Abstract:We introduce DataComp for Language Models (DCLM), a testbed for controlled dataset experiments with the goal of improving language models. As part of DCLM, we provide a standardized corpus of 240T tokens extracted from Common Crawl, effective pretraining recipes based on the OpenLM framework, and a broad suite of 53 downstream evaluations. Participants in the DCLM benchmark can experiment with data curation strategies such as deduplication, filtering, and data mixing at model scales ranging from 412M to 7B parameters. As a baseline for DCLM, we conduct extensive experiments and find that model-based filtering is key to assembling a high-quality training set. The resulting dataset, DCLM-Baseline enables training a 7B parameter language model from scratch to 64% 5-shot accuracy on MMLU with 2.6T training tokens. Compared to MAP-Neo, the previous state-of-the-art in open-data language models, DCLM-Baseline represents a 6.6 percentage point improvement on MMLU while being trained with 40% less compute. Our baseline model is also comparable to Mistral-7B-v0.3 and Llama 3 8B on MMLU (63% & 66%), and performs similarly on an average of 53 natural language understanding tasks while being trained with 6.6x less compute than Llama 3 8B. Our results highlight the importance of dataset design for training language models and offer a starting point for further research on data curation.
Abstract:Deep learning is closing the gap with humans on several object recognition benchmarks. Here we investigate this gap in the context of challenging images where objects are seen from unusual viewpoints. We find that humans excel at recognizing objects in unusual poses, in contrast with state-of-the-art pretrained networks (EfficientNet, SWAG, ViT, SWIN, BEiT, ConvNext) which are systematically brittle in this condition. Remarkably, as we limit image exposure time, human performance degrades to the level of deep networks, suggesting that additional mental processes (requiring additional time) take place when humans identify objects in unusual poses. Finally, our analysis of error patterns of humans vs. networks reveals that even time-limited humans are dissimilar to feed-forward deep networks. We conclude that more work is needed to bring computer vision systems to the level of robustness of the human visual system. Understanding the nature of the mental processes taking place during extra viewing time may be key to attain such robustness.
Abstract:Utilizing massive web-scale datasets has led to unprecedented performance gains in machine learning models, but also imposes outlandish compute requirements for their training. In order to improve training and data efficiency, we here push the limits of pruning large-scale multimodal datasets for training CLIP-style models. Today's most effective pruning method on ImageNet clusters data samples into separate concepts according to their embedding and prunes away the most prototypical samples. We scale this approach to LAION and improve it by noting that the pruning rate should be concept-specific and adapted to the complexity of the concept. Using a simple and intuitive complexity measure, we are able to reduce the training cost to a quarter of regular training. By filtering from the LAION dataset, we find that training on a smaller set of high-quality data can lead to higher performance with significantly lower training costs. More specifically, we are able to outperform the LAION-trained OpenCLIP-ViT-B32 model on ImageNet zero-shot accuracy by 1.1p.p. while only using 27.7% of the data and training compute. Despite a strong reduction in training cost, we also see improvements on ImageNet dist. shifts, retrieval tasks and VTAB. On the DataComp Medium benchmark, we achieve a new state-of-the-art ImageNet zero-shot accuracy and a competitive average zero-shot accuracy on 38 evaluation tasks.
Abstract:Vision-Language Models (VLMs) are pretrained on large, diverse, and noisy web-crawled datasets. This underscores the critical need for dataset pruning, as the quality of these datasets is strongly correlated with the performance of VLMs on downstream tasks. Using CLIPScore from a pretrained model to only train models using highly-aligned samples is one of the most successful methods for pruning.We argue that this approach suffers from multiple limitations including: 1) false positives due to spurious correlations captured by the pretrained CLIP model, 2) false negatives due to poor discrimination between hard and bad samples, and 3) biased ranking towards samples similar to the pretrained CLIP dataset. We propose a pruning method, SIEVE, that employs synthetic captions generated by image-captioning models pretrained on small, diverse, and well-aligned image-text pairs to evaluate the alignment of noisy image-text pairs. To bridge the gap between the limited diversity of generated captions and the high diversity of alternative text (alt-text), we estimate the semantic textual similarity in the embedding space of a language model pretrained on billions of sentences. Using DataComp, a multimodal dataset filtering benchmark, we achieve state-of-the-art performance on the large scale pool, and competitive results on the medium scale pool, surpassing CLIPScore-based filtering by 1.7% and 2.6% on average, on 38 downstream tasks.
Abstract:Progress in machine learning has been driven in large part by massive increases in data. However, large web-scale datasets such as LAION are largely uncurated beyond searches for exact duplicates, potentially leaving much redundancy. Here, we introduce SemDeDup, a method which leverages embeddings from pre-trained models to identify and remove semantic duplicates: data pairs which are semantically similar, but not exactly identical. Removing semantic duplicates preserves performance and speeds up learning. Analyzing a subset of LAION, we show that SemDeDup can remove 50% of the data with minimal performance loss, effectively halving training time. Moreover, performance increases out of distribution. Also, analyzing language models trained on C4, a partially curated dataset, we show that SemDeDup improves over prior approaches while providing efficiency gains. SemDeDup provides an example of how simple ways of leveraging quality embeddings can be used to make models learn faster with less data.
Abstract:Deep networks should be robust to rare events if they are to be successfully deployed in high-stakes real-world applications (e.g., self-driving cars). Here we study the capability of deep networks to recognize objects in unusual poses. We create a synthetic dataset of images of objects in unusual orientations, and evaluate the robustness of a collection of 38 recent and competitive deep networks for image classification. We show that classifying these images is still a challenge for all networks tested, with an average accuracy drop of 29.5% compared to when the objects are presented upright. This brittleness is largely unaffected by various network design choices, such as training losses (e.g., supervised vs. self-supervised), architectures (e.g., convolutional networks vs. transformers), dataset modalities (e.g., images vs. image-text pairs), and data-augmentation schemes. However, networks trained on very large datasets substantially outperform others, with the best network tested$\unicode{x2014}$Noisy Student EfficentNet-L2 trained on JFT-300M$\unicode{x2014}$showing a relatively small accuracy drop of only 14.5% on unusual poses. Nevertheless, a visual inspection of the failures of Noisy Student reveals a remaining gap in robustness with the human visual system. Furthermore, combining multiple object transformations$\unicode{x2014}$3D-rotations and scaling$\unicode{x2014}$further degrades the performance of all networks. Altogether, our results provide another measurement of the robustness of deep networks that is important to consider when using them in the real world. Code and datasets are available at https://github.com/amro-kamal/ObjectPose.