All problems
AI
Neural Network Quantization for Edge AI
Code: ANX-RP-004
Investigate and develop mixed-precision quantization techniques that maintain model accuracy while enabling inference on microcontroller-class hardware.
Why it matters
Deploying AI on IoT devices requires models small enough to fit in kilobytes of memory. Current quantization methods cause significant accuracy drops on complex tasks.
Research gap
Post-training quantization methods don't generalize well across architectures. Task-aware quantization for multi-task models is underexplored.
Research questions
- 01 What is the minimum bit-width for BERT-family models on keyword spotting?
- 02 Can we use NAS to find quantization-friendly architectures?
- 03 How do quantization errors compound across multi-stage pipelines?
Suggested technologies
Required skills
Team and supervisor
1-2 researchers, supervised by Muhammad Qasim.
Suitable for: AI, CS, SE.
Expected deliverables
- Quantization toolkit for edge deployment
- Accuracy vs latency benchmark across hardware targets
- Paper on task-aware mixed-precision strategies
Publication
Mixed-Precision Quantization of BERT for Keyword Spotting on Microcontrollers
Published · 2026-01-10
Work on this problem
Apply to the lab and this problem becomes your supervised research project.