myCSUSMPast Projects
Project Summary
This project resulted in a published research paper that earned a best paper award at SEET 2026.
Srivatsan Suresh Babu, Sophia Shahryar, and Justin Morris. "Energy-Efficient Hyperdimensional Computing on Apple Neural Engine Chips." Software Engineering of Emerging Technologies (2026). Best Paper Award.
Hyperdimensional Computing (HDC) is a lightweight machine learning algorithm that is more energy efficient than deep neural networks. The algorithm leverages high-dimensional representations and matrix multiplications for classification. While there is prior work on the efficiency of HDC on specialized accelerators such as the Tensor Processing Unit (TPU), there is a lack of exploration of HDC on the Apple Neural Engine (ANE). In this paper, we evaluate the performance and energy efficiency of HDC inference across different Apple Silicon devices, such as the M3 Pro, M4, and A19 Pro chips. We express HDC inference as a two-layer neural network composed of a fixed random projection and a linear similarity layer, enabling conversion from PyTorch to CoreML for hardware-accelerated execution. Using the scikit-learn handwritten digits dataset, we benchmark inference timing and energy consumption across CPUs, GPUs, and the ANE across different batch sizes and hypervector dimensionalities. Our results demonstrate that the ANE consistently achieves better energy efficiency for large batch sizes and high-dimensional hypervectors, with up to 58% energy reduction compared to CPU execution. These findings highlight the suitability of domain-specific accelerators for HDC workloads and demonstrate the practicality of deploying HDC efficiently on modern edge devices.
Reflective Statement
Working on the "Energy-Efficient Hyperdimensional Computing on Apple Neural Engine Chips" project completely shaped the engineer I am today. My friend Sophia and I wanted to prepare early in our careers, so we asked Justin if we could join his research on hyperdimensional computing. As a freshman with no prior knowledge of machine learning, where even the concept of a training dataset was foreign, I remember reading his paper and finding it super confusing.
That was only the algorithm portion. Launching it on the Apple Neural Engine was even harder. I was used to following YouTube tutorials, but this time I had to read documentation that sometimes didn’t even work. Since I didn’t know anyone specializing in these frameworks, I had to learn most of it on my own. After around a year of tackling those hurdles, I finally figured out how to use the ANE. Around that same time, I started my machine learning internship at Pacira, an offer I directly credit to this research role.
After my internship, my knowledge of machine learning skyrocketed. Coming back to the research with multiple projects under my belt, I easily mapped the HDC algorithm onto the Apple Neural Engine chip, completed all the tests within a semester, and focused on writing the paper the next. Though our paper was rejected the first time, our persistence paid off: getting into SEET 2026 led to a Best Paper Award at the Student Research Competition (SRC).
Looking back, I’d like to give a massive portion of credit to Professor Justin Morris. He helped me get into my internships (FPGA and ML), this research role, a position at a robotics startup, and now, multiple interviews. Whenever I'm in an interview, I find myself talking about the exact experiences and growth given to me by Justin Morris and SKALE.
Skill and Tool Assessment
Skills Gains
- Machine Learning Acceleration
- Performance Profiling
- Technical Writing
- Academic Research
Tools Learned to Use
- Domain Specific Languages (PyTorch)
- Swift Programming
- Matlab Technical Plotting






