Team10 - Library Master
Challenge: Scanning for Lifeforms
Keyword: IoT, Few-shot Learning, Model Compression, NeuroPilot, Biodiversity exploration

Many previous research works have shown the effeciency of Few-shot learning when lacking of labeled data. In our challenge, biodiversity exploration for rare species, we are facing the dilemma which is similar to Few-shot learning situation. Therefore Few-shot Learning strategy may classify more precise among common species and rare one than naïve classification methods.

• To evaluate this speculate, we have done several experiments to show the advantages of Few-shot learning.
• In order to improve the performance when implementing on edge devices, we also show the model compression work we have done.

• Dockerfile

FROM pytorch/pytorch:1.6.0-cuda10.1-cudnn7-runtime

RUN pip install scipy numpy tensorboardX pandas scikit-learn Pillow h5py

• naïve classification
• Few-shot learning
• Edge implement

• The figure(day view) on the left shows the superior of model trained on Few-shot learning, the right one(night) shows the further impact compare to naïve classification model
• The table shows that our model compression work significantly improve model inference time and power consumption by model quantization

• Deep Learning for Large Scale Bio-diversity Monitoring
• A deep active learning system for species identification and counting in camera trap images
• Scene‐specific convolutional neural networks for video‐based biodiversity detection
• Regularizing Meta-Learning via Gradient Dropout