Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks.
Sleep-like unsupervised replay reduces catastrophic forgetting in artificial neural networks.
复制标题
类似睡眠的无监督重放减少了人工神经网络中的灾难性遗忘。
DOI：
10.1038/s41467-022-34938-7
复制
发表时间：
2022-12-15
影响因子：
16.6
通讯作者：
Bazhenov, Maxim
中科院分区：
文献类型：
--
作者：
Tadros, Timothy;Krishnan, Giri P.;Ramyaa, Ramyaa;Bazhenov, Maxim
众所周知，人工神经网络会遭受灾难性的遗忘：当顺序学习多个任务时，它们在最近的任务上表现良好，而牺牲了以前学习的任务。在大脑中，睡眠通过重放最近和旧的冲突记忆痕迹，在增量学习中发挥着重要作用。在这里，我们测试了一个假设，即在人工神经网络中实现类似睡眠的阶段可以在新的训练过程中保护旧的记忆，并减轻灾难性的遗忘。睡眠被实现为离线训练，具有局部无监督的赫布可塑性规则和噪声输入。在增量学习框架中，睡眠能够恢复否则会忘记的旧任务。以前学习的记忆在睡眠中自发地重放，形成每类输入的独特表征。与旧任务相关的表征稀疏性和神经元活动增加，而与新任务相关的活动减少。这项研究表明，模拟睡眠动力学的自发重放可以减轻人工神经网络中的灾难性遗忘。众所周知，人工神经网络在最近学习的任务上表现良好，同时忘记了以前学习的任务。作者提出了一种无监督的睡眠回放算法来恢复旧任务的突触连接，这些突触连接可能在新任务训练后受损。
Artificial neural networks are known to suffer from catastrophic forgetting: when learning multiple tasks sequentially, they perform well on the most recent task at the expense of previously learned tasks. In the brain, sleep is known to play an important role in incremental learning by replaying recent and old conflicting memory traces. Here we tested the hypothesis that implementing a sleep-like phase in artificial neural networks can protect old memories during new training and alleviate catastrophic forgetting. Sleep was implemented as off-line training with local unsupervised Hebbian plasticity rules and noisy input. In an incremental learning framework, sleep was able to recover old tasks that were otherwise forgotten. Previously learned memories were replayed spontaneously during sleep, forming unique representations for each class of inputs. Representational sparseness and neuronal activity corresponding to the old tasks increased while new task related activity decreased. The study suggests that spontaneous replay simulating sleep-like dynamics can alleviate catastrophic forgetting in artificial neural networks. Artificial neural networks are known to perform well on recently learned tasks, at the same time forgetting previously learned ones. The authors propose an unsupervised sleep replay algorithm to recover old tasks synaptic connectivity that may have been damaged after new task training.
登录
查看更多内容
影响因子：
25
作者：
Ji, Daoyun;Wilson, Matthew A.
通讯作者：
Wilson, Matthew A.
影响因子：
2.9
作者：
Hayes, Tyler L.;Krishnan, Giri P.;Bazhenov, Maxim;Siegelmann, Hava T.;Sejnowski, Terrence J.;Kanan, Christopher
通讯作者：
Kanan, Christopher
深度神经网络： 生物视觉和大脑信息处理建模的新框架
DOI：
10.1146/annurev-vision-082114-035447
发表时间：
2015-01-01
期刊：
ANNUAL REVIEW OF VISION SCIENCE, VOL 1
影响因子：
--
作者：
Kriegeskorte, Nikolaus
通讯作者：
Kriegeskorte, Nikolaus
影响因子：
2.7
作者：
HENNEVIN, E;HARS, B;BLOCH, V
通讯作者：
BLOCH, V
影响因子：
7.7
作者：
Krishnan, Giri P.;Chauvette, Sylvain;Bazhenov, Maxim
通讯作者：
Bazhenov, Maxim