AFEC: Active Forgetting of Negative Transfer in Continual Learning
Dec 6, 2021
Speakers
Liyuan Wang
Speaker · 0 followers
About
Continual learning aims to learn a sequence of tasks from dynamic data distributions. Without access to the old training samples, knowledge transfer from the old tasks to each new task is difficult to determine, which might be either positive or negative. If the old knowledge interferes with the learning of a new task, i.e. the forward knowledge transfer is negative, then precisely remembering the old tasks will further aggravate the interference, thus decreasing the performance of continual learning. By contrast, biological neural networks can actively forget the old knowledge that conflicts with the learning of a new experience, through regulating the learning-triggered synaptic expansion-convergence. Inspired by the biological active forgetting, we propose to actively forget the old knowledge that limits the learning of new tasks to benefit continual learning. Under the framework of Bayesian continual learning, we develop a novel approach named Active Forgetting with synaptic Expansion-Convergence (AFEC), which is formally consistent with the biological weight updating mechanism. We extensively evaluate AFEC on a variety of continual learning benchmarks, including CIFAR10 regression tasks, visual classification tasks and Atari reinforcement tasks, where AFEC achieves the state-of-the-art performance through effectively improving the learning of new tasks. Further, AFEC can be a plug-and-play approach that significantly boosts the performance of representative continual learning strategies.Continual learning aims to learn a sequence of tasks from dynamic data distributions. Without access to the old training samples, knowledge transfer from the old tasks to each new task is difficult to determine, which might be either positive or negative. If the old knowledge interferes with the learning of a new task, i.e. the forward knowledge transfer is negative, then precisely remembering the old tasks will further aggravate the interference, thus decreasing the performance of continual lea…
Organizer
NeurIPS 2021
Account · 1.9k followers
About NeurIPS 2021
Neural Information Processing Systems (NeurIPS) is a multi-track machine learning and computational neuroscience conference that includes invited talks, demonstrations, symposia and oral and poster presentations of refereed papers. Following the conference, there are workshops which provide a less formal setting.
Like the format? Trust SlidesLive to capture your next event!
Professional recording and live streaming, delivered globally.
Sharing
Recommended Videos
Presentations on similar topic, category or speaker
Rethinking and Reweighting the Univariate Losses for Multi-Label Ranking: Consistency and Generalization
Watch later
Guoqiang Wu, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Can contrastive learning avoid shortcut solutions?
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Scalable Online Planning via Reinforcement Learning Fine-Tuning
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Large Scale Coordination Transfer for Cooperative Multi-Agent Reinforcement Learning
Watch later
Ethan Wang, …
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Sanity Checks for Lottery Tickets: Does Your Winning Ticket Really Win the Jackpot?
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%
Accelerating Quadratic Programming with Reinforcement Learning
Watch later
Total of 0 viewers voted for saving the presentation to eternal vault which is 0.0%