Categorical cross-entropy is a loss function used to train classification models by measuring the divergence between the predicted probability distribution over discrete class labels and the true one-hot encoded label distribution, penalizing confident incorrect predictions more heavily. In WiFi CSI sensing, it is the standard optimization objective for multi-class tasks such as people counting, activity recognition, and occupancy detection, guiding models like CNN+LSTM architectures to produce well-calibrated class probabilities across mutually exclusive categories. Key variants include sparse categorical cross-entropy, which accepts integer class indices rather than one-hot vectors for computational efficiency, and weighted categorical cross-entropy, which adjusts per-class penalty terms to address class imbalance commonly encountered in real-world CSI datasets where some occupancy levels or activities are underrepresented.
Dictionary term