Riemannian Walk for Incremental Learning: Understanding Forgetting and Intransigence

By Arslan Chaudhry, Puneet Kumar Dokania, Thalaiyasingam Ajanthan, Philip H. S. Torr. 2018

1,405 citations recorded by the source corpus.

Incremental learning (IL) has received a lot of attention recently, however, the literature lacks a precise problem definition, proper evaluation settings, and metrics tailored specifically for the IL problem. One of the main objectives of this work is to fill these gaps so as to provide a common ground for better understanding of IL. The main challenge for an IL algorithm is to update the classifier whilst preserving existing knowledge. We observe that, in addition to forgetting, a known issue while preserving knowledge, IL also suffers from a problem we call intransigence, inability of a model to update its knowledge. We introduce two metrics to quantify forgetting and intransigence that allow us to understand, analyse, and gain better insights into the behaviour of IL algorithms. We present RWalk, a generalization of EWC++ (our efficient version of EWC [Kirkpatrick2016EWC]) and Path Integral [Zenke2017Continual] with a theoretically grounded KL-divergence based perspective. We provide a thorough analysis of various IL algorithms on MNIST and CIFAR-100 datasets. In these experiments, RWalk obtains superior results in terms of accuracy, and also provides a better trade-off between forgetting and intransigence.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Toward Understanding Catastrophic Forgetting in Continual Learning (2019, methodology): “Besides these two measures, we can also design other measures for properties such as intransigence [8].”
A Comprehensive Survey of Continual Learning: Theory, Method and Application (2023, methodology): “Besides, weight regularization can also be combined with experience replay for better performance and generality [44], [52].”
Progress & Compress: A scalable framework for continual learning (2018, background): “A large body of literature recognises the importance of the continual learning problem, and there has been some increased interest in the topic recently (e.g. Rusu et al., 2016a; Shin et al., 2017; Lopez-Paz et al., 2017; Nguyen et al., 2017; Kirkpatrick et al., 2017; Chaudhry et al., 2018).”
Continual Learning via Sequential Function-Space Variational Inference (2023, methodology): “A popular family of objectives penalizes changes in parameters from one task to another (Ahn et al., 2019; Aljundi et al., 2018; Chaudhry et al., 2018; Kirkpatrick et al., 2017; Lee et al., 2017; Liu et al., 2018; Loo et al., 2020; Nguyen et al., 2018; Park et al., 2019; Ritter et al., 2018;…”
Incorporating neuro-inspired adaptability for continual learning in artificial intelligence (2023, methodology): “In particular, since the loss functions of such methods [7–10] typically have a similar form and differ only in the metric for estimating the parameter importance [49] (Method 4.1, Eq.”
A soft nearest-neighbor framework for continual semi-supervised learning (2022, neutral): “The first category comprises regularization-based methods [13,14,20,22,24,25,30,45,54], which address the problem of catastrophic forgetting by introducing appropriate regularization terms in the objective function or identifying a set of parameters that are most relevant for certain tasks.”
On the Effectiveness of Lipschitz-Driven Rehearsal in Continual Learning (2022, neutral): “W.r.t. the latter, TIL has been criticized as a less challenging and realistic benchmark [26, 74, 4]; we therefore conduct our main experiments on state-of-the-art CL models in the CIL setting 1 .”
Class-Incremental Learning with Cross-Space Clustering and Controlled Transfer (2022, methodology): “These methods include EWC [25], SI [58], MAS [4] and RWalk [9].”
Online Continual Learning with Contrastive Vision Transformer (2022, methodology): “To compare the alleviating forgetting capability, we assess the average forgetting [9,12] that measures the performance degradation in subsequent tasks.”
Incremental Task Learning with Incremental Rank Updates (2022, methodology): “Incremental task learning is a subcategory of continual learning or lifelong learning approaches aim to address the problem of catastrophic forgetting by adapting the network or training process to learn new tasks without forgetting the previously learned ones [23,16,3,2,4,6,28,29,12].”
Class-incremental Novel Class Discovery (2022, background): “The IL meth-ods are designed so as to prevent catastrophic forgetting [13] of the model on the old tasks and at the same time flexible enough to learn on new tasks [6].”
CoSCL: Cooperation of Small Continual Learners is Stronger than a Big One (2022, neutral): “8) RWALK [7] λ(128) λ(6) λ(48) λ(5) P&C [40] λ(40000) λ(20000) λ(1) λ(80) ∗AFEC [45] λe(1) λe(1) λe(0.”
Continual Learning through Retrieval and Imagination (2022, background): “Second, to compare the preventing forgetting capability, we assess the average forgetting (Chaudhry et al. 2018) that measures the performance degradation in subsequent tasks.”
CLiMB: A Continual Learning Benchmark for Vision-and-Language Tasks (2022, methodology): “This property of models to learn new tasks poorly in a continual learning setting is also called intransigence [Chaudhry et al., 2018].”
NISPA: Neuro-Inspired Stability-Plasticity Adaptation for Continual Learning in Sparse Networks (2022, background): “Regularization-based CL approaches modulate gradient updates, aiming to identify and preserve the weights that are more important for each task (Kirkpatrick et al., 2017; Zenke et al., 2017; Aljundi et al., 2018; Li & Hoiem, 2017; Serra et al., 2018; Chaudhry et al., 2018).”
Transfer without Forgetting (2022, methodology): “Regularization-based methods [31,37,13,12] typically identify subsets of weights that are highly functional for the representations of previous tasks, with the purpose to prevent their drastic modification through apposite optimization constraints.”