Master Thesis – Designing Ternary Neural Networks for Edge Vision

Running computer-vision models on edge devices requires balancing accuracy against limited memory, power and processing capacity. This thesis at Axis Communications explores neural networks whose weights are restricted to −1, 0 and +1, investigating how their architecture and training can be adapted for object detection. The aim is to establish whether these ternary models offer practical advantages on available edge hardware compared with higher-precision alternatives.

The assignment

• Study techniques for training ternary neural networks and apply them to existing object detectors.
• Investigate architectural and training changes that help preserve accuracy, including alternative network blocks, quantization-aware training, activation precision and mixed-precision designs.
• Benchmark the models against appropriate higher-precision or INT8 reference models.
• Measure accuracy, model size, CPU latency and throughput, and assess energy consumption if suitable measurement equipment is available.
• Identify model and implementation choices that improve performance on real edge devices.
• Discuss possible future applications in compute-in-memory hardware, such as memristor crossbars; working with this hardware is not required.

Who we are looking for

• Two master's students in Computer Science, Electrical Engineering, Computer Engineering or a related discipline.
• A keen interest in deep learning, computer vision, model compression, embedded systems or computer architecture.
• Some programming knowledge, ideally in Python, and experience using PyTorch or a comparable deep-learning framework.
• Affiliation with a Swedish university or university college, including through an exchange programme.

What we offer

• An opportunity to evaluate low-precision vision models using available edge hardware.
• A friendly, open working culture that values collaboration and encourages new ideas.

Practical details

Location: Lund. Scope: two students, with 30 credits per student. Apply through the thesis advert by 27 November 2026; applications may be written in Swedish or English. Apply as a pair, submitting separate applications and identifying your thesis partner in each. Include a CV, cover letter and university transcript. Questions: Magnus Svensson, Engineering Manager, Magnus.Svensson@axis.com, +46 46 2879187.