4 ms·

Are there some well-known examples of success in it?

• thethimble 1y ago
  Vision transformers effectively encode a grid of pixel patches. It’s ultimately a matter of ensuring the position encoding incorporates both X and Y and position. For LLMs we only have one axis of position and - more importantly - the vast majority of training data only is oriented in this way.