Brian Liao

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2020-78

May 28, 2020

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-78.pdf

We present MicroBotNet, a neural network architecture for image classification with fewer than 1 million multiply-and-accumulate (MAC) operations. We wish to explore architectures suitable for microrobots with a goal of less than 1 µJ per forward-pass. We estimate this to be feasible with 1 million MAC operations. MicroBotNet achieves 80.47% accuracy with 740,000 MAC operations on CIFAR-10. Additionally, 60% of weights are quantized to {1, 0, +1} using Trained Ternary Quantization. We also evaluate MicroBotNet on our Micro Robot Dataset, which is composed of 10 image classes a microrobot may encounter such as acorns, mushrooms, and ladybugs. After applying transfer learning, MicroBotNet achieves 67.80% accuracy on the Micro Robot Dataset. Finally, we test MicroBotNet on acorn images simulating a microrobot. MicroBotNet correctly identify the acorn in 7 out of 8 cases when approaching an acorn and 15 out of 16 cases from different angles using a best of last three frames filter.

Advisors: Kristofer Pister

---

BibTeX citation:

@mastersthesis{Liao:EECS-2020-78,
    Author= {Liao, Brian},
    Title= {MicroBotNet: Low Power Neural Networks for Microrobots},
    School= {EECS Department, University of California, Berkeley},
    Year= {2020},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-78.html},
    Number= {UCB/EECS-2020-78},
    Abstract= {We present MicroBotNet, a neural network architecture for image classification with fewer than 1 million
multiply-and-accumulate (MAC) operations. We wish to explore architectures suitable for microrobots with a goal of
less than 1 µJ per forward-pass. We estimate this to be
feasible with 1 million MAC operations. MicroBotNet achieves
80.47% accuracy with 740,000 MAC operations on CIFAR-10.
Additionally, 60% of weights are quantized to {1, 0, +1} using
Trained Ternary Quantization. We also evaluate MicroBotNet
on our Micro Robot Dataset, which is composed of 10 image
classes a microrobot may encounter such as acorns, mushrooms,
and ladybugs. After applying transfer learning, MicroBotNet
achieves 67.80% accuracy on the Micro Robot Dataset. Finally,
we test MicroBotNet on acorn images simulating a microrobot.
MicroBotNet correctly identify the acorn in 7 out of 8 cases
when approaching an acorn and 15 out of 16 cases from
different angles using a best of last three frames filter.},
}

EndNote citation:

%0 Thesis
%A Liao, Brian 
%T MicroBotNet: Low Power Neural Networks for Microrobots
%I EECS Department, University of California, Berkeley
%D 2020
%8 May 28
%@ UCB/EECS-2020-78
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2020/EECS-2020-78.html
%F Liao:EECS-2020-78