Showing papers by "Omer Levy published in 2023"

Scaling Laws for Generative Mixed-Modal Language Models

Abstract: Generative language models define distributions over sequences of tokens that can represent essentially any combination of data modalities (e.g., any permutation of image tokens from VQ-VAEs, speech tokens from HuBERT, BPE tokens for language or code, and so on). To better understand the scaling properties of such mixed-modal models, we conducted over 250 experiments using seven different modalities and model sizes ranging from 8 million to 30 billion, trained on 5-100 billion tokens. We report new mixed-modal scaling laws that unify the contributions of individual modalities and the interactions between them. Specifically, we explicitly model the optimal synergy and competition due to data and model size as an additive term to previous uni-modal scaling laws. We also find four empirical phenomena observed during the training, such as emergent coordinate-ascent style training that naturally alternates between modalities, guidelines for selecting critical hyper-parameters, and connections between mixed-modal competition and training stability. Finally, we test our scaling law by training a 30B speech-text model, which significantly outperforms the corresponding unimodal models. Overall, our research provides valuable insights into the design and training of mixed-modal generative models, an important new class of unified models that have unique distributional properties.

Pick-a-Pic: An Open Dataset of User Preferences for Text-to-Image Generation

Abstract: The ability to collect a large dataset of human preferences from text-to-image users is usually limited to companies, making such datasets inaccessible to the public. To address this issue, we create a web app that enables text-to-image users to generate images and specify their preferences. Using this web app we build Pick-a-Pic, a large, open dataset of text-to-image prompts and real users' preferences over generated images. We leverage this dataset to train a CLIP-based scoring function, PickScore, which exhibits superhuman performance on the task of predicting human preferences. Then, we test PickScore's ability to perform model evaluation and observe that it correlates better with human rankings than other automatic evaluation metrics. Therefore, we recommend using PickScore for evaluating future text-to-image generation models, and using Pick-a-Pic prompts as a more relevant dataset than MS-COCO. Finally, we demonstrate how PickScore can enhance existing text-to-image models via ranking.

LIMA: Less Is More for Alignment

Abstract: Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback. Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

Correspondence between the layered structure of deep language models and temporal structure of natural language processing in the human brain

Abstract: Deep language models (DLMs) provide a novel computational paradigm for how the brain processes natural language. Unlike symbolic, rule-based models described in psycholinguistics, DLMs encode words and their context as continuous numerical vectors. These “embeddings” are constructed by a sequence of computations organized in “layers” to ultimately capture surprisingly sophisticated representations of linguistic structures. How does this layered hierarchy map onto the human brain during natural language comprehension? In this study, we used electrocorticography (ECoG) to record neural activity in language areas along the superior temporal gyrus and inferior frontal gyrus while human participants listened to a 30-minute spoken narrative. We supplied this same narrative to a high-performing DLM (GPT2-XL) and extracted the contextual embeddings for each word in the story across all 48 layers of the model. We next trained a set of linear encoding models to predict the temporally-evolving neural activity from the embeddings at each layer. We found a striking correspondence between the layer-by-layer sequence of embeddings from GPT2-XL and the temporal sequence of neural activity in language areas. In addition, we found evidence for the gradual accumulation of recurrent information along the linguistic processing hierarchy. However, we also noticed additional neural processes in the brain, but not in DLMs, during the processing of surprising (unpredictable) words. These findings point to a connection between human language processing and DLMs where the layer-by-layer accumulation of contextual information in DLM embeddings matches the temporal dynamics of neural activity in high-order language areas.

ZeroSCROLLS: A Zero-Shot Benchmark for Long Text Understanding

Abstract: We introduce ZeroSCROLLS, a zero-shot benchmark for natural language understanding over long texts, which contains only test sets, without training or development data. We adapt six tasks from the SCROLLS benchmark, and add four new datasets, including two novel information fusing tasks, such as aggregating the percentage of positive reviews. Using ZeroSCROLLS, we conduct a comprehensive evaluation of both open-source and closed large language models, finding that Claude outperforms ChatGPT, and that GPT-4 achieves the highest average score. However, there is still room for improvement on multiple open challenges in ZeroSCROLLS, such as aggregation tasks, where models struggle to pass the naive baseline. As the state of the art is a moving target, we invite researchers to evaluate their ideas on the live ZeroSCROLLS leaderboard

X&Fuse: Fusing Visual Information in Text-to-Image Generation

Abstract: We introduce X&Fuse, a general approach for conditioning on visual information when generating images from text. We demonstrate the potential of X&Fuse in three different text-to-image generation scenarios. (i) When a bank of images is available, we retrieve and condition on a related image (Retrieve&Fuse), resulting in significant improvements on the MS-COCO benchmark, gaining a state-of-the-art FID score of 6.65 in zero-shot settings. (ii) When cropped-object images are at hand, we utilize them and perform subject-driven generation (Crop&Fuse), outperforming the textual inversion method while being more than x100 faster. (iii) Having oracle access to the image scene (Scene&Fuse), allows us to achieve an FID score of 5.03 on MS-COCO in zero-shot settings. Our experiments indicate that X&Fuse is an effective, easy-to-adapt, simple, and general approach for scenarios in which the model may benefit from additional visual information.

Translating Akkadian to English with neural machine translation

Abstract: Abstract Cuneiform is one of the earliest writing systems in recorded human history (ca. 3,400 BCE–75 CE). Hundreds of thousands of such texts were found over the last two centuries, most of which are written in Sumerian and Akkadian. We show the high potential in assisting scholars and interested laypeople alike, by using natural language processing (NLP) methods such as convolutional neural networks (CNN), to automatically translate Akkadian from cuneiform Unicode glyphs directly to English (C2E) and from transliteration to English (T2E). We show that high-quality translations can be obtained when translating directly from cuneiform to English, as we get 36.52 and 37.47 Best Bilingual Evaluation Understudy 4 (BLEU4) scores for C2E and T2E, respectively. For C2E, our model is better than the translation memory baseline in 9.43, and for T2E, the difference is even higher and stands at 13.96. The model achieves best results in short- and medium-length sentences (c. 118 or less characters). As the number of digitized texts grows, the model can be improved by further training as part of a human-in-the-loop system which corrects the results.

Vision Transformers with Mixed-Resolution Tokenization

Abstract: Vision Transformer models process input images by dividing them into a spatially regular grid of equal-size patches. Conversely, Transformers were originally introduced over natural language sequences, where each token represents a subword - a chunk of raw data of arbitrary size. In this work, we apply this approach to Vision Transformers by introducing a novel image tokenization scheme, replacing the standard uniform grid with a mixed-resolution sequence of tokens, where each token represents a patch of arbitrary size. Using the Quadtree algorithm and a novel saliency scorer, we construct a patch mosaic where low-saliency areas of the image are processed in low resolution, routing more of the model's capacity to important image regions. Using the same architecture as vanilla ViTs, our Quadformer models achieve substantial accuracy gains on image classification when controlling for the computational budget. Code and models are publicly available at https://github.com/TomerRonen34/mixed-resolution-vit .