Description
An autoencoder is a neural network trained to reconstruct its input through a low-dimensional bottleneck, yielding an unsupervised representation in the bottleneck layer. In CSI sensing it is the simplest pretraining trick: train a CSI autoencoder on unlabeled traffic, freeze the encoder, and use the bottleneck features for downstream classification or anomaly detection.
When it's used
- Unsupervised CSI feature extraction
- Anomaly detection (large reconstruction error → unexpected event)
- Denoising raw CSI before classifier ingestion
- Compression for low-bandwidth deployments
Limitations
- Supplanted by contrastive / masked SSL on most benchmarks
- Bottleneck size choice is empirical
- VAE versions add training instability