Background
A company’s predictive maintenance model ran accurately in the cloud but needed to run on low-power edge devices with strict memory and compute limits. Standard compression techniques reduced model size but degraded accuracy below an acceptable threshold.
The Challenge
The team needed to determine whether the model could be compressed enough to fit the device constraints while retaining the accuracy required for reliable failure prediction — an outcome that wasn’t predictable using standard compression methods alone.
Technological Uncertainty
It was not known whether any combination of pruning, quantization, and architecture modification could meet the memory and latency budget without an unacceptable accuracy loss, given the specific sensor data characteristics involved.
Experimental Development
The team systematically tested combinations of structured pruning, quantization levels, and knowledge distillation from the full-size model, measuring accuracy, memory footprint, and inference latency at each configuration.
What Failed?
Standard post-training quantization alone reduced size sufficiently but caused an unacceptable accuracy drop on rare failure signatures, and an initial distillation attempt using a generic teacher-student setup failed to transfer that rare-event sensitivity.
Technological Advancement
The team developed a custom distillation approach that specifically preserved sensitivity to rare failure signatures during compression, generating new technical knowledge about maintaining rare-event accuracy under aggressive model compression.
Potentially Relevant SR&ED Activities
- Systematic testing of pruning and quantization configurations against accuracy and latency targets
- Development of a custom knowledge-distillation approach for rare-event preservation
- Benchmarking of memory footprint and inference latency across iterations on target hardware
What Would Generally Not Qualify
Applying the finalized compressed model to additional identical edge devices, without further modification, would not itself qualify as ongoing eligible work.
Documentation
Benchmark logs comparing accuracy, memory, and latency across compression configurations, along with technical notes explaining why standard approaches were insufficient, would support this claim.
