Embedding models, which transform inputs into dense vector representations, are pivotal for capturing semantic information across various domains and modalities. Text embedding models represent words and sentences as vectors, strategically positioning semantically similar texts in close proximity within the embedding space (Gao et al., 2021; Le and Mikolov, 2014; Reimers and Gurevych, 2019). Recent research has focused on developing general-purpose embedding models capable of excelling in diverse downstream tasks, including information retrieval, clustering, and classification (Cer et al., 2018; Muennighoff et al., 2023). Leveraging their vast pre-training knowledge, large language models (LLMs) have emerged as a promising avenue for constructing such general-purpose embedding models, with the potential to significantly enhance performance across a broad spectrum of applications (Anil et al., 2023a,b; Brown et al., 2020). The integration of LLMs has revolutionized the development of high-quality embedding models through two primary approaches. Firstly, LLMs have been employed to refine training datasets by generating higher quality examples. Techniques such as hard negative mining (Lee et al., 2024) and synthetic data generation (Dai et al., 2022; Wang et al., 2023) enable the distillation of LLM knowledge into smaller, more efficient embedding models, leading to substantial performance gains. Secondly, recognizing that the embedding model parameters are frequently initialized from language models (Devlin et al., 2019; Karpukhin et al., 2020), researchers have explored leveraging LLM parameters directly for initialization (Ni et al., 2021).

Visual imagery does not consist of solitary objects, but instead reflects the composition of a multitude of fluid concepts. While there have been great advances in visual representation learning, such advances have focused on building better representations for a small number of discrete objects bereft of an understanding of how these objects are interacting. One can observe this limitation in representations learned through captions or contrastive learning -- where the learned model treats an image essentially as a bag of words. Several works have attempted to address this limitation through the development of bespoke learned architectures to directly address the shortcomings in compositional learning. In this work, we focus on simple, and scalable approaches. In particular, we demonstrate that by substantially improving weakly labeled data, i.e. captions, we can vastly improve the performance of standard contrastive learning approaches. Previous CLIP models achieved near chance rate on challenging tasks probing compositional learning. However, our simple approach boosts performance of CLIP substantially and surpasses all bespoke architectures. Furthermore, we showcase our results on a relatively new captioning benchmark derived from DOCCI. We demonstrate through a series of ablations that a standard CLIP model trained with enhanced data may demonstrate impressive performance on image retrieval tasks.

Learning from noisy data has become essential for adapting deep learning models to real-world applications. Traditional methods often involve first evaluating the noise and then applying strategies such as discarding noisy samples, re-weighting, or re-labeling. However, these methods can fall into a vicious cycle when the initial noise evaluation is inaccurate, leading to suboptimal performance. To address this, we propose a novel approach that leverages dataset distillation for noise removal. This method avoids the feedback loop common in existing techniques and enhances training efficiency, while also providing strong privacy protection through offline processing. We rigorously evaluate three representative dataset distillation methods (DATM, DANCE, and RCIG) under various noise conditions, including symmetric noise, asymmetric noise, and real-world natural noise. Our empirical findings reveal that dataset distillation effectively serves as a denoising tool in random noise scenarios but may struggle with structured asymmetric noise patterns, which can be absorbed into the distilled samples. Additionally, clean but challenging samples, such as those from tail classes in imbalanced datasets, may undergo lossy compression during distillation. Despite these challenges, our results highlight that dataset distillation holds significant promise for robust model training, especially in high-privacy environments where noise is prevalent. The source code is available at https://github.com/Kciiiman/DD_LNL.

Transformer models are deployed in a wide range of settings, from multi-accelerator clusters to standalone mobile phones. The diverse inference constraints in these scenarios necessitate practitioners to train foundation models such as PaLM 2, Llama, & ViTs as a series of models of varying sizes. Due to significant training costs, only a select few model sizes are trained and supported, limiting more fine-grained control over relevant tradeoffs, including latency, cost, and accuracy. This work introduces MatFormer, a nested Transformer architecture designed to offer elasticity in a variety of deployment constraints. Each Feed Forward Network (FFN) block of a MatFormer model is jointly optimized with a few nested smaller FFN blocks. This training procedure allows for the Mix'n'Match of model granularities across layers -- i.e., a trained universal MatFormer model enables extraction of hundreds of accurate smaller models, which were never explicitly optimized. We empirically demonstrate MatFormer's effectiveness across different model classes (decoders & encoders), modalities (language & vision), and scales (up to 2.6B parameters). We find that a 2.6B decoder-only MatFormer language model (MatLM) allows us to extract smaller models spanning from 1.5B to 2.6B, each exhibiting comparable validation loss and one-shot downstream evaluations to their independently trained counterparts. Furthermore, we observe that smaller encoders extracted from a universal MatFormer-based ViT (MatViT) encoder preserve the metric-space structure for adaptive large-scale retrieval. Finally, we showcase that speculative decoding with the accurate and consistent submodels extracted from MatFormer can further reduce inference latency.

Recent advancements in deep neural networks for graph-structured data have led to state-of-the-art performance on recommender system benchmarks. However, making these methods practical and scalable to web-scale recommendation tasks with billions of items and hundreds of millions of users remains a challenge. Here we describe a large-scale deep recommendation engine that we developed and deployed at Pinterest. We develop a data-efficient Graph Convolutional Network (GCN) algorithm PinSage, which combines efficient random walks and graph convolutions to generate embeddings of nodes (i.e., items) that incorporate both graph structure as well as node feature information. Compared to prior GCN approaches, we develop a novel method based on highly efficient random walks to structure the convolutions and design a novel training strategy that relies on harder-and-harder training examples to improve robustness and convergence of the model. We also develop an efficient MapReduce model inference algorithm to generate embeddings using a trained model. We deploy PinSage at Pinterest and train it on 7.5 billion examples on a graph with 3 billion nodes representing pins and boards, and 18 billion edges. According to offline metrics, user studies and A/B tests, PinSage generates higher-quality recommendations than comparable deep learning and graph-based alternatives. To our knowledge, this is the largest application of deep graph embeddings to date and paves the way for a new generation of web-scale recommender systems based on graph convolutional architectures.