{"id":"73957eaa-cf2c-4b8f-aac0-6c598be6e523","arxiv_id":"2608.03448","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A quantized ResNet running on an STM32 microcontroller reconstructs target displacement from self-mixing interferometry signals in real time with inference below 1 ms.","lead":"The paper demonstrates a compact sensor that uses a laser and a neural network on a microcontroller to track target motion in real time. This matters because it shows a low-power, standalone device can replace complex fringe-counting analysis for displacement measurements, potentially enabling portable vibration and flow sensors.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported displacement accuracy is validated only against a speaker-voltage proxy; a biased λ/V calibration would shift all PCC/MAE numbers, though the 0.64 ms latency claim is unaffected.","rationale":"The reader's weakest_assumption is exactly the load-bearing concern I find: the ground truth is a calibrated speaker voltage, not an independent displacement measurement. The central claim has two coupled parts: (1) the network runs in 0.64 ms on the STM32 and (2) the quantized model preserves accuracy (PCC 0.889 vs 0.884). Part (1) is solid: Table III reports direct timing measurements, and the live demo shows a working continuous pipeline. Part (2) is only as trustworthy as the ground truth. The comparison int16 vs float32 is still meaningful because both are evaluated against the same labels, but the absolute accuracy and the physical meaning of 'displacement inference' are open. The paper's own Section V limitation statement corroborates this. I do not see an internal logical flaw or a deceptive practice; the engineering contribution is real. The proposed LDV/piezo-capacitive check would settle whether the proxy bias matters. Since the reader already assigned CONDITIONAL for this reason, no adjustment is needed.","tokens_in":13654,"tokens_out":9525,"duration_ms":101267,"concrete_test":"Mount the same target speaker on an optical bench and measure its true displacement with an independent reference (e.g., a commercial laser Doppler vibrometer, or a piezo stage with a capacitive sensor) while acquiring the SMI signal and the speaker voltage. Use the aperiodic motions from the generalization set, and recompute the int16 ResNet's PCC/MAE against the independent displacement. If MAE stays within ~0.353 ± 10% and no frequency-dependent bias exceeds 10%, the proxy is adequate. If the error rises substantially (e.g., MAE > 0.5 λ/ms or a clear 10%+ bias at some frequencies), the reported accuracy is an artifact of the speaker-voltage calibration.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The quantitative part of the central claim—PCC 0.889 / MAE 0.353 λ/ms for the int16 ResNet—rests on ground-truth labels that are not independent optical measurements. Section II.B states that the generalization labels come from 'a calibration of the λ/V scaling factor or a speaker which is linear (within 10%)'; Section IV.E repeats that the live 'true displacement' is the time-derivative of the speaker driving voltage scaled by the same λ/V factor. The network is trained to map the SMI signal to this speaker-voltage-derived displacement, and the evaluation compares against the same proxy. Consequently, any frequency-dependent nonlinearity, hysteresis, or calibration bias in the speaker is absorbed into the learned mapping and cannot be detected by the reported metrics or by the agreement in Fig. 8. The paper itself acknowledges in Section V that accuracy is 'ultimately bound by the ground truth accuracy of the training set.' This is not an internal inconsistency, but it is an external-validity risk: the sensor may accurately reproduce the proxy while being biased with respect to true mechanical displacement. The real-time feasibility claim (0.64 ms inference on STM32U575ZI, Table III) is a timer-based measurement and does not depend on label accuracy.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an embedded self-mixing interferometry (SMI) sensor that combines a semiconductor laser front-end with an STM32U575ZI microcontroller running a quantized 16-bit ResNet. The network reconstructs target displacement from 1-ms windows of the interferometric signal, and the authors demonstrate fully on-device real-time operation with 0.64 ms inference latency per window, which is below the 1 ms real-time constraint. They compare the ResNet against a prior CNN, evaluate robustness to input amplitude variations via an input normalization layer, study 8/16-bit quantization trade-offs, measure memory/latency/energy, and provide a live reconstruction example. The central claims are that the int16 ResNet meets real-time constraints and achieves generalization performance comparable to float32 (PCC 0.889 vs. 0.884, MAE 0.353 vs. 0.364 λ/ms).","tokens_in":14006,"tokens_out":7371,"duration_ms":73565,"significance":"If the results hold, this is a valuable engineering contribution: it is the first demonstration, to my knowledge, of SMI displacement reconstruction performed fully on a low-power microcontroller in real time using a neural network. The 0.64 ms inference time on a 160 MHz Cortex-M33 with 6937 parameters is a concrete, measured result that invites direct comparison with future work. The amplitude-normalization layer is a simple and practical way to make the sensor robust to target reflectivity variations. The paper also provides useful memory, latency, and energy data for two MCU families. However, the quantitative accuracy numbers are tied to a speaker-voltage proxy for ground truth, and the energy measurement protocol appears to mix batch and streaming modes; these issues need clarification before the claims can be fully accepted.","major_comments":[{"comment":"The reported accuracy metrics (PCC, MAE, LRS) are computed against a ground truth that is derived from the voltage applied to the speaker, scaled by a λ/V factor claimed linear within 10%. The network is trained on this same proxy and evaluated against it, so any speaker nonlinearity, hysteresis, or calibration bias is absorbed into the learned mapping and cannot be detected by the reported numbers. While Section V acknowledges that accuracy is ultimately bounded by the training-set ground truth, this limitation is not reflected in the abstract or the central accuracy claims. Please provide either an independent optical calibration for at least a subset of the evaluation data, or a quantitative uncertainty analysis showing how a ±10% calibration error propagates to PCC/MAE. This is load-bearing for the sensor's measurement-accuracy claim, though not for the latency claim.","section":"Section IV.D, Tables IV-V"},{"comment":"The energy measurements use a 500 ms input buffer and report 'Total time' values (e.g., 797 ms for the STM32U575ZI ResNet) that exceed the 500 ms window. This suggests the measurement does not reflect the double-buffered streaming operation described in Section III.B and used in the live demo. As written, the configuration that is claimed to be real-time appears to take 797 ms to process 500 ms of signal, which is contradictory. Please clarify the measurement timeline, specify whether the MCU is active during data collection, and report energy under the actual real-time streaming mode. If the current numbers are intended as a conservative batch-mode estimate, state that explicitly and give the streaming-mode energy as well.","section":"Section IV.B, Table I"},{"comment":"The conclusion that int16 quantization does not significantly degrade performance and int8 causes a loss is based on single training runs and single inference passes, with no confidence intervals or repeated experiments. The differences in PCC between float32 and int16 are about 0.005, which could easily be within training variability. Please provide multiple training runs with different seeds (or another statistical treatment) to support the claim that int16 is the preferred format. This is directly relevant to the deployment choice made in the paper.","section":"Section IV.E and Conclusion"},{"comment":"The sentence in the conclusion stating that 'a displacement with frequency up to 500Hz could in principle be measured (from the sampling theorem)' is misleading. The network is trained only on target motions in the 5–100 Hz range (Section II.B and V), and the 48 kHz sampling rate alone does not guarantee that the learned mapping generalizes to 500 Hz. Please remove or heavily qualify this statement, or provide a dedicated validation at higher frequencies.","section":"Section II.B"}],"minor_comments":[{"comment":"The row 'CNN [24] float32 0.8551 0.74' appears incomplete; it is missing the MSE and MAE columns that are filled for the other rows. Please correct the formatting or provide the missing values.","section":"Table I"},{"comment":"The abbreviation 'RMPSprop' is likely a typo for 'RMSprop'. Also, the explanation of which model received random scaling during training could be clearer: the text says 'both networks are trained' with random scaling, but then says scaling 'has no impact' on the normalized network; please spell out the exact training protocol for each model.","section":"Section IV.A"},{"comment":"The phrase 'the first demonstration' is a strong precedence claim. Unless a comprehensive prior-art search confirms uniqueness, please soften to 'a first demonstration' or provide specific comparative evidence.","section":"Section IV.E"},{"comment":"There is a typo: 'an ajustable constant-current power supply' should be 'an adjustable constant-current power supply'.","section":"Section IV.D.1"},{"comment":"The phrase 'this dataset comes from a calibration of the λ/V scaling factor or a speaker which is linear' is ungrammatical. Presumably 'of' was intended. Please rephrase to make clear whether the calibration and linearity are separate checks or alternatives.","section":"Section II.B"}],"recommendation":"major_revision","confidential_remarks":"The paper is well within the scope of this venue, and the real-time inference result is significant. The self-citations (refs [17], [18], [24], [26]) are relevant and the dataset is self-produced, which is acceptable for a prototype paper. I do not see a circularity problem beyond the speaker-proxy issue flagged in the major comments. The main risk is that the accuracy claims are over-stated given the proxy ground truth, and the energy measurement protocol needs clarification. These are fixable with additional analysis and revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things before opening this one. First, the real-time claim is credible and well-measured: a 16-bit quantized ResNet with 6937 parameters runs in 0.64 ms per 1 ms window on an STM32U575ZI using CMSIS-NN, and the live demo in Fig. 8 looks convincing. Second, the accuracy numbers (PCC 0.889, MAE 0.353 λ/ms) are only as good as the ground truth, which is derived from speaker drive voltage scaled by a λ/V factor, not from independent optical measurement. That is not a fatal flaw for a prototype paper, but it is a real limitation that should be stated more prominently per section.\n\nWhat is genuinely new: this is the first demonstration of fully embedded real-time SMI displacement reconstruction, combining laser, ADC, and MCU in one loop. The construction is careful: they compare two MCUs, with and without CMSIS-NN, and report ROM/RAM footprints and energy consumption. The input normalization layer is a sensible addition for amplitude robustness and is shown to extend the operating range over four orders of magnitude. They also retrain the older CNN under the same conditions for a fair comparison. The out-of-distribution aperiodic test set, taken with a different feedback condition and C parameter, gives a real generalization check rather than an in-sample curve fit. The dataset is public, which is good practice.\n\nSoft spots: the ground-truth proxy issue is the main one. The speaker is linear within 10% over 10–100 Hz, but that linearity uncertainty propagates directly into the reported errors. No error bars are given on PCC or MAE, so we do not know whether the int16 vs float32 differences are meaningful (they look like noise). The paper acknowledges the ground-truth limitation in the conclusion, but it should be in the abstract and results discussion. Also, the live demo only shows one frequency-modulated signal; it is not a stress test across the claimed 5–100 Hz and up to 2.5 λ/ms range. Minor: the sentence claiming the main limitation is the neural network rather than ground truth needs reconciling with the earlier statement that accuracy is bound by ground truth.\n\nThe self-citations are appropriate here, as they are building on their own prior CNN work and dataset. The paper is not overclaiming novelty; it explicitly says this is an updated prototype. It is an engineering contribution, not a physics breakthrough, but it is a solid systems-level result.\n\nWho is this for? People working on embedded ML for optical sensors, and the SMI community interested in practical real-time processing. I would bring it to a reading group as an example of careful embedded deployment. I would not cite the accuracy numbers without noting the calibration caveat.\n\nRecommendation: send it to peer review. It deserves referee time, and the main revision request should be independent displacement validation (e.g., an interferometric reference) and error bars, not a desk rejection.","headline":"Genuine real-time embedded SMI sensor demo with solid latency numbers, but the accuracy metrics rest on a speaker-voltage proxy that could hide calibration bias.","tokens_in":14432,"tokens_out":1441,"would_cite":true,"duration_ms":17718,"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":"A 6,937-parameter neural network, quantized to 16 bits and run on a 160 MHz microcontroller, reconstructs target displacement from laser self-mixing signals in 0.64 ms per 1 ms window, meeting the real-time constraint while matching float32","keywords":["self-mixing interferometry","embedded neural network","displacement sensing","real-time inference","ResNet","quantization","microcontroller","CMSIS-NN"],"falsifier":"Replace the calibrated speaker with an independent displacement reference, such as a commercial laser interferometer or a capacitive sensor measuring the same target, and compare the network's inferred displacement to that reference across the full operating envelope (5–100 Hz, below 2.5 $\\lambda$/ms, feedback parameter C near unity); any systematic disagreement larger than the reported mean absolute error of roughly 0.35 $\\lambda$/ms would indicate that the speaker-voltage proxy is contaminating the accuracy claims.","tokens_in":13586,"feed_emoji":"📏","tokens_out":5846,"duration_ms":59859,"temperature":0.7,"pith_summary":"Self-mixing interferometry measures target motion by feeding a laser's own reflected light back into the laser and reading the resulting voltage changes, but decoding that signal is notoriously difficult. This paper shows that a very small residual neural network with 6,937 parameters after batch-normalization fusion can perform that decoding accurately enough for real-time use when quantized to 16-bit fixed point and run on a 160 MHz microcontroller with CMSIS-NN optimizations. The network processes each 1 ms, 48-sample window in 0.64 ms, leaving time to spare, and its accuracy on out-of-distribution aperiodic signals (Pearson correlation 0.889, mean absolute error 0.353 $\\lambda$/ms) essentially matches the full-precision model (0.884, 0.364 $\\lambda$/ms). A per-window input normalization layer keeps performance flat across four orders of magnitude of signal amplitude, relaxing the analog front-end constraints. If this holds, standalone, battery-powered vibration and displacement sensors with built-in neural decoding are within reach.","feed_headline":"Tiny neural network runs laser displacement sensor in real time","feed_subtitle":"A 16-bit ResNet on an STM32 measures target motion at kHz rate, using under 1 ms per sample.","key_machinery":"A residual convolutional neural network with three residual blocks (kernel length 3, residual branches kernel length 1), global average pooling, and a dense regression head, preceded by a fixed SampleNormLayer that scales each 48-sample window to [0, 1]. The network is trained on periodic motions, post-training quantized to 16-bit fixed point (Q7.9), batch-normalization is fused into the preceding convolutions, and the Qualia toolchain generates portable C code that uses CMSIS-NN's SIMD-optimized convolution kernels. The normalization layer provides amplitude robustness, the compressed ResNet fits the microcontroller's memory and compute budget, and 16-bit quantization preserves accuracy whi","core_discovery":"The paper establishes that a ResNet with 6,937 parameters, operating on non-overlapping 1 ms windows of 48 samples at 48 kHz, can reconstruct target displacement from a self-mixing interferometry signal on a microcontroller within the real-time constraint of 1 ms. With 16-bit post-training quantization (Q7.9 format), fused batch-normalization layers, and CMSIS-NN kernels on an STM32U575ZI, inference takes 0.64 ms and uses 22.9 kB RAM and 76.4 kB ROM, while generalization performance on aperiodic out-of-distribution displacements is statistically indistinguishable from float32 (PCC 0.889 vs 0.884; MAE 0.353 vs 0.364 $\\lambda$/ms). The per-window sample normalization layer makes the network in","pith_inferences":["Because the network learns fringe shape rather than signal amplitude, the same architecture could likely be retrained to estimate other self-mixing parameters (for example the feedback parameter C or target distance) from the same voltage signal, turning the sensor into a multi-parameter instrument.","The reported accuracy is measured against a speaker-voltage proxy for displacement, so an independent optical reference, such as a commercial laser interferometer, would be a stronger validation; the 10% linearity tolerance of the speaker calibration is likely the dominant systematic uncertainty in the absolute numbers.","The sub-millisecond headroom suggests the same pipeline could be ported to other microcontroller families or to hardware accelerators, and at higher sampling rates the architecture could be deepened or the window length extended, trading the current latency margin for higher resolution or a wider frequency band."],"forward_implications":["A standalone, battery-powered displacement/vibration sensor is feasible: total system power is about 600 mW, dominated by the unoptimized analog front end, and one prediction costs 0.167 $\\mu$Wh, so AA batteries could sustain roughly 10 hours of continuous operation.","The sensor is robust to target reflectivity and laser-dependent signal amplitude variations because of the per-window normalization layer, so the same network can be used across different target materials and even different laser systems.","Real-time operation at 1 ms per window implies the system could in principle measure displacement frequencies up to about 500 Hz, bounded by the 48 kHz sampling rate and the training domain (5–100 Hz, below 2.5 $\\lambda$/ms, feedback parameter C near unity).","8-bit quantization is not worth the accuracy loss (linear-regression slope drops by 0.029 compared with float32) under this simple uniform quantization scheme; the authors choose 16-bit as the deployment configuration.","The 0.64 ms inference leaves time within each 1 ms window that could in principle be used for additional on-device computations, such as the quality-of-service estimation the authors say they are currently investigating."],"supporting_citations":[{"why":"Supplies the original convolutional-neural-network approach to self-mixing displacement sensing, the experimental optical setup, and the training/validation methodology that this paper inherits and improves.","marker":"[17]"},{"why":"Provides the training and generalization datasets used for all experiments, including the periodic training subset and the aperiodic out-of-distribution subset with speaker-calibrated ground truth.","marker":"[26]"},{"why":"The ResNet architecture on which the improved embedded model is based, replacing the earlier CNN for higher efficiency.","marker":"[25]"},{"why":"Defines the post-training quantization method used here: power-of-two scale factor, symmetric range, nearest rounding, and the Q7.9 fixed-point format.","marker":"[23]"},{"why":"The Qualia framework that generates the portable C inference library from the trained quantized model for microcontroller deployment.","marker":"[22]"},{"why":"The CMSIS-NN library whose 16-bit SIMD-optimized convolutions bring inference latency below the 1 ms real-time budget.","marker":"[28]"},{"why":"The authors' earlier prototype and CNN deployment, which serves as the baseline for the embedded-platform and energy-consumption comparisons.","marker":"[24]"}],"fun_headline_variants":["Laser sensor embeds ResNet for real-time displacement tracking","Neural network on STM32 reconstructs motion from laser self-mixing","Edge AI: 7k-parameter ResNet runs laser interferometry in 0.64 ms","Low-power laser sensor with embedded neural net hits kHz rate","16-bit quantized ResNet on STM32 tracks laser target in real time"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The ground-truth displacement labels in both training and evaluation are derived from the voltage applied to a calibrated speaker, assuming a linear $\\lambda/V$ scaling factor accurate to within 10% over the operating frequency band; if that proxy is biased or nonlinear, the network learns and the reported accuracy numbers inherit the bias.","fun_headline_variants_meta":{"raw":{"variants":["Laser sensor embeds ResNet for real-time displacement tracking","Neural network on STM32 reconstructs motion from laser self-mixing","Edge AI: 7k-parameter ResNet runs laser interferometry in 0.64 ms","Low-power laser sensor with embedded neural net hits kHz rate","16-bit quantized ResNet on STM32 tracks laser target in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000803,"raw_usage":{"total_tokens":3389,"prompt_tokens":791,"completion_tokens":2598,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":535,"completion_tokens_details":{"reasoning_tokens":2498}},"tokens_in":535,"tokens_out":2598,"duration_ms":19589,"temperature":1.0,"reasoning_tokens":2498,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T18:57:13.664834+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace the calibrated speaker with an independent displacement reference, such as a commercial laser interferometer or a capacitive sensor measuring the same target, and compare the network's inferred displacement to that reference across the full operating envelope (5–100 Hz, below 2.5 $\\lambda$/ms, feedback parameter C near unity); any systematic disagreement larger than the reported mean absolute error of roughly 0.35 $\\lambda$/ms would indicate that the speaker-voltage proxy is contaminating the accuracy claims.","supporting_citations":[{"cited_title":"Convolutional neural network for self- mixing interferometric displacement sensing,","cited_arxiv_id":null,"evidence_quote":"Supplies the original convolutional-neural-network approach to self-mixing displacement sensing, the experimental optical setup, and the training/validation methodology that this paper inherits and improves."},{"cited_title":"Displacement measurement via self mixing interferometry and neural network training set,","cited_arxiv_id":null,"evidence_quote":"Provides the training and generalization datasets used for all experiments, including the periodic training subset and the aperiodic out-of-distribution subset with speaker-calibrated ground truth."},{"cited_title":"Quantization and deployment of deep neural networks on microcontrollers,","cited_arxiv_id":null,"evidence_quote":"Defines the post-training quantization method used here: power-of-two scale factor, symmetric range, nearest rounding, and the Q7.9 fixed-point format."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Qualia framework that generates the portable C inference library from the trained quantized model for microcontroller deployment."},{"cited_title":"Enabling deep learning at the iot edge,","cited_arxiv_id":null,"evidence_quote":"The CMSIS-NN library whose 16-bit SIMD-optimized convolutions bring inference latency below the 1 ms real-time budget."},{"cited_title":"Integrating embedded neural networks and self-mixing interferometry for smart sensors design,","cited_arxiv_id":null,"evidence_quote":"The authors' earlier prototype and CNN deployment, which serves as the baseline for the embedded-platform and energy-consumption comparisons."}],"review_version":1}