RMM: Reinforced Memory Management for Class-Incremental Learning
Authors: Yaoyao Liu, Bernt Schiele, Qianru Sun
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | For evaluation, we plug RMM into two top-performing baselines (LUCIR+AANets and POD+AANets [30]) and conduct experiments on three benchmarks (CIFAR-100, Image Net-Subset, and Image Net-Full). Our results show clear improvements, e.g., boosting POD+AANets by 3.6%, 4.4%, and 1.9% in the 25-Phase settings of the above benchmarks, respectively. |
| Researcher Affiliation | Academia | 1Max Planck Institute for Informatics, Saarland Informatics Campus 2School of Computing and Information Systems, Singapore Management University {yaoyao.liu, schiele}@mpi-inf.mpg.de qianrusun@smu.edu.sg |
| Pseudocode | Yes | Algorithm 1 Learning policy functions in RMM |
| Open Source Code | Yes | The code is available at https://class-il.mpi-inf.mpg.de/rmm/. |
| Open Datasets | Yes | CIFAR-100 [24] contains 60, 000 samples of 32 32 color images from 100 classes. There are 500 training and 100 test samples for each class. Image Net (ILSVRC 2012) [44] contains around 1.3 million samples of 224 224 color images from 1, 000 classes. |
| Dataset Splits | Yes | When building the tasks, we randomly choose 10% training samples of each class (from D0) to compose a pseudo validation set (note that we are not allowed to use the original validation set in training). |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers, such as programming language versions, library versions, or specific solver versions. |
| Experiment Setup | Yes | On CIFAR-100 (Image Net-Subset/Full), we train it for 160 (90) epochs in each phase, and divide the learning rate by 10 after 80 (30) and then after 120 (60) epochs. Then, we fine-tune the model for 20 epochs using only exemplars (including the preserved exemplars of the new data to be used in future phases). We use an SGD optimizer and an ADAM optimizer for the classification model and policy functions, respectively. |