Description
LSTM is a gated recurrent neural network designed to keep a long-term memory cell across time steps without vanishing gradients. In CSI sensing, LSTMs are the workhorse for sequence-level tasks — HAR, gesture-sequence classification, occupancy time-series — where convolutional receptive fields are too short and Transformers too heavy. Bi-directional and stacked variants are common.
When it's used
- Activity-sequence classification on CSI windows
- Occupancy / crowd-count time-series regression
- Sequence backbone before a Transformer is justified by data size
- Hybrid CNN-LSTM pipelines
Limitations
- Sequential training is slow vs Transformers
- Long-range dependencies still limited above ~1000 steps
- More finicky to tune than CNN baselines