AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

UniT: Multimodal Multitask Learning With a Unified Transformer

Ronghang Hu, Amanpreet Singh

ICCV 2021 · Architectures and efficiency › Attention and transformers · approach: Transformer block and depth design · 336 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
joint training with separate decoders instead of single-task trainingUniT · VQAv2 (COCO+VG+VQAv2)accuracy: 66.38 -> 68.84Improved
share one decoder across tasks instead of separate decodersUniT · COCO det.mAP: 40.8 -> 37.2Hurt
initialize shared-decoder joint model from COCO single-task modelUniT shared decoder · COCO det.mAP: 37.2 -> 40.8Improved
add VG detection and COCO detection together to VQAv2 trainingUniT shared (COCO init.) · VG det.mAP: 3.83 -> 4.53Improved
use VG detection instead of COCO detection as the detection partner for VQAv2UniT shared (COCO init.) · VQAv2accuracy: 66.88 -> 68.49Improved
joint multi-task training on 8 datasets instead of single-task trainingUniT shared (COCO init.) · SNLI-VEaccuracy: 70.52 -> 73.16Improved
joint multi-task training on 8 datasets instead of single-task trainingUniT shared (COCO init.) · COCO det.mAP: 40.6 -> 39.0Hurt
per-task finetuning of shared modelUniT · COCO det.mAP: 39.0 -> 42.3Improved
per-task finetuning of shared modelUniT · SST-2accuracy: 89.29 -> 88.06Hurt
increase decoder layers to 8 (ablation)UniT · COCO det. (ablation)mAP: 38.79 -> 40.13Improved
double decoder layers to 12 with 400k iterations (ablation)UniT · COCO det. (ablation)mAP: 38.79 -> 39.02No clear effect
use larger decoder hidden size 768 instead of 256 (ablation)UniT · COCO det. (ablation)mAP: 36.32 -> 38.79Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy