Alisha Menon

EECS Department, University of California, Berkeley

Technical Report No. UCB/EECS-2022-13

May 1, 2022

http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-13.pdf

To address prosthetic applications and the issues that cause users to abandon them, particularly to lessen the burden of prosthetic control on users, this work aims to implement the concept of “shared control” for prosthetic reflex task response. The proposed algorithm, Hyperdimensional Computing (HDC), has previously been demonstrated to recall reactive behavior [1] and also improves sensor fusion performance by inherently binding features extracted from arbitrary data streams. Towards exploring HDC sensor fusion, both early and late sensor fusion inference ASICs are synthesized in the TSMC 28HPM process and various optimizations to the algorithm are proposed and evaluated that have significant implications on multi-modal fusion power and area requirements. Additionally, vectorization of the many-channel HDC architecture is explored through implementation on the Rocketchip RISC-V processor [2] by optimizing an existing C implementation of HDC inference [3]. A strategic combination of algorithm tweaks, code optimizations, and vector acceleration using the Hwacha vector processor [4] yielded multiple orders of magnitude (∼200×) speedups for HDC classification compared to [3]. Towards the goal of developing a platform to explore HDC recall of reactive behavior for prosthetic applications, force sensors have been selected and characterized, and a PCB has been fabricated to integrate the force sensors into the existing EMG classification adapter board [5]. The proposed algorithm was demonstrated to be well-suited for sensor fusion tasks leaving exciting avenues for future work in integrating the sensor fusion properties with the recall of reactive behavior platform.

Advisors: Jan M. Rabaey

---

BibTeX citation:

@mastersthesis{Menon:EECS-2022-13,
    Author= {Menon, Alisha},
    Title= {HD Recall of Reactive Behavior through Multi-Modal Sensor Fusion},
    School= {EECS Department, University of California, Berkeley},
    Year= {2022},
    Month= {May},
    Url= {http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-13.html},
    Number= {UCB/EECS-2022-13},
    Abstract= {To address prosthetic applications and the issues that cause users to abandon them, particularly to lessen the burden of prosthetic control on users, this work aims to implement the concept of “shared control” for prosthetic reflex task response. The proposed algorithm, Hyperdimensional Computing (HDC), has previously been demonstrated to recall reactive behavior [1] and also improves sensor fusion performance by inherently binding features extracted from arbitrary data streams. Towards exploring HDC sensor fusion, both early and late sensor fusion inference ASICs are synthesized in the TSMC 28HPM process and various optimizations to the algorithm are proposed and evaluated that have significant implications on multi-modal fusion power and area requirements. Additionally, vectorization of the many-channel HDC architecture is explored through implementation on the Rocketchip RISC-V processor [2] by optimizing an existing C implementation of HDC inference [3]. A strategic combination of algorithm tweaks, code optimizations, and vector acceleration using the Hwacha vector processor [4] yielded multiple orders of magnitude (∼200×) speedups for HDC classification compared to [3]. Towards the goal of developing a platform to explore HDC recall of reactive behavior for prosthetic applications, force sensors have been selected and characterized, and a PCB has been fabricated to integrate the force sensors into the existing EMG classification adapter board [5]. The proposed algorithm was demonstrated to be well-suited for sensor fusion tasks leaving exciting avenues for future work in integrating the sensor fusion properties with the recall of reactive behavior platform.},
}

EndNote citation:

%0 Thesis
%A Menon, Alisha 
%T HD Recall of Reactive Behavior through Multi-Modal Sensor Fusion
%I EECS Department, University of California, Berkeley
%D 2022
%8 May 1
%@ UCB/EECS-2022-13
%U http://www2.eecs.berkeley.edu/Pubs/TechRpts/2022/EECS-2022-13.html
%F Menon:EECS-2022-13