TriRE: A Multi-Mechanism Learning Paradigm for Continual Knowledge Retention and Promotion

Authors: Preetha Vijayan, Prashant Bhat, Bahram Zonooz, Elahe Arani

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental Setup: We expand the Mammoth CL repository in Py Torch [9]. On the basis of Class-IL and Task-IL scenarios, we assess the existing CL techniques against the proposed one. Experimental Results: We compare Tri RE with contemporary rehearsal-based and weight regularization methods in Class-IL and Task-IL settings. As shown in Table 1, Tri RE consistently outperforms
Researcher Affiliation Collaboration Preetha Vijayan1, , Prashant Bhat2,3,*, Bahram Zonooz2,3, , Elahe Arani2, 1Nav Info Europe 2Eindhoven University of Technology (TU/e) 3Tom Tom
Pseudocode Yes Algorithm 1 Proposed Approach Tri RE
Open Source Code Yes Code is available at https://github.com/Neur AI-Lab/Tri RE
Open Datasets Yes To evaluate the performance of our method in Task-IL and Class-IL scenarios, we employ three different datasets: Seq-CIFAR10, Seq-CIFAR100, and Seq-Tiny Image Net. These datasets are derived from CIFAR10, CIFAR100, and Tiny Image Net, respectively.
Dataset Splits Yes The hyperparameters required to replicate the results of Tri RE can be found in Table 5. These hyperparameters were determined through a tuning process involving different random initializations and a small portion of the training set reserved for validation.
Hardware Specification Yes Table 7 specifically lists the training times required to learn a single task (the first) for 3 epochs on an NVIDIA RTX 2080 Ti for Seq-CIFAR100 dataset with a buffer size of 200.
Software Dependencies No The paper mentions using 'Py Torch' but does not specify a version number for it or any other software dependencies.
Experiment Setup Yes The hyperparameters required to replicate the results of Tri RE can be found in Table 5... All experiments were conducted using a batch size of 32 and trained for 50 epochs. Tri RE was optimized using the Adam optimizer [24] implemented in Py Torch. Furthermore, the number of epochs allocated to each phase specified in Algorithm 1 was consistently set at a ratio of E1 : E2 : E3 = 3 : 1 : 1.