pith. sign in

arxiv: 2607.02465 · v1 · pith:X3GRTMN4new · submitted 2026-07-02 · 💻 cs.AR

Probabilistic Memory for Trustworthy Edge Intelligence

Likai Pei , Jiahao Zheng , Xueji Zhao , Emilie Ye , Jianbo Liu , Hanqing Tao , Ming-Yen Lee , Ruiyang Qin
show 4 more authors
Yiyu Shi Shimeng Yu X. Sharon Hu Ningyuan Cao
This is my paper

Pith reviewed 2026-07-03 03:32 UTC · model grok-4.3

classification 💻 cs.AR
keywords probabilistic memoryGaussian random number generationedge intelligenceBayesian neural networksmemory architecturehardware accelerationtrustworthy AIprobabilistic computation
0
0 comments X

The pith

Probabilistic memory stores distribution parameters and samples Gaussians directly at native memory bandwidth.

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

The paper introduces p-MEM as a memory primitive that holds parameters such as mean and standard deviation instead of fixed values, then generates samples on demand at full memory speed. Deterministic storage appears as the special case of zero variance. Simulations across device types and nodes show GRNG throughput above 1000 GSa/s/mm² when including array access. When p-MEM is added to CPU and GPU systems, Bayesian neural network workloads see instruction counts drop by up to 2.19x and 4.37x, sampling latency by 562x and 3.45x, and energy by 295.5x and 3.53x. A reader would care because the approach removes the main hardware barrier to uncertainty-aware, robust, and private computation on edge devices.

Core claim

p-MEM is a unified memory primitive that stores distribution parameters and samples directly at the native memory bandwidth, where deterministic data becomes the zero-variance special case. Using a layout-validated simulator, p-MEM achieves more than 1000 GSa/s/mm² GRNG throughput including memory-array access. Integrated into CPU/GPU systems it reduces instruction count by up to 2.19x/4.37x, sampling latency by 562x/3.45x, and energy by 295.5x/3.53x for Bayesian neural network workloads.

What carries the argument

p-MEM, a memory array that stores mean and standard deviation parameters and performs on-array Gaussian sampling at native bandwidth.

Load-bearing premise

The layout-validated simulator accurately captures real silicon behavior for the chosen device technologies, memory specifications, and technology nodes without unmodeled parasitics or fabrication variations.

What would settle it

Fabricate a p-MEM test chip in one of the simulated technology nodes and measure its actual GRNG throughput, latency, and energy against the simulator predictions under identical workload conditions.

Figures

Figures reproduced from arXiv: 2607.02465 by Emilie Ye, Hanqing Tao, Jiahao Zheng, Jianbo Liu, Likai Pei, Ming-Yen Lee, Ningyuan Cao, Ruiyang Qin, Shimeng Yu, X. Sharon Hu, Xueji Zhao, Yiyu Shi.

Figure 1
Figure 1. Figure 1: (a) Motivation for probabilistic memory to support prob [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Compiled-instruction analysis on RISC-V CPU and NVIDIA [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overall architecture of the proposed p-MEM. (a) Weight decomposition and hybrid deterministic / probabilistic data storage. (b) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Framework overview of proposed p-MEM. 22nm. In contrast, RRAM and FeFET models are derived from fabri￾cated data provided in the NeuroSim framework. The technology file defines electrical parameters, the config file specifies 𝜇 and 𝜎 bit precision and system-level settings, and the parameter file contains customizable physical attributes such as capacitance and resistance. RNG specifications follow a fixed… view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of pSRAM-D, pSRAM-A, and conventional [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 8
Figure 8. Figure 8: Bandwidth per area comparison across different array sizes [PITH_FULL_IMAGE:figures/full_fig_p005_8.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of SRAM, RRAM, and FeRAM in area, latency, [PITH_FULL_IMAGE:figures/full_fig_p005_7.png] view at source ↗
read the original abstract

Probabilistic computation plays an important role in trustworthy edge intelligence to quantify uncertainty, enhance robustness, reconstruct data, and protect privacy, but its adoption is limited by the orders-of-magnitude data throughput gap between Gaussian random number generation (GRNG) and computation, as well as instruction overhead. This paper introduces probabilistic memory (p-MEM), a unified memory primitive that stores distribution parameters, such as mean and standard deviation, and samples directly at the native memory bandwidth, where deterministic data becomes the zero-variance special case. Using a layout-validated p-MEM simulator, we comprehensively explore device choices, memory specifications, and technology nodes, showing that p-MEM can achieve more than 1000 GSa/s/mm^2 GRNG throughput, including memory-array access. Integrated into CPU/GPU systems, p-MEM reduces instruction count by up to 2.19x/4.37x, sampling latency by 562x/3.45x, and energy by 295.5x/3.53x for Bayesian neural network workloads, providing a scalable hardware substrate for trustworthy probabilistic AI.

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

1 major / 1 minor

Summary. The paper introduces probabilistic memory (p-MEM) as a unified memory primitive that stores distribution parameters (e.g., mean and standard deviation) and performs direct sampling at native memory bandwidth, treating deterministic data as the zero-variance case. Using a layout-validated p-MEM simulator, it explores device choices, memory specifications, and technology nodes to claim >1000 GSa/s/mm² GRNG throughput (including array access). When integrated into CPU/GPU systems, p-MEM yields up to 2.19×/4.37× instruction count reduction, 562×/3.45× sampling latency reduction, and 295.5×/3.53× energy reduction for Bayesian neural network workloads.

Significance. If the simulator results translate to silicon, p-MEM would offer a scalable hardware substrate for trustworthy probabilistic edge AI by closing the GRNG throughput gap. A strength is the comprehensive parameter exploration across devices, specs, and nodes using layout validation; no machine-checked proofs or parameter-free derivations are present.

major comments (1)
  1. [p-MEM simulator and integration results sections] The headline claims (>1000 GSa/s/mm² throughput and the 2.19–562× integrated speedups) are generated exclusively by the authors' layout-validated simulator. Layout validation confirms geometric/electrical correctness but does not capture process variation, mismatch, temperature-dependent leakage, or interconnect parasitics that appear post-fabrication; this assumption is load-bearing for all quantitative results.
minor comments (1)
  1. [Abstract] Abstract does not include error bars, sensitivity analysis on device parameters, or direct silicon baseline comparisons for the largest speedups.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for highlighting the scope of our simulator validation. We agree that layout validation provides geometric and electrical correctness but does not model post-fabrication effects, and we will revise the manuscript to explicitly address this limitation and its implications for the quantitative claims.

read point-by-point responses

  1. Referee: [p-MEM simulator and integration results sections] The headline claims (>1000 GSa/s/mm² throughput and the 2.19–562× integrated speedups) are generated exclusively by the authors' layout-validated simulator. Layout validation confirms geometric/electrical correctness but does not capture process variation, mismatch, temperature-dependent leakage, or interconnect parasitics that appear post-fabrication; this assumption is load-bearing for all quantitative results.

    Authors: We acknowledge that the reported throughput and speedup figures are derived from a layout-validated simulator that ensures geometric and electrical fidelity at the cell and array level but omits post-silicon phenomena including process variation, device mismatch, temperature-dependent leakage, and interconnect parasitics. These omissions mean the numbers represent idealized pre-fabrication performance. In the revised manuscript we will (1) add a new subsection under Evaluation that enumerates these assumptions, (2) qualify all headline metrics as upper-bound estimates under nominal conditions, and (3) include a brief sensitivity discussion based on published variation models for the explored device technologies. This revision directly addresses the concern while preserving the value of the comprehensive pre-silicon design-space exploration. revision: yes

Circularity Check

0 steps flagged

No circularity: performance claims rest on simulator outputs, not self-referential equations or citations

full rationale

The paper's headline throughput and speedup numbers are produced by running the authors' layout-validated p-MEM simulator across device choices and nodes. No equations, fitted parameters, or self-citations are shown that would make any claimed quantity equivalent to its own inputs by construction. The work does not invoke prior self-authored uniqueness theorems, smuggle ansatzes, or rename known results as new derivations. The simulator is treated as an external evaluation tool rather than a definitional loop, satisfying the criteria for a self-contained (non-circular) result.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The central claims rest on the accuracy of the authors' simulator for multiple technology nodes and the assumption that the reported speedups translate to real systems; no free parameters are explicitly fitted in the abstract, but device-level parameters are explored.

axioms (1)
  • domain assumption The simulator's device models and layout validation faithfully represent fabricated silicon behavior across the explored technology nodes.
    Invoked when the abstract states performance numbers are obtained from the layout-validated simulator.
invented entities (1)
  • probabilistic memory (p-MEM) no independent evidence
    purpose: A memory primitive that stores distribution parameters and performs sampling at native bandwidth.
    The new hardware abstraction introduced to close the GRNG throughput gap.

pith-pipeline@v0.9.1-grok · 5755 in / 1343 out tokens · 27682 ms · 2026-07-03T03:32:59.279367+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

40 extracted references

  1. [1]

    A Survey on Trustworthy Edge Intelligence: From Security and Reliability to Transparency and Sustainability.IEEE Communications Surveys & Tutorials, 27(3):1729–1757, June 2025

    Xiaojie Wang, Beibei Wang, Yu Wu, Zhaolong Ning, Song Guo, and Fei Richard Yu. A Survey on Trustworthy Edge Intelligence: From Security and Reliability to Transparency and Sustainability.IEEE Communications Surveys & Tutorials, 27(3):1729–1757, June 2025

    2025

  2. [2]

    Enciso, Boyang Cheng, Jianbo Liu, Steven Davis, Zhenge Jia, Michael Niemier, Yiyu Shi, X

    Likai Pei, Yifan Qin, Zephan M. Enciso, Boyang Cheng, Jianbo Liu, Steven Davis, Zhenge Jia, Michael Niemier, Yiyu Shi, X. Sharon Hu, and Ningyuan Cao. To- wards uncertainty-quantifiable biomedical intelligence: Mixed-signal compute-in- entropy for bayesian neural networks. In2024 IEEE/ACM International Conference On Computer Aided Design (ICCAD), 2024

    2024

  3. [3]

    Privacy-by-Sensing with Time-domain Differentially-Private Compressed Sensing

    Jianbo Liu, Boyang Cheng, Pengyu Zeng, Steven Davis, Muya Chang, and Ningyuan Cao. Privacy-by-Sensing with Time-domain Differentially-Private Compressed Sensing. In2023 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1–6, April 2023

    2023

  4. [4]

    15.3 A 65nm Uncertainty-Quantifiable Ventricular Arrhythmia Detection Engine with \mathbf1.75\boldsymbol\mu\mathbfJ Per Inference

    Jianbo Liu, Zephan Enciso, Boyang Cheng, Likai Pei, Steven Davis, Yifan Qin, Zhenge Jia, Xiaobo Sharon Hu, Yiyu Shi, and Ningyuan Cao. 15.3 A 65nm Uncertainty-Quantifiable Ventricular Arrhythmia Detection Engine with \mathbf1.75\boldsymbol\mu\mathbfJ Per Inference. In2025 IEEE International Solid-State Circuits Conference (ISSCC), volume 68, pages 1–3, Fe...

    2025

  5. [5]

    Ahmed Abdelrazik, Mahmoud Eldesouky, Ibrahim Antoun, Edward Y. M. Lau, Abdulmalik Koya, Zakariyya Vali, Safiyyah A. Suleman, James Donaldson, and G. André Ng. Wearable Devices for Arrhythmia Detection: Advancements and Clinical Implications.Sensors (Basel, Switzerland), 25(9):2848, April 2025

    2025

  6. [6]

    Metwally, Dalia Perelman, Heyjun Park, Yue Wu, Alokkumar Jha, Seth Sharp, Alessandra Celli, Ekrem Ayhan, Fahim Abbasi, Anna L

    Ahmed A. Metwally, Dalia Perelman, Heyjun Park, Yue Wu, Alokkumar Jha, Seth Sharp, Alessandra Celli, Ekrem Ayhan, Fahim Abbasi, Anna L. Gloyn, Tracey McLaughlin, and Michael P. Snyder. Prediction of metabolic subphenotypes of type 2 diabetes via continuous glucose monitoring and machine learning.Nature Biomedical Engineering, 9(8):1222–1239, August 2025

    2025

  7. [7]

    Diaz C, Patricio Colmegna, Elliot Pryor, and Marc D

    Jenny L. Diaz C, Patricio Colmegna, Elliot Pryor, and Marc D. Breton. A Performance-Based Adaptation Index for Automated Insulin Delivery Systems. Journal of Diabetes Science and Technology, page 19322968251315499, February 2025

    2025

  8. [8]

    Huili Yu, Kevin Meier, Matthew Argyle, and Randal W. Beard. Cooperative Path Planning for Target Tracking in Urban Environments Using Unmanned Air and Ground Vehicles.IEEE/ASME Transactions on Mechatronics, 20(2):541–552, April 2015

    2015

  9. [9]

    Towards Uncertainty- aware Robotic Perception via Mixed-signal BNN Engine Leveraging Probabilistic Quantum Tunneling

    Likai Pei, Yu Zhou, Xingtian Wang, Xueji Zhao, Wanxin Huang, Boyang Cheng, Halid Mulaosmanovic, Stefan Duenkel, Dominik Kleimaier, Sven Beyer, Kai Ni, Mengxue Hou, Michael Niemier, and Ningyuan Cao. Towards Uncertainty- aware Robotic Perception via Mixed-signal BNN Engine Leveraging Probabilistic Quantum Tunneling. In2025 62nd ACM/IEEE Design Automation C...

    2025

  10. [10]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes, Decem- ber 2022

    2022

  11. [11]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. InProceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, pages 6840–6851, Red Hook, NY, USA, December

  12. [12]

    Curran Associates Inc

  13. [13]

    Adaptive Computation and Machine Learning Series

    Daphne Koller and Nir Friedman.Probabilistic Graphical Models: Principles and Techniques. Adaptive Computation and Machine Learning Series. MIT Press, Cambridge, MA, USA, July 2009

    2009

  14. [14]

    An Explainable Bayesian Decision Tree Algorithm.Frontiers in Applied Mathematics and Statistics, 7, March 2021

    Giuseppe Nuti, Lluís Antoni Jiménez Rugama, and Andreea-Ingrid Cross. An Explainable Bayesian Decision Tree Algorithm.Frontiers in Applied Mathematics and Statistics, 7, March 2021

    2021

  15. [15]

    The Algorithmic Foundations of Differential Privacy.Found

    Cynthia Dwork and Aaron Roth. The Algorithmic Foundations of Differential Privacy.Found. Trends Theor. Comput. Sci., 9(3-4):211–407, August 2014

    2014

  16. [16]

    In-Situ Privacy via Mixed-Signal Perturbation and Hardware-Secure Data Reversibility

    Steven Davis, Jianbo Liu, Boyang Cheng, Muya Chang, and Ningyuan Cao. In-Situ Privacy via Mixed-Signal Perturbation and Hardware-Secure Data Reversibility. IEEE Transactions on Circuits and Systems I: Regular Papers, 71(6):2538–2549, June 2024

    2024

  17. [17]

    Stochastic Mixed- Signal Circuit Design for In-Sensor Privacy

    Ningyuan Cao, Jianbo Liu, Boyang Cheng, and Muya Chang. Stochastic Mixed- Signal Circuit Design for In-Sensor Privacy. InProceedings of the 41st IEEE/ACM International Conference on Computer-Aided Design, ICCAD ’22, pages 1–9, New York, NY, USA, December 2022. Association for Computing Machinery

    2022

  18. [18]

    Enciso, Boyang Cheng, Likai Pei, Jianbo Liu, Steven Davis, Michael Niemier, and Ningyuan Cao

    Zephan M. Enciso, Boyang Cheng, Likai Pei, Jianbo Liu, Steven Davis, Michael Niemier, and Ningyuan Cao. A 65 nm Bayesian Neural Network Accelerator with 360 fJ/Sample In-Word GRNG for AI Uncertainty Estimation, January 2025

    2025

  19. [19]

    Richard Dorrance, Deepak Dasalukunte, Hechen Wang, Renzhi Liu, and Brent R. Carlton. An Energy-Efficient Bayesian Neural Network Accelerator With CiM and a Time-Interleaved Hadamard Digital GRNG Using 22-nm FinFET.IEEE Journal of Solid-State Circuits, 58(10):2826–2838, October 2023

    2023

  20. [20]

    20.1 NVE: A 3nm 23.2TOPS/W 12b-Digital-CIM-Based Neural Engine for High-Resolution Visual- Quality Enhancement on Smart Devices

    Ming-En Shih, Shih-Wei Hsieh, Ping-Yuan Tsai, Ming-Hung Lin, Pei-Kuei Tsung, En-Jui Chang, Jenwei Liang, Shu-Hsin Chang, Chung-Lun Huang, You-Yu Nian, Zhe Wan, Sushil Kumar, Cheng-Xin Xue, Gajanan Jedhe, Hidehiro Fujiwara, Haruki Mori, Chih-Wei Chen, Po-Hua Huang, Chih-Feng Juan, Chung-Yi Chen, Tsung-Yao Lin, Ch Wang, Chih-Cheng Chen, and Kevin Jou. 20.1 ...

    2024

  21. [21]

    23.8 An 88.36TOPS/W Bit-Level-Weight-Compressed Large-Language-Model Accelerator with Cluster-Aligned INT-FP-GEMM and Bi-Dimensional Workflow Reformulation

    Yubin Qin, Yang Wang, Jiachen Wang, Zhiwei Lin, Yushu Zhao, Shaojun Wei, Yang Hu, and Shouyi Yin. 23.8 An 88.36TOPS/W Bit-Level-Weight-Compressed Large-Language-Model Accelerator with Cluster-Aligned INT-FP-GEMM and Bi-Dimensional Workflow Reformulation. In2025 IEEE International Solid-State Circuits Conference (ISSCC), volume 68, pages 420–422, February 2025

    2025

  22. [22]

    A 65-nm Digital Stochastic Compute-in-Memory CNN Processor With 8-bit Precision.IEEE Journal of Solid-State Circuits, 60(10):3749–3761, October 2025

    Jiyue Yang, Tianmu Li, Wojciech Romaszkan, Puneet Gupta, and Sudhakar Pa- marti. A 65-nm Digital Stochastic Compute-in-Memory CNN Processor With 8-bit Precision.IEEE Journal of Solid-State Circuits, 60(10):3749–3761, October 2025

    2025

  23. [23]

    ReRAM-Based Pseudo-True Random Number Genera- tor With High Throughput and Unpredictability Characteristics.IEEE Transac- tions on Electron Devices, 68(4):1593–1597, April 2021

    Po-Hao Tseng, Ming-Hsiu Lee, Yu-Hsuan Lin, Hsiang-Lan Lung, Keh-Chung Wang, and Chih-Yuan Lu. ReRAM-Based Pseudo-True Random Number Genera- tor With High Throughput and Unpredictability Characteristics.IEEE Transac- tions on Electron Devices, 68(4):1593–1597, April 2021

    2021

  24. [24]

    Self-Heating Phase- Change Memory-Array Demonstrator for True Random Number Generation

    Enrico Piccinini, Rossella Brunetti, and Massimo Rudan. Self-Heating Phase- Change Memory-Array Demonstrator for True Random Number Generation. IEEE Transactions on Electron Devices, 64(5):2185–2192, May 2017

    2017

  25. [25]

    Probabilistic embeddings for cross-modal retrieval

    Sanghyuk Chun, Seong Joon Oh, Rafael Sampaio De Rezende, Yannis Kalantidis, and Diane Larlus. Probabilistic embeddings for cross-modal retrieval. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8415–8424, 2021

    2021

  26. [26]

    Mervin E. Muller. An Inverse Method for The Generation of Random Normal Deviates on Large-Scale Computers.Mathematical Tables and Other Aids to Computation, 12(63):167–174, 1958

    1958

  27. [27]

    Revisiting Central Limit Theorem: Accurate Gaussian Random Number Generation in VLSI.IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 23(5):842–855, May 2015

    Jamshaid Sarwar Malik, Ahmed Hemani, Jameel Nawaz Malik, Ben Silmane, and Nasirud Din Gohar. Revisiting Central Limit Theorem: Accurate Gaussian Random Number Generation in VLSI.IEEE Transactions on Very Large Scale Integration (VLSI) Systems, 23(5):842–855, May 2015

    2015

  28. [28]

    G. E. P. Box and Mervin E. Muller. A Note on the Generation of Random Normal Deviates.The Annals of Mathematical Statistics, 29(2):610–611, June 1958

    1958

  29. [29]

    The Ziggurat Method for Generating Random Variables.Journal of Statistical Software, 5:1–7, October 2000

    George Marsaglia and Wai Wan Tsang. The Ziggurat Method for Generating Random Variables.Journal of Statistical Software, 5:1–7, October 2000

    2000

  30. [30]

    Practical True Random Number Generator Using CMOS Image Sensor Dark Noise.IEEE Access, 7:91407–91413, 2019

    Byung Kwon Park, Hojoong Park, Yong-Su Kim, Ju-Sung Kang, Yongjin Yeom, Changhui Ye, Sung Moon, and Sang-Wook Han. Practical True Random Number Generator Using CMOS Image Sensor Dark Noise.IEEE Access, 7:91407–91413, 2019

    2019

  31. [31]

    Massey Jr

    Frank J. Massey Jr. The Kolmogorov-Smirnov Test for Goodness of Fit.Journal of the American Statistical Association, 46(253):68–78, March 1951

    1951

  32. [32]

    On Chi-Squared Goodness-of-Fit Test for Normality

    Mikhail Nikulin, Léo Gerville-Réache, and Xuan Quang Tran. On Chi-Squared Goodness-of-Fit Test for Normality. InStatistical Models and Methods for Relia- bility and Survival Analysis, chapter 14, pages 213–227. John Wiley & Sons, Ltd, 2013

    2013

  33. [33]

    Neu- roSim Simulator for Compute-in-Memory Hardware Accelerator: Validation and Benchmark.Frontiers in Artificial Intelligence, 4, June 2021

    Anni Lu, Xiaochen Peng, Wantong Li, Hongwu Jiang, and Shimeng Yu. Neu- roSim Simulator for Compute-in-Memory Hardware Accelerator: Validation and Benchmark.Frontiers in Artificial Intelligence, 4, June 2021

    2021

  34. [34]

    Predictive technology model for nano-cmos design exploration.ACM Journal on Emerging Technologies in Computing Systems (JETC), 3(1):1–es, 2007

    Wei Zhao and Yu Cao. Predictive technology model for nano-cmos design exploration.ACM Journal on Emerging Technologies in Computing Systems (JETC), 3(1):1–es, 2007

    2007

  35. [35]

    Xiangyu Dong, Cong Xu, Yuan Xie, and Norman P. Jouppi. Nvsim: A circuit- level performance, energy, and area model for emerging nonvolatile memory. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 31(7):994–1007, 2012

    2012

  36. [36]

    Lee, Kobbi Nissim, Sofya Raskhod- nikova, and Adam Smith

    Shiva Prasad Kasiviswanathan, Homin K. Lee, Kobbi Nissim, Sofya Raskhod- nikova, and Adam Smith. What Can We Learn Privately?, February 2010

    2010

  37. [37]

    Learning multiple layers of features from tiny images

    Alex Krizhevsky. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009

    2009

  38. [38]

    Mobilenets: Effi- cient convolutional neural networks for mobile vision applications

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. Mobilenets: Effi- cient convolutional neural networks for mobile vision applications. InProceedings of the IEEE conference on computer vision and pattern recognition. IEEE, 2017

    2017

  39. [39]

    Msr-vtt: A large video description dataset for bridging video and language

    Jun Xu, Tao Mei, Ting Yao, and Yong Rui. Msr-vtt: A large video description dataset for bridging video and language. In2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5288–5296, 2016

    2016

  40. [40]

    Imagebind: One embedding space to bind them all

    Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan Misra. Imagebind: One embedding space to bind them all. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15180–15190, 2023. 8

    2023