Description
A neural network is the parametric model class — composition of affine maps with non-linearities — underlying every deep-learning method. The thesis uses the term in two ways: (a) as a generic synonym for "the learned model" in surveys and high-level claims, and (b) as the concrete fully-connected (MLP) architecture used as a baseline against cnn, lstm, and transformer-attention variants on CSI tasks.
When it's used
- Baseline classifier on top of hand-crafted CSI features
- Final regression head on top of CNN/Transformer backbones
- Theoretical statements about expressivity and capacity
Limitations
- Plain MLPs underperform structured architectures on CSI sequences
- Overfit small datasets without regularisation
- No built-in inductive bias for spatial / temporal structure
Source Papers
- di2023_285b ↗ — neural-network ingredients of physics-informed crowd model
- meegahapola2024_a321 ↗ — neural-net components in domain adaptation
- chaudhari2024_6efc ↗ — neural-network classifiers on CSI features
- chen2018_97e0 ↗ — neural-net occupancy estimator
- khan2024_43e8 ↗ — Bayesian-neural-net WiFi sensing