Self-Guided Masked Autoencoder

Authors: Jeongwoo Shin, Inseo Lee, Junho Lee, Joonseok Lee

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Comprehensive experiments on various downstream tasks verify the effectiveness of the proposed method.
Researcher Affiliation Collaboration Jeongwoo Shin1, Inseo Lee1, Junho Lee1, Joonseok Lee1,2 1Seoul National University, 2Google Research
Pseudocode No No explicit pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We plan to publicly provide our code used in this paper.
Open Datasets Yes We pre-train all competing models for 400 epochs on Image Net-1K [12], and fine-tune on 3 downstream tasks: image classification, object detection, and semantic segmentation. [...] We use CIFAR-100 [30], i Naturalist 2019 [48], and CUB200-2011 [49] for image classification. We fine-tune our model on COCO [36] for object detection, and on ADE20K [62] for semantic segmentation.
Dataset Splits Yes all experiments have been conducted on 10% of Image Net-1K training set, unless noted otherwise. [...] We additionally measure the feature variance (σF ) and variance of the pairwise similarities (σS), on the Image Net-1K validation set:
Hardware Specification Yes We conduct experiments on 8 NVidia A6000 GPUs (48GB).
Software Dependencies No The paper does not specify particular software dependencies with version numbers.
Experiment Setup Yes We pre-train all competing models for 400 epochs on Image Net-1K [12] [...] We fine-tune a Mask R-CNN model [24] end-to-end on COCO with a Vi T backbone for 90K iterations [...] We fix the masking ratio to 0.75 for all the experiments in this section.