Mitigating Gradient Bias in Multi-objective Learning: A Provably Convergent Approach

Authors: Heshan Devaka Fernando, Han Shen, Miao Liu, Subhajit Chaudhury, Keerthiram Murugesan, Tianyi Chen

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

Reproducibility Variable Result LLM Response
Research Type Experimental Simulations on supervised and reinforcement learning demonstrate the effectiveness of our method relative to state-of-the-art methods.
Researcher Affiliation Collaboration 1Rensselaer Polytechnic Institute 2IBM Thomas J. Watson Research Center
Pseudocode Yes Algorithm 1 Mo Co: Stochastic Multi-objective gradient with Correction
Open Source Code No The paper mentions using a third-party codebase (MTRL codebase) but does not provide specific links or explicit statements for its own implementation of the described methodology.
Open Datasets Yes NYU-v2 (Silberman et al., 2012) and City Scapes (Cordts et al., 2015) datasets. ... Office-31 (Saenko et al., 2010) and Office-home(Venkateswara et al., 2017) datasets. ... Met-world environment (Yu et al., 2020b).
Dataset Splits Yes As per the implementation in (Lin & Zhang, 2022), 60% of the total dataset is used for training, 20% for validation, and the rest 20% for testing.
Hardware Specification No The paper does not provide specific details about the hardware used, such as GPU or CPU models. It only mentions funding sources.
Software Dependencies No The paper mentions software components like 'MTRL codebase', 'soft actor-critic (SAC)', 'Adam', 'Seg Net', 'MTAN', and 'Res Net18' but does not specify their version numbers.
Experiment Setup Yes For all the algorithms, we use the initial learning rate of 0.001, exponentially decaying at a rate of 0.05. In this example for Mo Co, we use βk = 5/k0.5, where k is the number of iterations. ... All the MTL methods in comparison are trained for 200 epochs, using a batch size of 8. We use Adam as the optimizer with a learning rate of 0.0001 for the first 100 epochs, and with a learning rate of 0.00005 for the rest of the epochs. ... All the methods are trained for 2 million steps with a batch size of 1280. ... Table 10: Summary of hyper-parameter choices for Mo Co in each experiment.