Batch normalization is a neural network training technique that normalizes the activations of each layer across a mini-batch of samples by adjusting and scaling them to have zero mean and unit variance, using learnable parameters to restore representational capacity. In WiFi CSI-based sensing tasks such as people counting and occupancy detection, it is particularly valuable because it stabilizes training dynamics, accelerates convergence, and reduces sensitivity to initialization, which is critical when working with noisy, high-dimensional CSI signals that vary across environments and hardware configurations. Common variants include layer normalization, which normalizes across features rather than the batch dimension and is often preferred in recurrent architectures like LSTM, and instance normalization, which is applied per-sample and per-channel, each offering trade-offs in stability and generalization relevant to the sequential and multi-receiver data structures common in CSI sensing research.
Dictionary term