entropy regularization

**Entropy Regularization** is a **technique that adds the entropy of the model's output distribution to the training objective** — encouraging higher entropy (more exploration, less certainty) or lower entropy (more decisive predictions) depending on the application. **Entropy Regularization Forms** - **Maximum Entropy**: Add $+eta H(p)$ to reward higher entropy — prevents premature convergence to deterministic policies. - **Minimum Entropy**: Add $-eta H(p)$ to penalize high entropy — encourages decisive, low-entropy predictions. - **Semi-Supervised**: Use entropy minimization on unlabeled data — push unlabeled predictions toward confident (low-entropy) decisions. - **Conditional Entropy**: Regularize the conditional entropy $H(Y|X)$ — controls per-input prediction sharpness. **Why It Matters** - **RL Exploration**: Maximum entropy RL (SAC) prevents premature policy collapse — maintains exploration. - **Semi-Supervised**: Entropy minimization is a key component of semi-supervised learning. - **Calibration**: Entropy regularization helps produce well-calibrated probability predictions. **Entropy Regularization** is **controlling the model's decisiveness** — using entropy to balance between confident predictions and exploratory uncertainty.