Human infant learning integrates information across senses -- sight, sound, touch, etc. -- but current state of the art machine learning models usually use only one of these types. It remains to be seen whether integrating data across modes is necessary for achieving human-level intelligence.

In contemporary machine learning (ML) research, we are mostly interested in image, text, graph, and video data. State of the art models in each of these domains train only on inputs of that specific domain; let's call this uni-modal training. By extension, if a model were to train on two or more of these input types, while evaluating on only one, we'll call that multi-modal training with uni-modal evaluation. For the purposes of this question, we are only interested in uni-modal evaluation tasks, so robotics and driving benchmarks are out of the question.

Question Description: When will a multi-modal trained model out-perform the previous state of the art on one of the following uni-modal benchmarks:

1. ImageNet
2. WikiText-103
3. Cityscapes
4. Additional uni-modal benchmarks from paperswithcode.com may be added to reflect trends in machine learning research. I will review paperswithcode.com two and four years after this question opens to request that moderators add the two most popular benchmarks which have more new entries (since June 1, 2020) than at least two thirds of the above benchmarks. If one of the newly added benchmarks involves data of the same type as one of the above benchmarks (i.e. image classification, text, image segmentation), and has more new entries, then the old benchmark will be superseded, and removed from the list.

Resolution Condition: This question resolves as the first date on which one of the benchmarks above has a #1 ranked paper which sets the record using a multi-modal trained model. If no such paper is listed before 2030, then the question resolves as >01/01/2030.

Specifics and Caveats:

1. Multi-modal pre-training counts towards resolution.

2. For text tasks, training on video counts if, and only if the image stream is used -- i.e. not just the audio stream.

3. For image tasks, training on video counts if, and only if the audio stream is used -- i.e. not just the image stream.

4. If paperswithcode.com shuts down or permanently stops updating their data, then the question resolves as ambiguous.