DANA: Dimension-Adaptive Neural Architecture for Multivariate Sensor Data
summary
Proposes a concrete method—Dimension-Adaptive Pooling (DAP)—that sits between the last convolutional map and the first dense/recurrent layer to accept variable sensor counts and sampling rates while outputting fixed-size tensors, plus a training recipe—Dimension-Adaptive Training (DAT)—that randomly samples input dimensions and accumulates gradients across batches. Combines them into DANA, converts existing DNNs without extra parameters, and shows on four real-world human-activity-recognition datasets and three synthetic sets up to ~20% higher accuracy under missing sensors and changing sampling rates.
AI-written summary · about
related talks
47mInterpreting Deep Neural Networks (DNNs)
06mAdversarial Multi-view Networks for Activity Recognition
15mADaPS: Adaptive Data Partitioning to Parallelize CNN Inference on Resource-Constrained(…)
01h 30mDeep Learning Techniques for Monocular Depth Estimation
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
01h 02mMulti-Dimension Mapping of Embedded Applications
01h 01mMapping multisets to vectors
18mDVABatch: Diversity-aware Multi-Entry Multi-Exit Batching for Efficient...