readme

IntelLabs · Aug 23, 2021 · 69bbd23 · 69bbd23
1 parent 37cc498
commit 69bbd23
Showing 1 changed file with 2 additions and 2 deletions.
diff --git a/CLOC/README.md b/CLOC/README.md
@@ -15,9 +15,9 @@ We perform an empirical study for online visual continual learning with natural
 
 1. We create a large scale continual learning benchmark called Continual LOCalization (CLOC), which contains roughly 39 million images taken from 8 years. We construct a continual classification task using the geo-locations of each image. CLOC is suitable for benchmarking continual learning due to its large scale and natural distribution shifts. We provide the code to construct CLOC here.
 
-2. We define online continual learning on CLOC, and distinguish learning efficacy (how fast can the model adapt to new data) from information retention (how well can the model remember old knowledge). Previous studies mostly focuses on offline continual learning, and optimizes for information retention. In this paper, we show that learning efficacy and information retention are somewhat conflicting, and propose serveral strategies to optimize learning efficacy.
+2. We define online continual learning (OCL) on CLOC, and distinguish learning efficacy (how fast can the model adapt to new data) from information retention (how well can the model remember old knowledge). Previous studies mostly focus on offline continual learning, and optimize for information retention. In this paper, we show that learning efficacy and information retention are somewhat conflicting, and propose serveral strategies to optimize learning efficacy.
 
-With the proposed strategies, our online continual learning model can perform on-par or better than supervised learning models given similar budgets.
+We show that the average online accuracy of our OCL model can perform on-par or better than the validation accuracy of supervised learning (SL) models, given similar budgets, as shown in the bottom right figure above.
 
 The code in this repo is free for non-commercial academic use. Any commercial use is strictly 
 prohibited without the authors' consent. Please acknowledge the authors by citing: