Direction-oriented Multi-objective Learning: Simple and Provable Stochastic Algorithms

Authors: Peiyao Xiao, Hao Ban, Kaiyi Ji

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that our methods achieve competitive or improved performance compared to existing gradient manipulation approaches in a series of tasks on multi-task supervised learning and reinforcement learning.
Researcher Affiliation Collaboration 1Department of CSE, University at Buffalo 2 Zuoyi Technology
Pseudocode Yes Algorithm 1 Stochastic Direction-oriented Multi-objective Gradient descent (SDMGrad)
Open Source Code Yes Code is available at https://github.com/ml-opt-lab/sdmgrad.
Open Datasets Yes For the supervised learning setting, we evaluate the performance on the Cityscapes [39] and NYUv2 [40] datasets. ... For the reinforcement learning setting, we evaluate the performance on the MT10 benchmarks, which include 10 robot manipulation tasks under the Meta-World environment [43].
Dataset Splits No The paper mentions training data and test data for Multi-Fashion+MNIST ('120000 training images and 20000 test images'), but does not provide explicit train/validation/test splits or ratios for all datasets used, nor for the validation set specifically.
Hardware Specification Yes The experiments on Cityscapes and NYU-v2 are run on RTX 3090 and Tesla V100 GPU, respectively. ... All experiments on MT10 are run on RTX 2080Ti GPU.
Software Dependencies No The paper mentions using 'the library released by [10]' and 'MTRL codebase[49]' but does not provide specific version numbers for these or other key software dependencies like Python, PyTorch, or TensorFlow.
Experiment Setup Yes We search the hyperparameter λ {0.1, 0.2, , 1.0} for our SDMGrad method... We train our method for 200 epochs, using Adam optimizer with learning rate 0.0001 for the first 100 epochs and 0.00005 for the rest. The batch size for Cityscapes and NYU-v2 are 8 and 2 respectively.