Efficient Mixture of Geographical Species for On Device Wildlife Monitoring
This work addresses the need for low-compute models for ecological conservation, though it appears incremental by applying conditional computation to a new edge use case.
The paper tackles the problem of efficient on-device wildlife monitoring by training a single species detector that uses conditional computation to bias subnetworks based on geographic location, achieving performance on datasets like iNaturalist and iWildcam.
Efficient on-device models have become attractive for near-sensor insight generation, of particular interest to the ecological conservation community. For this reason, deep learning researchers are proposing more approaches to develop lower compute models. However, since vision transformers are very new to the edge use case, there are still unexplored approaches, most notably conditional execution of subnetworks based on input data. In this work, we explore the training of a single species detector which uses conditional computation to bias structured sub networks in a geographically-aware manner. We propose a method for pruning the expert model per location and demonstrate conditional computation performance on two geographically distributed datasets: iNaturalist and iWildcam.