EPISODE · Apr 11, 2025 · 13 MIN
Continual Forgetting for Pre-trained Vision Models
from Neural intel Pod · host Neuralintel.org
The provided research paper explores the novel problem of continual forgetting in pre-trained vision models, where the goal is to sequentially remove specific unwanted knowledge while retaining the model's performance on other tasks. To address this, the authors introduce Group Sparse LoRA (GS-LoRA++), a parameter-efficient fine-tuning approach that uses low-rank adaptation with a group sparsity regularization to selectively modify the model's feed-forward networks. Furthermore, to handle practical scenarios like scarce forgetting data, they incorporate prototype regularization to guide the forgetting process. Extensive experiments across various vision tasks demonstrate the effectiveness and efficiency of their proposed method in achieving selective and continual knowledge erasure with minimal impact on remaining knowledge.
What this episode covers
The provided research paper explores the novel problem of continual forgetting in pre-trained vision models, where the goal is to sequentially remove specific unwanted knowledge while retaining the model's performance on other tasks. To address this, the authors introduce Group Sparse LoRA (GS-LoRA++), a parameter-efficient fine-tuning approach that uses low-rank adaptation with a group sparsity regularization to selectively modify the model's feed-forward networks. Furthermore, to handle practical scenarios like scarce forgetting data, they incorporate prototype regularization to guide the forgetting process. Extensive experiments across various vision tasks demonstrate the effectiveness and efficiency of their proposed method in achieving selective and continual knowledge erasure with minimal impact on remaining knowledge.
NOW PLAYING
Continual Forgetting for Pre-trained Vision Models
No transcript for this episode yet
Similar Episodes
Mar 14, 2026 ·23m
Mar 11, 2026 ·16m
Feb 28, 2026 ·14m