Scribd Logo
Upload

Welcome to Scribd!

  • 9 views

    ElieNicolas BNNs

    Uploaded by

    kzkt5ssfn6
    This document outlines a study comparing binary neural networks (BNNs) to convolutional neural networks (CNNs) on an FPGA. BNNs are expected to have much lower accuracy but be faster, smaller, and more efficient than CNNs. The analysis will evaluate the accuracy, size, speed, and performance of different BNN types against CNN baselines. The goal is to determine if optimized BNNs can replace CNNs for applications requiring efficient neural network inference on FPGAs and IoT devices.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd

    ElieNicolas BNNs

    Uploaded by

    kzkt5ssfn6
    9 views16 pages
    This document outlines a study comparing binary neural networks (BNNs) to convolutional neural networks (CNNs) on an FPGA. BNNs are expected to have much lower accuracy but be faster, smaller, and more efficient than CNNs. The analysis will evaluate the accuracy, size, speed, and performance of different BNN types against CNN baselines. The goal is to determine if optimized BNNs can replace CNNs for applications requiring efficient neural network inference on FPGAs and IoT devices.

    Original Title

    ElieNicolas_BNNs

    Copyright

    © © All Rights Reserved

    Available Formats

    PPTX, PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document outlines a study comparing binary neural networks (BNNs) to convolutional neural networks (CNNs) on an FPGA. BNNs are expected to have much lower accuracy but be faster, smaller, and more efficient than CNNs. The analysis will evaluate the accuracy, size, speed, and performance of different BNN types against CNN baselines. The goal is to determine if optimized BNNs can replace CNNs for applications requiring efficient neural network inference on FPGAs and IoT devices.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    9 views16 pages

    ElieNicolas BNNs

    Uploaded by

    kzkt5ssfn6
    This document outlines a study comparing binary neural networks (BNNs) to convolutional neural networks (CNNs) on an FPGA. BNNs are expected to have much lower accuracy but be faster, smaller, and more efficient than CNNs. The analysis will evaluate the accuracy, size, speed, and performance of different BNN types against CNN baselines. The goal is to determine if optimized BNNs can replace CNNs for applications requiring efficient neural network inference on FPGAs and IoT devices.

    Copyright:

    © All Rights Reserved
    Download as PPTX, PDF, TXT or read online from Scribd
    Download as pptx, pdf, or txt
    of 16

    December 9,

    2022
    Binary Neural Elie Nicolas

    Networks: An Extensive University of


    Balamand
    Investigation on FPGA Department
    of Computer
    Engineering
    Outline

    • Introduction
    • Methodology
    • Research Question
    • Analysis Technique
    • Results
    a. Accuracy
    b. Size
    c. Speed
    d. Performance
    • Conclusion
    Introduction

    What are Neural Networks? Convolutional Neural Networks


    • Are computing systems inspired by • Special type of neural networks that
    human learning in their biological uses convolution instead of general
    neurons that constitute of their brain multiplication.
    and nervous system. • Designed for image processing and
    • They imitate biological learning and recognition
    how neurons work in our body.
    Introduction (cont.)

    Binarized Neural Networks


    • A variant of convolutional neural
    networks in which all the weights
    and activations are purely in binary.
    • Its main idea is that since it is binary,
    it will use much less memory and
    processing than full precision.
    • Due to 1-bit of precision, a lot of
    information will be lost (loss of
    accuracy)
    Procedure/Methodology

    • In this project, we’ll be comparing the most optimized and


    efficient variants of Binary Neural Networks with Convolutional
    Neural Networks on a DE1-SoC FPGA to see if they are worth using
    nowadays instead of CNNs.
    • Key assumptions:
    1. BNNs are much faster and more efficient that CNNs
    2. BNNs will have a huge loss of accuracy compared to CNNs
    3. Modified and optimized versions of BNNs might have a smaller
    gap compared to CNNs
    Research Question

    • The main question we are trying to answer is:


    • Are Binary Neural Networks able to replace ordinary Convolutional
    Neural Networks for FPGAs and IoTs? If so, what are the best
    optimization technique for BNNs to replace and better other types
    of neural networks?
    Analysis Technique
    • A Comparative, Statistical approach is used.
    • Averaging of some figures is also used.
    Expected Results

    • CNNs having better accuracy than BNNs, even the most optimized
    types.
    • BNNs being faster that CNNs, with less complexity and memory
    usage.
    • Certain BNNs having an acceptable accuracy.
    Discussion of Results

    a. Accuracy b. Size

    c. Speed d. Performance
    a. Accuracy

    Full Accuracy BNN Type Accuracy


    Precision (%) (%)
    CNN Type XNOR-Net 87.7
    VGG-Small 92.6 PCNN 91.3
    NIN 89.1 RBNN 90.6
    ResNet-18 91.2 BONN 96.4
    WRN-22 90.8 ReActNet 88.9
    b. Size (Average)

    • Notice how BNNs are as 16 times


    smaller than CNNs in size.
    • Theoretically it should be 32 times
    smaller, but size the optimizations and
    variants take some space also
    c. Speed

    Full-Precision Time (ms) BNN Type Time (ms)


    CNN Type XNOR-Net 5.2
    VGG-Small 52 PCNN 7.5
    NIN 48 RBNN 4.3
    ResNet-18 57 BONN 6.1
    WRN-22 49.2 ReActNet 5.7
    d. Performance

    Full-Precision FLOPs BNN Type FLOPs


    CNN Type (x108) (x108)
    VGG-Small 16.7 XNOR-Net 1.4
    NIN 12.9 PCNN 1.2
    ResNet-18 13.6 RBNN 1.22
    WRN-22 12.7 BONN 1.5
    ReActNet 1.32
    Data observations

    Average of CNNs Average of BNNs

    Accuracy Speed Size Performance Needed


    Conclusion

    • BNNs are an interesting type of recently developed neural networks which


    provide a faster and more efficient neural network than ordinary CNNs.

    • Results suggest that BNNs might be useful and able to replace CNNs, but
    under certain conditions in which BNNs are optimized.
    References

    1. Kim, S., & Rutenbar, R. (2018, April). Accelerator Design with Effective Resource Utilization for Binary
    Convolutional Neural Networks on an FPGA. 2018 IEEE 26th Annual International Symposium on
    Field-Programmable Custom Computing Machines (FCCM). https://doi.org/10.1109/fccm.2018.00052
    2. Valencia, R., & Sham, C. W. (2020, December 16). FPGA deployment of neuroevolved Binary Neural
    Networks. 2020 IEEE Asia-Pacific Conference on Computer Science and Data Engineering (CSDE).
    https://doi.org/10.1109/csde50874.2020.9411568
    3. Zhou, Y., Redkar, S., & Huang, X. (2017, August). Deep learning binary neural network on an FPGA.
    2017 IEEE 60th International Midwest Symposium on Circuits and Systems (MWSCAS).
    https://doi.org/10.1109/mwscas.2017.8052915.
    4. Hatano, T., & Yamane, S. (2020, October 13). Boosting Binary Neural Networks for FPGA. 2020 IEEE
    9th Global Conference on Consumer Electronics (GCCE).
    https://doi.org/10.1109/gcce50665.2020.9292056
    5. Fukuda, Y., & Kawahara, T. (2018, May). Stochastic weights binary neural networks on FPGA. 2018 7th
    International Symposium on Next Generation Electronics (ISNE).
    https://doi.org/10.1109/isne.2018.8394726

    6. Yuan, C., & Agaian, S. (in press). A comprehensive review of Binary Neural Network. Cs.NE.
    https://doi.org/10.48550/arXiv.2110.06804

    You might also like