Edge AI: Approximate Computing for Efficient CNNs

Axis Communications AB; Lunds kommun; On-site; SEK 190,000 - 279,000; Full time

Job summary

Axis Communications AB in Lund, Sweden, invites two Master’s students in Electrical/Computer Engineering or Computer Science to pursue a Master Thesis on Approximative Computing for CNNs. The project focuses on deploying CNNs for object detection and semantic segmentation on edge devices, studying the impact of approximate arithmetic on accuracy, power, and area, and exploring recovery strategies.

Ideal candidates have Python proficiency and experience with PyTorch or TensorFlow, a basic

Qualifications

• Strong Python programming skills and experience with a deep learning framework such as PyTorch or TensorFlow.
• Solid understanding of convolutional neural networks and how they are trained and evaluated.
• Coursework or experience in digital design or computer architecture, including computer arithmetic.

Responsibilities

• Investigate how approximate arithmetic circuits affect CNN performance for object detection and segmentation.
• Quantify trade-offs between accuracy and power/area savings on edge devices.
• Explore methods to recover lost accuracy, e.g., layer-wise multiplier selection, calibration, and fine-tuning.

Skills

Python

CNN basics

Experimentation

Education

Master's degree in Electrical/Computer Engineering or related

Tools

PyTorch

TensorFlow

Verilog/VHDL

CUDA

Job description

Axis Communications AB in Lund, Sweden, invites two Master’s students in Electrical/Computer Engineering or Computer Science to pursue a Master Thesis on Approximative Computing for CNNs. The project focuses on deploying CNNs for object detection and semantic segmentation on edge devices, studying the impact of approximate arithmetic on accuracy, power, and area, and exploring recovery strategies.

Ideal candidates have Python proficiency and experience with PyTorch or TensorFlow, a basic