The extent to which text-only language models (LMs) learn to represent features of the non-linguistic world is an open question. Prior work has shown that pretrained LMs can be taught to caption images when a vision model's parameters are optimized to encode images in the language space. We test a stronger hypothesis: that the conceptual representations learned by frozen text-only models and vision-only models are similar enough that this can be achieved with a linear map. We show that the image representations from vision models can be transferred as continuous prompts to frozen LMs by training only a single linear projection. Using these to prompt the LM achieves competitive performance on captioning and visual question answering tasks compared to models that tune both the image encoder and text decoder (such as the MAGMA model). We compare three image encoders with increasing amounts of linguistic supervision seen during pretraining: BEIT (no linguistic information), NF-ResNET (lexical category information), and CLIP (full natural language descriptions). We find that all three encoders perform equally well at transferring visual property information to the language model (e.g., whether an animal is large or small), but that image encoders pretrained with linguistic supervision more saliently encode category information (e.g., distinguishing hippo vs. elephant) and thus perform significantly better on benchmark language-and-vision tasks. Our results indicate that LMs encode conceptual information structurally similarly to vision-based models, even those that are solely trained on images. Code is available here: https://github.com/jmerullo/limber
Metadata: arXiv. Abstracts retain their original authors’ rights.
In the conversation
Original HN comments, with labels from simple text rules. A label describes the whole comment and may miss context or irony. Read the discussion to make up your own mind. About the labels
>the sounds for man, and the word man can be mapped. The concept of man can be mapped from English to Chinese. Given enough of these clues, and a large enough data set it's not shocking that models get decent at inference between languages.
1. Many such concepts cannot be directly mapped between different languages especially with distant language pairs.
2. You can connect linguistic and visual models with a single linear projection layer. That's it. Entirely different modalities
https://arxiv.org/abs/2209.15162
https://arxiv.org/abs/2304.08485
I think you are severely underestimating the extent to which representations can group in neural networks.
They're mammals, they're social animals, they see. You're assuming a level of alienness we don't have the knowledge or understanding to truly ascertain.
Slightly off-topic, but I am just starting to learn about the field, and saw
> GPT4 still projects information into some highly non-linear latent space and samples from that
Do you have more information/know where I can find things to read about the "highly non-linear" part of that? I've been reading some stuff[1][2][3] about smaller models, and my impression has been that the latent space is shockingly linear for those ones. But counterexamples would be informative, especially if it's something that only happens with larger models.
-----
[1] https://arxiv.org/abs/2209.15162
[2] https://www.neelnanda.io/mechanistic-interpretability/othell...
[3] https://arxiv.org/pdf/1610.01644.pdf
The hypothesis that you can't learn some things from text - you need real life experience, is intuitive and I used to think it's true. But there are interesting results from just a few days ago saying that text by itself is also enough:
> We test a stronger hypothesis: that the conceptual representations learned by text only models are functionally equivalent (up to a linear transformation) to those learned by models trained on vision tasks. Specifically, we show that the image representations from vision models can be transferred as continuous prompts to frozen LMs by training only a single linear projection.
Linearly Mapping from Image to Text Space - https://arxiv.org/abs/2209.15162
There's something a bit more mindblowing than that. Language models and vision models learn representations so similar that you can connect them with just a linear projection between image embedding and text embedding space(no training of the image encoder or llm required).
https://arxiv.org/abs/2209.15162
https://llava-vl.github.io/
LLMs are already being grounded.
as well as this - https://llava-vl.github.io/, Just found this paper that demonstrated this a few months ago (that somehow language and vision models learn representations similar enough that linear projection is enough) https://arxiv.org/abs/2209.15162
Artificial intelligence (AI) researchers have been developing and refining large language models (LLMs) that exhibit remarkable capabilities across a variety of domains and tasks, challenging our understanding of learning and cognition.
Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task.
The Fermi paradox is the conflict between an expectation of a high {\em ex ante} probability of intelligent life elsewhere in the universe and the apparently lifeless universe we in fact observe.
Despite their massive size, successful deep artificial neural networks can exhibit a remarkably small difference between training and test performance.