Meta-Learning with Neural Bandit Scheduler

Authors: Yunzhe Qi, Yikun Ban, Tianxin Wei, Jiaru Zou, Huaxiu Yao, Jingrui He

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

Reproducibility Variable Result LLM Response
Research Type Experimental Theoretical analysis and extensive experiments are presented to show the effectiveness of our proposed framework. To demonstrate the effectiveness of BASS, we compare our method against seven strong baselines on three real data sets, with different specifications. In addition, complementary experiments as well as a case study on Ensemble Inference are also provided to better understand the property and behavior of BASS.
Researcher Affiliation Academia Yunzhe Qi University of Illinois at Urbana-Champaign Champaign, IL yunzheq2@illinois.edu Yikun Ban University of Illinois at Urbana-Champaign Champaign, IL yikunb2@illinois.edu Tianxin Wei University of Illinois at Urbana-Champaign Champaign, IL twei10@illinois.edu Jiaru Zou University of Illinois at Urbana-Champaign Champaign, IL jiaruz2@illinois.edu Huaxiu Yao University of North Carolina at Chapel Hill Chapel Hill, NC huaxiu@cs.unc.edu Jingrui He University of Illinois at Urbana-Champaign Champaign, IL jingrui@illinois.edu
Pseudocode Yes Algorithm 1 BAndit Ta Sk Scheduler (BASS)
Open Source Code Yes Code will be made available at https://github.com/yunzhe0306/Bandit_Task_Scheduler.
Open Datasets Yes We adopt Drug [36], Mini-Image Net (M-Image Net) [42] and CIFAR-100 (CIFAR) [25] data sets under the few-shot learning scenario.
Dataset Splits Yes For the data partitioning, we have the Mini-Image Net and CIFAR-100 data sets divided into the partitions 64 : 16 : 20, which correspond to the training set, validation set and the testing set respectively. ... Then, for the Drug data set, we partition the tasks into 4100 : 76 : 100 representing the training set, validation set and the testing set.
Hardware Specification Yes All the experiments are performed on a Linux machine with Intel Xeon CPU, 128GB RAM, and Tesla V100 GPU.
Software Dependencies No ANIL [40] is adopted as the backbone meta-learning framework. (No version specified). No other specific software dependencies with version numbers are mentioned.
Experiment Setup Yes Input: Task distribution P(T ). Iterations K. GD steps J. Number of chosen tasks B. Learning rates for meta-model η1, η2. Learning rates for BASS ηθ 1 , ηθ 2 . Exploration coefficient α (0, 1]. For our BASS, we apply two 2-layer FC networks for f1( ; θ1), f2( ; θ2) respectively, and set network width m = 200... For the learning rate, we find the learning rate for BASS with grid search from {0.01, 0.001, 0.0001}, and choose the learning rates for the meta-model η1 = 0.01, η2 = 0.001.