- Overcoming Catastrophic Forgetting by XAI
Authors: Giang Nguyen
Summary: Explaining the behaviors of deep neural networks, often thought of as black packing containers, is crucial particularly when they’re now being adopted over various points of human life. Taking the benefits of interpretable machine studying (interpretable ML), this work proposes a novel device known as Catastrophic Forgetting Dissector (or CFD) to clarify catastrophic forgetting in continuous studying settings. We additionally introduce a brand new methodology known as Vital Freezing based mostly on the observations of our device. Experiments on ResNet articulate how catastrophic forgetting occurs, notably exhibiting which parts of this well-known community are forgetting. Our new continuous studying algorithm defeats numerous current methods by a big margin, proving the potential of the investigation. Vital freezing not solely assaults catastrophic forgetting but in addition exposes explainability.
2. Scale back Catastrophic Forgetting of Dense Retrieval Coaching with Teleportation Negatives
Authors: Si Sun, Chenyan Xiong, Yue Yu, Arnold Overwijk, Zhiyuan Liu, Jie Bao
Summary: On this paper, we examine the instability in the usual dense retrieval coaching, which iterates between mannequin coaching and exhausting unfavorable choice utilizing the being-trained mannequin. We present the catastrophic forgetting phenomena behind the coaching instability, the place fashions be taught and overlook completely different unfavorable teams throughout coaching iterations. We then suggest ANCE-Tele, which accumulates momentum negatives from previous iterations and approximates future iterations utilizing lookahead negatives, as “teleportations” alongside the time axis to easy the educational course of. On internet search and OpenQA, ANCE-Tele outperforms earlier state-of-the-art techniques of comparable dimension, eliminates the dependency on sparse retrieval negatives, and is aggressive amongst techniques utilizing considerably extra (50x) parameters. Our evaluation demonstrates that teleportation negatives cut back catastrophic forgetting and enhance convergence pace for dense retrieval coaching. Our code is accessible at https://github.com/OpenMatch/ANCE-Tele