts DNN accuracy by 7.5 points","keywords":["FeFET","processing-in-memory","DNN accelerator","conductance nonlinearity","hardware-aware training","EMNIST","differential memory cell","over-parametrization"],"falsifier":"Program the weights from a HATI-trained network (α=0.4) into an actual FeFET array and measure EMNIST accuracy; if the measured accuracy does not land near the simulated 67.38%, or if the fitted α values from devices processed at different anneal temperatures fall outside the swept 0.25–2.0 range, the sigmoidal model or the simulation's treatment of device behavior is wrong.","tokens_in":4888,"feed_emoji":"🧠","tokens_out":5532,"duration_ms":50645,"temperature":0.7,"pith_summary":"This paper asks how much the imperfect analog behavior of ferroelectric-FET (FeFET) memory cells hurts a neural network that runs inside the memory array, and what design choices can compensate. Using the EMNIST letter-classification benchmark, it shows that the sigmoidal nonlinearity of FeFET conductance—modeled by a single parameter α—degrades accuracy substantially: at the experimentally fitted α=0.4, hardware-only inference (HAI) drops to about 58% versus 74.92% for a floating-point network. The paper's central demonstration is that hardware-aware training (HATI), where the device model is used during training, recovers some of the loss (to about 67.4%), and that making the network deeper (over-parametrized) further improves HATI accuracy, even though depth hurts the hardware-agnostic HAI model. The practical stake is that FeFET-based processing-in-memory could accelerate DNNs, but only if device linearity, ADC resolution, and bits-per-cell are co-designed with training.","feed_headline":"FeFET nonlinearity cuts DNN accuracy by 7.5 points","feed_subtitle":"Hardware-aware training and deeper networks recover part of that loss on EMNIST.","key_machinery":"The central object is the sigmoidal conductance model of Eq. (1): $G(x) = G_{\\min} + (G_{\\max}-G_{\\min})/(1+e^{-\\alpha x})$, where $x$ is the pulse-amplitude-coded input and $\\alpha$ quantifies the steepness and nonlinearity of the device's analog response. This model, validated against measured FeFET pulse data at $\\alpha=0.4$, replaces the convex/concave curves used in earlier resistive-memory studies and is physically motivated by the Gaussian distribution of coercive fields in the HfZrO$_2$ ferroelectric. The other key mechanism is the differential FeFET cell, which stores positive and negative weights in two devices so that matrix-vector multiplication is computed in O(1) by summing column currents, enabling negative activations without a separate array.","core_discovery":"The paper claims that FeFET conductance as a function of write-pulse amplitude is sigmoidal rather than convex or concave, because it reflects the Gaussian distribution of coercive fields in the ferroelectric layer; this nonlinearity is captured by a parameter α that ranges from 0.25 to 2.0. With this model, the authors show that classifying EMNIST with a fully connected network stored in FeFET cells loses about 20% accuracy as α varies from 0.25 to 2.0, and that even the best hardware-aware-trained network (α=0.4) reaches 67.38% versus the 74.92% floating-point baseline. They further demonstrate that splitting an 8-bit weight across two 4-bit FeFET cells beats storing it in one 8-bit cell, that a 10-bit ADC captures nearly all available accuracy, and that increasing the number of hidden layers helps HATI but hurts HAI because errors accumulate in the hardware-agnostic case.","pith_inferences":["The same over-parametrization plus hardware-aware training recipe likely extends to convolutional and residual networks, where the error-propagation effect of device nonlinearity may differ across layers; the paper does not test those architectures.","Because the sigmoidal model ties α to the coercive-field distribution, process-level engineering of that distribution or write-verify pulse schemes could linearize the effective conductance more cheaply than requiring intrinsically linear devices.","The study ignores cycle-to-cycle variation, read noise, and endurance; if those correlate with α, the reported HATI accuracy numbers may be optimistic in hardware.","The differential cell's support for negative activations suggests tanh or leaky-ReLU networks may interact differently with nonlinearity than the sigmoid-equipped networks studied, so the depth result may not directly transfer to such networks."],"forward_implications":["With the measured FeFET nonlinearity (α=0.4), even hardware-aware training leaves a roughly 7.5-point accuracy gap to a floating-point network on EMNIST (67.38% vs 74.92%).","Splitting an 8-bit weight into two 4-bit FeFET cells yields higher accuracy than a single 8-bit cell, at the cost of doubling the array area.","A 10-bit ADC is sufficient; going to 12 bits gives only marginal accuracy gains.","Deeper, over-parametrized networks improve HATI accuracy but degrade HAI accuracy, so design should pair deeper networks with hardware-aware training.","Nonlinearity has a strong effect: changing α from 0.25 to 2.0 drops accuracy by about 20%, so more linear devices remain desirable."],"supporting_citations":[{"why":"Provides the EMNIST Balanced dataset, the more complex benchmark whose accuracy quantifies the device-imperfection loss.","marker":"[1]"},{"why":"Supplies the measured FeFET pulse data used to validate the sigmoidal conductance model at α=0.4 and the device fabrication basis.","marker":"[2]"},{"why":"Establishes the crossbar-based O(1) matrix-vector multiplication scheme that the FeFET PIM architecture builds on.","marker":"[3]"},{"why":"Gives the convex/concave nonlinearity models of resistive synapses that the paper replaces with the sigmoidal FeFET fit.","marker":"[4]"},{"why":"Explains how the distribution of coercive fields in the ferroelectric determines the hysteresis shape, grounding the sigmoidal model physically.","marker":"[16]"}],"fun_headline_variants":["FeFET sigmoid transfer cuts DNN accuracy by 7.5","Hardware-aware training trims FeFET DNN accuracy loss","4-bit FeFET cells beat 8-bit for DNN accelerator","Deeper nets help HATI but hurt HAI in FeFET DNNs","10-bit ADC yields near-full FeFET DNN accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The sigmoidal conductance curve with a single fitted parameter α, validated only at α=0.4, is assumed to capture real FeFET behavior over the whole α range and across all operating conditions; if the true device response is asymmetric, noisier, or differently shaped, the accuracy numbers and design conclusions shift.","fun_headline_variants_meta":{"raw":{"variants":["FeFET sigmoid transfer cuts DNN accuracy by 7.5","Hardware-aware training trims FeFET DNN accuracy loss","4-bit FeFET cells beat 8-bit for DNN accelerator","Deeper nets help HATI but hurt HAI in FeFET DNNs","10-bit ADC yields near-full FeFET DNN accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001056,"raw_usage":{"total_tokens":4366,"prompt_tokens":816,"completion_tokens":3550,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":432,"completion_tokens_details":{"reasoning_tokens":3457}},"tokens_in":432,"tokens_out":3550,"duration_ms":25322,"temperature":1.0,"reasoning_tokens":3457,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:44:38.550365+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Program the weights from a HATI-trained network (α=0.4) into an actual FeFET array and measure EMNIST accuracy; if the measured accuracy does not land near the simulated 67.38%, or if the fitted α values from devices processed at different anneal temperatures fall outside the swept 0.25–2.0 range, the sigmoidal model or the simulation's treatment of device behavior is wrong.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the EMNIST Balanced dataset, the more complex benchmark whose accuracy quantifies the device-imperfection loss."},{"cited_title":"Acceleration of deep neural network training with resistive cross-point devices: Design considerations,","cited_arxiv_id":null,"evidence_quote":"Establishes the crossbar-based O(1) matrix-vector multiplication scheme that the FeFET PIM architecture builds on."},{"cited_title":"Schenk et al","cited_arxiv_id":null,"evidence_quote":"Explains how the distribution of coercive fields in the ferroelectric determines the hysteresis shape, grounding the sigmoidal model physically."}],"review_version":1}