Forget by Uncertainty: Orthogonal Entropy Unlearning for Quantized Neural Networks
Researchers have developed a new framework called Orthogonal Entropy Unlearning (OEU) designed to effectively remove specific data from quantized neural networks while preserving overall model accuracy. This method addresses limitations in existing unlearning techniques by maximizing prediction uncertainty on forgotten data, thus avoiding confident mispredictions. OEU also employs gradient orthogonal projection to prevent interference between forgetting and retaining data gradients, offering theoretical guarantees for utility preservation. AI
IMPACT Provides a novel method for data unlearning in quantized neural networks, crucial for privacy compliance in edge device deployments.