pith. sign in

arxiv: 2606.19574 · v1 · pith:TALVRBVYnew · submitted 2026-06-17 · 📡 eess.IV · cs.CV

FrequencyFormer: A Co-Designed Sensor-to-Processor Pipeline for Frequency-Domain Vision Transformer Inference

Chengwei Zhou , Ovishake Sen , Xuming Chen , Rishith Paramasivam , Shaahin Angizi , Swarup Bhunia , Baibhab Chatterjee , Gourav Datta This is my paper

Pith reviewed 2026-06-26 18:45 UTC · model grok-4.3

classification 📡 eess.IV cs.CV
keywords vision transformerfrequency domainDCT tokenizerin-sensor computingenergy efficiencyedge AIdata compressionsensor pipeline
0
0 comments X

The pith

Frequency-domain tokenization at the sensor replaces patch embedding to cut ViT data transfer by 128x while staying compatible with pretrained models.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that performing tokenization in the frequency domain directly at the sensor can address the main energy and bandwidth limits when running vision transformers on edge devices. A multi-scale DCT tokenizer compresses 224x224 images into far fewer tokens, cutting off-chip data movement by up to 128 times with only modest accuracy impact. This tokenization serves as a direct substitute for the usual spatial patch embedding, so the rest of a pretrained ViT can be used unchanged across classification, detection, and segmentation. The full pipeline adds low-power hardware for the transform and an improved communication link, yielding large measured gains in efficiency and energy.

Core claim

FrequencyFormer presents a sensor-to-processor pipeline in which a multi-scale DCT tokenizer converts input images into compact frequency-domain tokens that replace standard ViT patch embeddings. The tokenizer achieves up to 128x reduction in off-chip data volume and remains compatible with pretrained backbones for multiple vision tasks. A LUT-based near-sensor implementation performs the transform without multipliers, and a modified MIPI link further lowers transfer costs. The resulting system reports 28.8 TOPS/W, 230x lower communication energy, and 2.22x lower total sensor-side energy.

What carries the argument

The multi-scale DCT tokenizer that converts spatial images into compact frequency-domain tokens for direct use as ViT input.

If this is right

  • The pipeline reaches 28.8 TOPS/W efficiency.
  • Communication energy falls by 230 times compared with baseline transfer.
  • Total sensor-side energy drops by a factor of 2.22.
  • Off-chip data volume is reduced by up to 128 times with modest accuracy loss.
  • The tokenizer works as a drop-in replacement across classification, detection, and segmentation without model changes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Frequency-domain compression at the sensor could extend to other transformer-based models or non-image sensors where spectral representations are natural.
  • The multiplier-free LUT approach for DCT may suggest similar fixed-coefficient designs for other early-processing steps in edge accelerators.
  • If accuracy holds across more tasks, the method could support always-on vision applications on severely power-constrained devices.
  • The data-volume savings might combine with other compression techniques to further scale ViT deployment on bandwidth-limited links.

Load-bearing premise

That multi-scale DCT tokens can replace standard spatial patch embeddings in pretrained vision transformers without causing large accuracy losses or requiring retraining on classification, detection, or segmentation tasks.

What would settle it

Measuring accuracy on ImageNet with a pretrained ViT using the DCT tokenizer instead of patch embedding and finding a drop larger than a few percentage points, or hardware tests failing to show the claimed 230x communication energy reduction.

Figures

Figures reproduced from arXiv: 2606.19574 by Baibhab Chatterjee, Chengwei Zhou, Gourav Datta, Ovishake Sen, Rishith Paramasivam, Shaahin Angizi, Swarup Bhunia, Xuming Chen.

Figure 1
Figure 1. Figure 1: (a) Energy breakdown for edge vision backbones on 360×360 frames shows that as models are optimized (ResNet-50 FP32 → MobileNet￾v2 INT8/INT4 → EfficientFormer-L1 INT4), compute energy decreases while communication (MIPI) remains constant, increasing its share of total energy. Estimates use ASIC-level models including compute and memory. (b) Pro￾posed 3D-integrated pipeline with a stacked CMOS sensor and ne… view at source ↗
Figure 3
Figure 3. Figure 3: 8×8 Selection-Aware DCT Pruning with Harmonic-Aware Quanti￾zation. C8 is pruned to 5 rows and C8 ⊤ to 4 columns needed by the 𝐾=14 zigzag coefficients, reducing to a 5×4 output tile. Bit-widths decrease from INT8 to INT4 per row/column; the intermediate result CprunedX is requantized before the column-harmonic multiplication. The resulting 5 × 4 coefficient matrix Y, where each cell shows its effective pre… view at source ↗
Figure 4
Figure 4. Figure 4: (a) MIPI-CSI example; (b) Tx and Rx Architecture in MIPI [40] ; (c) Conventional Implementation of the Physical layer of the Tx/Rx, which is implemented with a simple inverter-based driver in the Tx and with a dynamic comparator at the Rx; (d) Proposed/Modified Implementation of the Physical layer of the Tx/Rx: the Tx remain unchanged, while the dynamic comparator at the Rx is preseded by a dynamic integra… view at source ↗
Figure 5
Figure 5. Figure 5: (a) Area overhead for uniform INT9 tokenizer multiplication showing our LUT implementation significantly reduces area compared to Traditional LUT (≈29.54×) and Array Multiplier (≈1.24×). (b) Energy per in￾ference comparison, where our LUT achieves up to ≈3.34× and ≈1.8× energy savings over Traditional LUT and Wallace Tree, respectively, demonstrating improved efficiency. Multi-scale branches [PITH_FULL_IM… view at source ↗
Figure 7
Figure 7. Figure 7: (a) 65nm test-chip micrograph for the communication PHY; (b) Power consumption at the Tx and Rx with IR (integrator+comparator) and without IR (comparator only). we demonstrate measured results from a 65nm test chip for same SNR at the input of the comparator. At 30 fps frame rate (≈36 Mbps raw data rate, which reduces to <360 kbps through the Frequen￾cyFormer processing pipeline), we have ≈ 2.77/2 = 1.38 … view at source ↗
read the original abstract

Deploying vision transformers (ViTs) on sensor-edge systems is limited not only by on-device compute, but also by the energy and bandwidth required to transmit high-dimensional image data from the sensor to the processor. While in-sensor and near-sensor computing reduce this cost through early feature extraction, existing methods often provide only modest compression. We observe that the frequency domain provides a naturally compact representation of visual information and can be exploited at the sensor level to reduce sensor-to-processor data movement. Building on this insight, we present FrequencyFormer, a co-designed sensor-to-processor pipeline for efficient ViT inference. FrequencyFormer includes: (1) a multi-scale DCT tokenizer that compresses a 224x224 image into compact frequency-domain tokens, achieving up to 128x reduction in off-chip data volume with modest accuracy loss; (2) a LUT-based near-sensor hardware implementation that leverages fixed DCT coefficients for multiplier-free, energy- and area-efficient tokenization; and (3) a modified MIPI-based low-power communication architecture that further reduces transfer energy. FrequencyFormer serves as a drop-in replacement for standard ViT patch embedding and remains compatible with pretrained backbones across classification, detection, and segmentation tasks. The pipeline achieves 28.8 TOPS/W, reduces communication energy by 230x, and lowers total sensor-side energy by 2.22x, demonstrating frequency-domain tokenization as a scalable foundation for in-sensor ViT deployment.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents FrequencyFormer, a co-designed sensor-to-processor pipeline for ViT inference. It features a multi-scale DCT tokenizer that compresses 224x224 images into frequency-domain tokens for up to 128x off-chip data reduction, a LUT-based near-sensor hardware implementation for multiplier-free tokenization, and a modified MIPI low-power communication architecture. The pipeline is claimed to serve as a drop-in replacement for standard ViT patch embedding, remaining compatible with pretrained backbones on classification, detection, and segmentation tasks, while achieving 28.8 TOPS/W, 230x communication energy reduction, and 2.22x total sensor-side energy reduction.

Significance. If the quantitative efficiency claims and the drop-in compatibility hold with supporting experiments, the work could provide a practical foundation for reducing sensor-to-processor data movement in edge ViT deployments by exploiting frequency-domain compactness, with potential applicability across multiple vision tasks and hardware constraints.

major comments (2)
  1. [Abstract] Abstract: The specific quantitative performance claims (128x data reduction, 230x communication energy reduction, 2.22x total energy reduction, 28.8 TOPS/W) are stated without any description of experimental methodology, baselines, datasets, error bars, or measurement setup, which is load-bearing for assessing whether the energy and accuracy results are reproducible or derived from the proposed pipeline.
  2. [Abstract] Abstract: The central assertion that the multi-scale DCT tokenizer acts as a drop-in replacement for standard patch embedding and remains compatible with pretrained ViT backbones (across classification, detection, and segmentation) without substantial accuracy loss or retraining lacks isolated validation experiments; the input distribution shift from learned spatial embeddings to fixed frequency coefficients could undermine the no-retraining claim that underpins the broad applicability of the 128x reduction and energy savings.
minor comments (1)
  1. [Abstract] The abstract would benefit from a brief statement of the evaluation tasks, datasets, and accuracy metrics used to support the 'modest accuracy loss' claim.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for highlighting these points on the abstract. The quantitative claims are supported by detailed experiments in the body of the manuscript, but we agree the abstract can be strengthened for clarity. We also recognize the value of more isolated validation for the drop-in replacement property and will incorporate additional experiments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The specific quantitative performance claims (128x data reduction, 230x communication energy reduction, 2.22x total energy reduction, 28.8 TOPS/W) are stated without any description of experimental methodology, baselines, datasets, error bars, or measurement setup, which is load-bearing for assessing whether the energy and accuracy results are reproducible or derived from the proposed pipeline.

    Authors: We agree the abstract's brevity omits key methodological details. The full manuscript (Sections 3.3, 4.1, and 4.4) specifies the datasets (ImageNet-1K, COCO, ADE20K), baselines (standard ViT patch embedding and prior near-sensor methods), hardware synthesis for TOPS/W, MIPI energy models, and accuracy metrics without retraining. Error bars from multiple runs are reported in the results tables. To address the concern directly in the abstract, we will add a concise clause referencing the evaluation protocol and datasets while respecting length constraints. revision: partial

  2. Referee: [Abstract] Abstract: The central assertion that the multi-scale DCT tokenizer acts as a drop-in replacement for standard patch embedding and remains compatible with pretrained ViT backbones (across classification, detection, and segmentation) without substantial accuracy loss or retraining lacks isolated validation experiments; the input distribution shift from learned spatial embeddings to fixed frequency coefficients could undermine the no-retraining claim that underpins the broad applicability of the 128x reduction and energy savings.

    Authors: The manuscript evaluates the tokenizer on pretrained backbones without retraining across tasks, reporting accuracy within 1-3% of the spatial baseline. However, we concur that dedicated isolation of the distribution shift would strengthen the claim. We will add an ablation experiment that freezes the backbone, swaps only the embedding layer (standard vs. multi-scale DCT), and measures accuracy delta on the same pretrained weights. This will be placed in Section 4.2 of the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; claims are empirical outcomes of co-design

full rationale

The paper presents FrequencyFormer as a hardware-software co-design using a multi-scale DCT tokenizer with fixed coefficients, LUT-based near-sensor implementation, and modified MIPI communication. All reported metrics (28.8 TOPS/W, 230x communication energy reduction, 2.22x sensor-side energy reduction, 128x data volume reduction) are stated as measured results of the pipeline rather than derived predictions. No equations, parameter-fitting steps, self-citations, or uniqueness theorems appear in the abstract or described claims that would reduce any result to its own inputs by construction. The drop-in compatibility assertion is presented as a design property verified through task experiments, with no visible reduction to fitted inputs or circular self-reference. The work is self-contained as an empirical engineering demonstration.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The abstract contains no mathematical derivations, parameter fittings, or postulated entities; no free parameters, axioms, or invented entities can be identified.

pith-pipeline@v0.9.1-grok · 5828 in / 1269 out tokens · 36746 ms · 2026-06-26T18:45:34.746704+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 14 canonical work pages · 7 internal anchors

  1. [1]

    Minhaz Abedin et al . 2022. Mr-pipa: An integrated multilevel rram (hfo x)- based processing-in-pixel accelerator.IEEE Journal on Exploratory Solid-State Computational Devices and Circuits8, 2 (2022), 59–67

    2022

  2. [2]

    Ahalapitiya Adibi, Muhammad Firdausi, et al . 2022. In-sensor image memo- rization and encoding via optical neurons for neuromorphic sensing.Nature Communications13, 1 (2022), 6812

    2022

  3. [3]

    Simoncelli

    Johannes Ballé, Valero Laparra, and Eero P. Simoncelli. 2017. End-to-end opti- mized image compression.arXiv preprint arXiv:1611.01704(2017)

    work page arXiv 2017

  4. [4]

    Johannes Ballé, David Minnen, Saurabh Singh, Sung Jin Hwang, and Nick John- ston. 2018. Variational image compression with a scale hyperprior.arXiv preprint arXiv:1802.01436(2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  5. [5]

    Daniel Bolya et al. 2023. Token merging: Your ViT but faster. InInternational Conference on Learning Representations

    2023

  6. [6]

    Laurie Bose et al. 2020. Fully embedding fast convolutional networks on pixel processor arrays. InEuropean Conference on Computer Vision. Springer, 488–503

    2020

  7. [7]

    Ningyuan Cao et al. 2022. A 65 nm Wireless Image SoC Supporting On-Chip DNN Optimization and Real-Time Computation-Communication Trade-Off via Actor-Critical Neuro-Controller.IEEE Journal of Solid-State Circuits57, 8 (2022), 2545–2559. https://doi.org/10.1109/JSSC.2022.3159473

    work page doi:10.1109/jssc.2022.3159473 2022

  8. [8]

    Yang Chai. 2023. In-sensor visual perception and inference.Intelligent Computing 2 (2023)

    2023

  9. [9]

    Denis Guangyin Chen et al. 2014. A 12 pJ/Pixel Analog-to-Information Converter Based 816×640 Pixel CMOS Image Sensor.IEEE Journal of Solid-State Circuits 49, 5 (2014), 1210–1222. https://doi.org/10.1109/JSSC.2014.2307063

    work page doi:10.1109/jssc.2014.2307063 2014

  10. [10]

    Wenlin Chen et al. 2016. Packing convolutional neural networks in the frequency domain. InNeurIPS Workshop

    2016

  11. [11]

    Aakanksha Chowdhery, Pete Warden, Jonathon Shlens, Andrew Howard, and Rocky Rhodes. 2019. Visual wake words dataset.arXiv preprint arXiv:1906.05721 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  12. [12]

    Gourav Datta et al. 2022. A processing-in-pixel-in-memory paradigm for resource- constrained tinyml applications.Scientific Reports12, 1 (2022), 14396

    2022

  13. [13]

    Peyman Dehghanzadeh, Ovishake Sen, Baibhab Chatterjee, and Swarup Bhunia

  14. [14]

    Comput.74, 4 (2025), 1348–1361

    LUNA-CiM: A Programmable Compute-in-Memory Fabric for Neural Network Acceleration.IEEE Trans. Comput.74, 4 (2025), 1348–1361

    2025

  15. [15]

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xi- aohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. 2020. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  16. [16]

    Max Ehrlich and Larry S. Davis. 2019. Deep residual learning in the JPEG transform domain.arXiv preprint arXiv:1901.07758(2019)

    work page arXiv 2019

  17. [17]

    Lionel Gueguen, Alex Sergeev, Ben Kadlec, Rosanne Liu, and Jason Yosinski. 2018. Faster neural networks straight from JPEG. InAdvances in Neural Information Processing Systems, Vol. 31

    2018

  18. [18]

    Song Han, Huizi Mao, and William J. Dally. 2016. Deep compression: Compressing deep neural networks with pruning, trained quantization and Huffman coding. arXiv preprint arXiv:1510.00149(2016)

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  19. [19]

    Lucas Hansen. 2015. Tiny ImageNet challenge submission.CS 231N(2015)

    2015

  20. [20]

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. 2017. Mask r-cnn. InProceedings of the IEEE international conference on computer vision. 2961–2969

    2017

  21. [21]

    Ibraheem, Mokhtar M

    Noor A. Ibraheem, Mokhtar M. Hasan, Rafiqul Z. Khan, and Pramod K. Mishra

  22. [22]

    Understanding color models: A review.ARPN Journal of Science and Technology2, 3 (2012), 265–275

    2012

  23. [23]

    Benoit Jacob et al. 2018. Quantization and training of neural networks for efficient integer-arithmetic-only inference. InIEEE Conference on Computer Vision and Pattern Recognition. 2704–2713

    2018

  24. [24]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparameterization with Gumbel-softmax.arXiv preprint arXiv:1611.01144(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  25. [25]

    Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. [n. d.]. CIFAR-10 (Canadian Institute for Advanced Research). ([n. d.]). http://www.cs.toronto.edu/~kriz/cifar. html

  26. [26]

    Seonghyeon Lee, Beomsik Kang, and Chang D. Yoo. 2024. DCT-ViT: High- frequency pruned vision transformer with discrete cosine transform.IEEE Access 12 (2024), 80386–80396

    2024

  27. [27]

    James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, and Santiago Ontanon. 2022. FNet: Mixing tokens with Fourier transforms. InConference of the North American Chapter of the Association for Computational Linguistics. 4296–4313

    2022

  28. [28]

    Xinyu Li, Yanyi Zhang, Jianbo Yuan, Hanlin Lu, and Yushi Zhu. 2023. Discrete Cosin TransFormer: Image modeling from frequency domain. InIEEE/CVF Winter Conference on Applications of Computer Vision. 3169–3178

    2023

  29. [29]

    Yanyu Li et al. 2022. Efficientformer: Vision transformers at mobilenet speed. Advances in neural information processing systems35 (2022), 12934–12949

    2022

  30. [30]

    Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. 2022. Exploring plain vision transformer backbones for object detection. InEuropean conference on computer vision. Springer, 280–296

    2022

  31. [31]

    Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. 2021. Pruning and quantization for deep neural network acceleration: A survey.Neu- rocomputing461 (2021), 370–403

    2021

  32. [32]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. 2014. Microsoft COCO: Common objects in context. InEuropean Conference on Computer Vision. Springer, 740–755

    2014

  33. [33]

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. 2014. Microsoft coco: Common objects in context. InEuropean conference on computer vision. Springer, 740–755

    2014

  34. [34]

    Min Liu, Yifan He, and Hailong Jiao. 2022. An LUT-based multiplier array for systolic array-based convolutional neural network accelerator. In2022 IEEE Asia Pacific Conference on Circuits and Systems (APCCAS). IEEE, 55–59

    2022

  35. [35]

    Ze Liu et al. 2021. Swin transformer: Hierarchical vision transformer using shifted windows. InProceedings of the IEEE/CVF international conference on computer vision. 10012–10022

    2021

  36. [36]

    Ilya Loshchilov and Frank Hutter. 2017. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101(2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  37. [37]

    Shovan Maity et al. 2019. BodyWire: A 6.3-pJ/b 30-Mb/s -30-dB SIR-Tolerant Broadband Interference-Robust Human Body Communication Transceiver Using Time Domain Interference Rejection.IEEE Journal of Solid-State Circuits54, 10 (2019), 2892–2906. https://doi.org/10.1109/JSSC.2019.2932852

    work page doi:10.1109/jssc.2019.2932852 2019

  38. [38]

    Lukas Mennel et al . 2020. Ultrafast machine vision with 2D material neural network image sensors.Nature579, 7797 (2020), 62–66

    2020

  39. [39]

    Chihiro Okada et al. 2021. 7.6 A High-Speed Back-Illuminated Stacked CMOS Image Sensor with Column-Parallel kT/C-Cancelling S&H and Delta-Sigma ADC. In2021 IEEE International Solid-State Circuits Conference (ISSCC), Vol. 64. 116–118. https://doi.org/10.1109/ISSCC42613.2021.9366024

    work page doi:10.1109/isscc42613.2021.9366024 2021

  40. [40]

    Hongyi Pan et al. 2024. Discrete cosine transform based decorrelated attention for vision transformers.arXiv preprint arXiv:2405.13901(2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    Badri Narayana Patro et al. 2025. SpectFormer: Frequency and attention is what you need in a vision transformer. InIEEE/CVF Winter Conference on Applications of Computer Vision. 3745–3755

    2025

  42. [42]

    Prabu Kumar, CTO, e-con Systems (2026)). [n. d.]. What is a MIPI Cam- era? How does MIPI Camera Work? [Online]. Available: https://www.e- consystems.com/blog/camera/technology/what-is-a-mipi-camera-how-does- mipi-camera-work/. Accessed: 04/13/2026

    2026

  43. [43]

    Zequn Qin et al. 2021. FcaNet: Frequency channel attention networks. InIEEE/CVF International Conference on Computer Vision. 783–792

    2021

  44. [44]

    Akshay Krishna Ramanathan et al. 2020. Look-up table based energy efficient processing in cache support for neural network acceleration. In2020 53rd Annual IEEE/ACM International Symposium on Microarchitecture (MICRO). IEEE, 88–101

    2020

  45. [45]

    Yongming Rao et al. 2021. DynamicViT: Efficient vision transformers with dy- namic token sparsification. InAdvances in Neural Information Processing Systems, Vol. 34. 13937–13949

    2021

  46. [46]

    Behzad Razavi. 2015. The StrongARM Latch [A Circuit for All Seasons].IEEE Solid-State Circuits Magazine7 (2015), 12–17. https://api.semanticscholar.org/ CorpusID:9477992

    2015

  47. [47]

    Arman Roohi et al . 2023. Pipsim: A behavior-level modeling tool for CNN processing-in-pixel accelerators.IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems43, 1 (2023), 141–150

    2023

  48. [48]

    Ovishake Sen et al. 2025. Look-Up Table based Energy-Efficient Architecture for Neural Accelerators (LANA).IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems(2025)

    2025

  49. [49]

    Hugo Touvron et al. 2021. Training data-efficient image transformers & distilla- tion through attention. InInternational Conference on Machine Learning. PMLR, 10347–10357

    2021

  50. [50]

    Karlo Vasquez, Yeshwanth Venkatesha, Anurag Bhattacharjee, Abhiroop Moitra, and Priyadarshini Panda. 2021. Activation density based mixed-precision quan- tization for energy efficient neural networks. InDesign, Automation & Test in Europe Conference. 1360–1365

    2021

  51. [51]

    Gomez, Łukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. InAdvances in Neural Information Processing Systems, Vol. 30

    2017

  52. [52]

    Gregory K. Wallace. 1992. The JPEG still picture compression standard.IEEE Transactions on Consumer Electronics38, 1 (1992), xviii–xxxiv

    1992

  53. [53]

    Xiao Wang et al. 2024. Integrated photonic encoder for low power and high- speed image processing.Nature Communications15, 1 (2024), 4510. https: //doi.org/10.1038/s41467-024-48099-2

    work page doi:10.1038/s41467-024-48099-2 2024

  54. [54]

    Han Xu et al. 2020. Utilizing direct photocurrent computation and 2D kernel scheduling to improve in-sensor-processing efficiency. In2020 57th ACM/IEEE Design Automation Conference (DAC). IEEE, 1–6

    2020

  55. [55]

    Han Xu et al. 2021. Senputing: An ultra-low-power always-on vision perception chip featuring the deep fusion of sensing and computing.IEEE Transactions on Circuits and Systems I: Regular Papers69, 1 (2021), 232–243

    2021

  56. [56]

    Kai Xu et al. 2020. Learning in the frequency domain. InIEEE/CVF Conference on Computer Vision and Pattern Recognition. 1740–1749

    2020

  57. [57]

    Fei Zhou and Yang Chai. 2020. Near-sensor and in-sensor computing.Nature Electronics3, 11 (2020), 664–671. 9

    2020