Master Thesis - Approximative Computing for CNNs

Running convolutional neural networks on edge devices requires careful use of energy and silicon area, with multiply-accumulate operations accounting for much of the hardware cost. Approximate arithmetic can reduce these demands, but its effects on object detection and semantic segmentation are not yet well understood. This thesis at Axis Communications will investigate the balance between prediction accuracy and hardware efficiency when approximate arithmetic circuits are used for these tasks.

The assignment

• Use existing open-source approximate-multiplier libraries and simulation frameworks to study CNNs for object detection and semantic segmentation.
• Measure how approximate multiplication changes detection and segmentation accuracy.
• Compare prediction quality against estimated reductions in power consumption and circuit area.
• Investigate accuracy-recovery approaches, including choosing multipliers per layer, calibration and fine-tuning.

Who we are looking for

• Two master's students studying electrical engineering, computer engineering, computer science or a related subject.
• Strong Python programming skills and experience using a deep learning framework such as PyTorch or TensorFlow.
• Basic knowledge of CNNs, including training and evaluation.
• Studies or practical experience in digital design or computer architecture, covering computer arithmetic.
• Experience with detection or segmentation models, quantization, hardware-aware optimization, Verilog/VHDL, synthesis tools, C/C++ or CUDA is advantageous.
• An interest in the interaction between hardware and machine learning, energy-efficient computing and edge AI, with the ability to draw well-supported conclusions from experiments.
• Pairs with complementary machine learning and hardware expertise are particularly encouraged.
• Applicants must be affiliated with a Swedish university or university college, either directly or through an exchange programme.

What we offer

• An open, friendly and collaborative working culture that encourages new ideas and ongoing development of skills and working methods.

Practical details

Location: Lund. Scope: two students, each completing a 30-credit master's thesis. Apply through the thesis advert in Swedish or English and include a CV and university grade transcript. Apply as a pair, submitting separate applications that clearly identify the other applicant. The listed application deadline is 14 October 2026. Questions about the assignment can be sent to Tiger Moberg at tiger.moberg@axis.com.