We present a method to formulate algorithm discovery as program search, and apply it to discover optimization algorithms for deep neural network training. We leverage efficient search techniques to explore an infinite and sparse program space. To bridge the large generalization gap between proxy and target tasks, we also introduce program selection and simplification strategies. Our method discovers a simple and effective optimization algorithm, $\textbf{Lion}$ ($\textit{Evo$\textbf{L}$ved S$\textbf{i}$gn M$\textbf{o}$me$\textbf{n}$tum}$). It is more memory-efficient than Adam as it only keeps track of the momentum. Different from adaptive optimizers, its update has the same magnitude for each parameter calculated through the sign operation. We compare Lion with widely used optimizers, such as Adam and Adafactor, for training a variety of models on different tasks. On image classification, Lion boosts the accuracy of ViT by up to 2% on ImageNet and saves up to 5x the pre-training compute on JFT. On vision-language contrastive learning, we achieve 88.3% $\textit{zero-shot}$ and 91.1% $\textit{fine-tuning}$ accuracy on ImageNet, surpassing the previous best results by 2% and 0.1%, respectively. On diffusion models, Lion outperforms Adam by achieving a better FID score and reducing the training compute by up to 2.3x. For autoregressive, masked language modeling, and fine-tuning, Lion exhibits a similar or better performance compared to Adam. Our analysis of Lion reveals that its performance gain grows with the training batch size. It also requires a smaller learning rate than Adam due to the larger norm of the update produced by the sign function. Additionally, we examine the limitations of Lion and identify scenarios where its improvements are small or not statistically significant. Lion is also successfully deployed in production systems such as Google search ads CTR model.
Vision-language pre-training like CLIP has shown promising performance on various downstream tasks such as zero-shot image classification and image-text retrieval. Most of the existing CLIP-alike works usually adopt relatively large image encoders like ResNet50 and ViT, while the lightweight counterparts are rarely discussed. In this paper, we propose a multi-level interaction paradigm for training lightweight CLIP models. Firstly, to mitigate the problem that some image-text pairs are not strictly one-to-one correspondence, we improve the conventional global instance-level alignment objective by softening the label of negative samples progressively. Secondly, a relaxed bipartite matching based token-level alignment objective is introduced for finer-grained alignment between image patches and textual words. Moreover, based on the observation that the accuracy of CLIP model does not increase correspondingly as the parameters of text encoder increase, an extra objective of masked language modeling (MLM) is leveraged for maximizing the potential of the shortened text encoder. In practice, an auxiliary fusion module injecting unmasked image embedding into masked text embedding at different network stages is proposed for enhancing the MLM. Extensive experiments show that without introducing additional computational cost during inference, the proposed method achieves a higher performance on multiple downstream tasks.
Like masked language modeling (MLM) in natural language processing, masked image modeling (MIM) aims to extract valuable insights from image patches to enhance the feature extraction capabilities of the underlying deep neural network (DNN). Contrasted with other training paradigms like supervised learning and unsupervised contrastive learning, masked image modeling (MIM) pretraining typically demands significant computational resources in order to manage large training data batches (e.g., 4096). The significant memory and computation requirements pose a considerable challenge to its broad adoption. To mitigate this, we introduce a novel learning framework, termed~\textit{Block-Wise Masked Image Modeling} (BIM). This framework involves decomposing the MIM tasks into several sub-tasks with independent computation patterns, resulting in block-wise back-propagation operations instead of the traditional end-to-end approach. Our proposed BIM maintains superior performance compared to conventional MIM while greatly reducing peak memory consumption. Moreover, BIM naturally enables the concurrent training of numerous DNN backbones of varying depths. This leads to the creation of multiple trained DNN backbones, each tailored to different hardware platforms with distinct computing capabilities. This approach significantly reduces computational costs in comparison with training each DNN backbone individually. Our framework offers a promising solution for resource constrained training of MIM.
Providing a personalized user experience on information dense webpages helps users in reaching their end-goals sooner. We explore an automated approach to identifying user personas by leveraging high dimensional trajectory information from user sessions on webpages. While neural collaborative filtering (NCF) approaches pay little attention to token semantics, our method introduces SessionBERT, a Transformer-backed language model trained from scratch on the masked language modeling (mlm) objective for user trajectories (pages, metadata, billing in a session) aiming to capture semantics within them. Our results show that representations learned through SessionBERT are able to consistently outperform a BERT-base model providing a 3% and 1% relative improvement in F1-score for predicting page links and next services. We leverage SessionBERT and extend it to provide recommendations (top-5) for the next most-relevant services that a user would be likely to use. We achieve a HIT@5 of 58% from our recommendation model.
We describe our team's contribution to the STRICT-SMALL track of the BabyLM Challenge. The challenge requires training a language model from scratch using only a relatively small training dataset of ten million words. We experiment with three variants of cognitively-motivated curriculum learning and analyze their effect on the performance of the model on linguistic evaluation tasks. In the vocabulary curriculum, we analyze methods for constraining the vocabulary in the early stages of training to simulate cognitively more plausible learning curves. In the data curriculum experiments, we vary the order of the training instances based on i) infant-inspired expectations and ii) the learning behavior of the model. In the objective curriculum, we explore different variations of combining the conventional masked language modeling task with a more coarse-grained word class prediction task to reinforce linguistic generalization capabilities. Our results did not yield consistent improvements over our own non-curriculum learning baseline across a range of linguistic benchmarks; however, we do find marginal gains on select tasks. Our analysis highlights key takeaways for specific combinations of tasks and settings which benefit from our proposed curricula. We moreover determine that careful selection of model architecture, and training hyper-parameters yield substantial improvements over the default baselines provided by the BabyLM challenge.
This paper presents the Ensemble Nucleotide Byte-level Encoder-Decoder (ENBED) foundation model, analyzing DNA sequences at byte-level precision with an encoder-decoder Transformer architecture. ENBED uses a sub-quadratic implementation of attention to develop an efficient model capable of sequence-to-sequence transformations, generalizing previous genomic models with encoder-only or decoder-only architectures. We use Masked Language Modeling to pre-train the foundation model using reference genome sequences and apply it in the following downstream tasks: (1) identification of enhancers, promotors and splice sites, (2) recognition of sequences containing base call mismatches and insertion/deletion errors, an advantage over tokenization schemes involving multiple base pairs, which lose the ability to analyze with byte-level precision, (3) identification of biological function annotations of genomic sequences, and (4) generating mutations of the Influenza virus using the encoder-decoder architecture and validating them against real-world observations. In each of these tasks, we demonstrate significant improvement as compared to the existing state-of-the-art results.
Real-world NLP applications often deal with nonstandard text (e.g., dialectal, informal, or misspelled text). However, language models like BERT deteriorate in the face of dialect variation or noise. How do we push BERT's modeling capabilities to encompass nonstandard text? Fine-tuning helps, but it is designed for specializing a model to a task and does not seem to bring about the deeper, more pervasive changes needed to adapt a model to nonstandard language. In this paper, we introduce the novel idea of sandwiching BERT's encoder stack between additional encoder layers trained to perform masked language modeling on noisy text. We find that our approach, paired with recent work on including character-level noise in fine-tuning data, can promote zero-shot transfer to dialectal text, as well as reduce the distance in the embedding space between words and their noisy counterparts.
Techniques in causal analysis of language models illuminate how linguistic information is organized in LLMs. We use one such technique, AlterRep, a method of counterfactual probing, to explore the internal structure of multilingual models (mBERT and XLM-R). We train a linear classifier on a binary language identity task, to classify tokens between Language X and Language Y. Applying a counterfactual probing procedure, we use the classifier weights to project the embeddings into the null space and push the resulting embeddings either in the direction of Language X or Language Y. Then we evaluate on a masked language modeling task. We find that, given a template in Language X, pushing towards Language Y systematically increases the probability of Language Y words, above and beyond a third-party control language. But it does not specifically push the model towards translation-equivalent words in Language Y. Pushing towards Language X (the same direction as the template) has a minimal effect, but somewhat degrades these models. Overall, we take these results as further evidence of the rich structure of massive multilingual language models, which include both a language-specific and language-general component. And we show that counterfactual probing can be fruitfully applied to multilingual models.
We present Lil-Bevo, our submission to the BabyLM Challenge. We pretrained our masked language models with three ingredients: an initial pretraining with music data, training on shorter sequences before training on longer ones, and masking specific tokens to target some of the BLiMP subtasks. Overall, our baseline models performed above chance, but far below the performance levels of larger LLMs trained on more data. We found that training on short sequences performed better than training on longer sequences.Pretraining on music may help performance marginally, but, if so, the effect seems small. Our targeted Masked Language Modeling augmentation did not seem to improve model performance in general, but did seem to help on some of the specific BLiMP tasks that we were targeting (e.g., Negative Polarity Items). Training performant LLMs on small amounts of data is a difficult but potentially informative task. While some of our techniques showed some promise, more work is needed to explore whether they can improve performance more than the modest gains here. Our code is available at https://github.com/venkatasg/Lil-Bevo and out models at https://huggingface.co/collections/venkatasg/babylm-653591cdb66f4bf68922873a
We analyze the masked language modeling pretraining objective function from the perspective of the distributional hypothesis. We investigate whether better sample efficiency and the better generalization capability of models pretrained with masked language modeling can be attributed to the semantic similarity encoded in the pretraining data's distributional property. Via a synthetic dataset, our analysis suggests that distributional property indeed leads to the better sample efficiency of pretrained masked language models, but does not fully explain the generalization capability. We also conduct analyses over two real-world datasets and demonstrate that the distributional property does not explain the generalization ability of pretrained natural language models either. Our results illustrate our limited understanding of model pretraining and provide future research directions.
Humans subconsciously engage in geospatial reasoning when reading articles. We recognize place names and their spatial relations in text and mentally associate them with their physical locations on Earth. Although pretrained language models can mimic this cognitive process using linguistic context, they do not utilize valuable geospatial information in large, widely available geographical databases, e.g., OpenStreetMap. This paper introduces GeoLM, a geospatially grounded language model that enhances the understanding of geo-entities in natural language. GeoLM leverages geo-entity mentions as anchors to connect linguistic information in text corpora with geospatial information extracted from geographical databases. GeoLM connects the two types of context through contrastive learning and masked language modeling. It also incorporates a spatial coordinate embedding mechanism to encode distance and direction relations to capture geospatial context. In the experiment, we demonstrate that GeoLM exhibits promising capabilities in supporting toponym recognition, toponym linking, relation extraction, and geo-entity typing, which bridge the gap between natural language processing and geospatial sciences. The code is publicly available at https://github.com/knowledge-computing/geolm.
Numerous debiasing techniques have been proposed to mitigate the gender bias that is prevalent in pretrained language models. These are often evaluated on datasets that check the extent to which the model is gender-neutral in its predictions. Importantly, this evaluation protocol overlooks the possible adverse impact of bias mitigation on useful gender knowledge. To fill this gap, we propose DiFair, a manually curated dataset based on masked language modeling objectives. DiFair allows us to introduce a unified metric, gender invariance score, that not only quantifies a model's biased behavior, but also checks if useful gender knowledge is preserved. We use DiFair as a benchmark for a number of widely-used pretained language models and debiasing techniques. Experimental results corroborate previous findings on the existing gender biases, while also demonstrating that although debiasing techniques ameliorate the issue of gender bias, this improvement usually comes at the price of lowering useful gender knowledge of the model.
Sequential tabular data is one of the most commonly used data types in real-world applications. Different from conventional tabular data, where rows in a table are independent, sequential tabular data contains rich contextual and sequential information, where some fields are dynamically changing over time and others are static. Existing transformer-based approaches analyzing sequential tabular data overlook the differences between dynamic and static fields by replicating and filling static fields into each transformer, and ignore temporal information between rows, which leads to three major disadvantages: (1) computational overhead, (2) artificially simplified data for masked language modeling pre-training task that may yield less meaningful representations, and (3) disregarding the temporal behavioral patterns implied by time intervals. In this work, we propose FATA-Trans, a model with two field transformers for modeling sequential tabular data, where each processes static and dynamic field information separately. FATA-Trans is field- and time-aware for sequential tabular data. The field-type embedding in the method enables FATA-Trans to capture differences between static and dynamic fields. The time-aware position embedding exploits both order and time interval information between rows, which helps the model detect underlying temporal behavior in a sequence. Our experiments on three benchmark datasets demonstrate that the learned representations from FATA-Trans consistently outperform state-of-the-art solutions in the downstream tasks. We also present visualization studies to highlight the insights captured by the learned representations, enhancing our understanding of the underlying data. Our codes are available at https://github.com/zdy93/FATA-Trans.
The Document-based Visual Question Answering competition addresses the automatic detection of parent-child relationships between elements in multi-page documents. The goal is to identify the document elements that answer a specific question posed in natural language. This paper describes the PoliTo's approach to addressing this task, in particular, our best solution explores a text-only approach, leveraging an ad hoc sampling strategy. Specifically, our approach leverages the Masked Language Modeling technique to fine-tune a BERT model, focusing on sentences containing sensitive keywords that also occur in the questions, such as references to tables or images. Thanks to the effectiveness of this approach, we are able to achieve high performance compared to baselines, demonstrating how our solution contributes positively to this task.
Language models have become the backbone of today's AI systems. However, their predominant left-to-right generation limits the use of bidirectional context, which is essential for tasks that involve filling text in the middle. We propose the Fill-in Language Model (FiLM), a new language modeling approach that allows for flexible generation at any position without adhering to a specific generation order. Its training extends the masked language modeling objective by adopting varying mask probabilities sampled from the Beta distribution to enhance the generative capabilities of FiLM. During inference, FiLM can seamlessly insert missing phrases, sentences, or paragraphs, ensuring that the outputs are fluent and are coherent with the surrounding context. In both automatic and human evaluations, FiLM outperforms existing infilling methods that rely on left-to-right language models trained on rearranged text segments. FiLM is easy to implement and can be either trained from scratch or fine-tuned from a left-to-right language model. Notably, as the model size grows, FiLM's perplexity approaches that of strong left-to-right language models of similar sizes, indicating FiLM's scalability and potential as a large language model.
Click-through rate (CTR) prediction has become increasingly indispensable for various Internet applications. Traditional CTR models convert the multi-field categorical data into ID features via one-hot encoding, and extract the collaborative signals among features. Such a paradigm suffers from the problem of semantic information loss. Another line of research explores the potential of pretrained language models (PLMs) for CTR prediction by converting input data into textual sentences through hard prompt templates. Although semantic signals are preserved, they generally fail to capture the collaborative information (e.g., feature interactions, pure ID features), not to mention the unacceptable inference overhead brought by the huge model size. In this paper, we aim to model both the semantic knowledge and collaborative knowledge for accurate CTR estimation, and meanwhile address the inference inefficiency issue. To benefit from both worlds and close their gaps, we propose a novel model-agnostic framework (i.e., ClickPrompt), where we incorporate CTR models to generate interaction-aware soft prompts for PLMs. We design a prompt-augmented masked language modeling (PA-MLM) pretraining task, where PLM has to recover the masked tokens based on the language context, as well as the soft prompts generated by CTR model. The collaborative and semantic knowledge from ID and textual features would be explicitly aligned and interacted via the prompt interface. Then, we can either tune the CTR model with PLM for superior performance, or solely tune the CTR model without PLM for inference efficiency. Experiments on four real-world datasets validate the effectiveness of ClickPrompt compared with existing baselines.
Target proteins that lack accessible binding pockets and conformational stability have posed increasing challenges for drug development. Induced proximity strategies, such as PROTACs and molecular glues, have thus gained attention as pharmacological alternatives, but still require small molecule docking at binding pockets for targeted protein degradation. The computational design of protein-based binders presents unique opportunities to access "undruggable" targets, but have often relied on stable 3D structures or structure-influenced latent spaces for effective binder generation. In this work, we introduce PepMLM, a target sequence-conditioned generator of de novo linear peptide binders. By employing a novel span masking strategy that uniquely positions cognate peptide sequences at the C-terminus of target protein sequences, PepMLM fine-tunes the state-of-the-art ESM-2 pLM to fully reconstruct the binder region, achieving low perplexities matching or improving upon validated peptide-protein sequence pairs. After successful in silico benchmarking with AlphaFold-Multimer, outperforming RFDiffusion on structured targets, we experimentally verify PepMLM's efficacy via fusion of model-derived peptides to E3 ubiquitin ligase domains, demonstrating endogenous degradation of emergent viral phosphoproteins and Huntington's disease-driving proteins. In total, PepMLM enables the generative design of candidate binders to any target protein, without the requirement of target structure, empowering downstream therapeutic applications.
This thesis focuses on improving the pre-training of natural language models using unsupervised raw data to make them more efficient and aligned with downstream applications. In the first part, we introduce three alternative pre-training objectives to BERT's Masked Language Modeling (MLM), namely Random Token Substitution (RTS), Cluster-based Random Token Substitution (C-RTS), and Swapped Language Modeling (SLM). These objectives involve token swapping instead of masking, with RTS and C-RTS aiming to predict token originality and SLM predicting the original token values. Results show that RTS and C-RTS require less pre-training time while maintaining performance comparable to MLM. Surprisingly, SLM outperforms MLM on certain tasks despite using the same computational budget. In the second part, we proposes self-supervised pre-training tasks that align structurally with downstream applications, reducing the need for labeled data. We use large corpora like Wikipedia and CC-News to train models to recognize if text spans originate from the same paragraph or document in several ways. By doing continuous pre-training, starting from existing models like RoBERTa, ELECTRA, DeBERTa, BART, and T5, we demonstrate significant performance improvements in tasks like Fact Verification, Answer Sentence Selection, and Summarization. These improvements are especially pronounced when limited annotation data is available. The proposed objectives also achieve state-of-the-art results on various benchmark datasets, including FEVER (dev set), ASNQ, WikiQA, and TREC-QA, as well as enhancing the quality of summaries. Importantly, these techniques can be easily integrated with other methods without altering the internal structure of Transformer models, making them versatile for various NLP applications.
The meanings of words and phrases depend not only on where they are used (contexts) but also on who use them (writers). Pretrained language models (PLMs) are powerful tools for capturing context, but they are typically pretrained and fine-tuned for universal use across different writers. This study aims to improve the accuracy of text understanding tasks by personalizing the fine-tuning of PLMs for specific writers. We focus on a general setting where only the plain text from target writers are available for personalization. To avoid the cost of fine-tuning and storing multiple copies of PLMs for different users, we exhaustively explore using writer-specific prompts to personalize a unified PLM. Since the design and evaluation of these prompts is an underdeveloped area, we introduce and compare different types of prompts that are possible in our setting. To maximize the potential of prompt-based personalized fine-tuning, we propose a personalized intermediate learning based on masked language modeling to extract task-independent traits of writers' text. Our experiments, using multiple tasks, datasets, and PLMs, reveal the nature of different prompts and the effectiveness of our intermediate learning approach.
Text-based person re-identification (TBPReID) aims to retrieve person images represented by a given textual query. In this task, how to effectively align images and texts globally and locally is a crucial challenge. Recent works have obtained high performances by solving Masked Language Modeling (MLM) to align image/text parts. However, they only performed uni-directional (i.e., from image to text) local-matching, leaving room for improvement by introducing opposite-directional (i.e., from text to image) local-matching. In this work, we introduce Bidirectional Local-Matching (BiLMa) framework that jointly optimize MLM and Masked Image Modeling (MIM) in TBPReID model training. With this framework, our model is trained so as the labels of randomly masked both image and text tokens are predicted by unmasked tokens. In addition, to narrow the semantic gap between image and text in MIM, we propose Semantic MIM (SemMIM), in which the labels of masked image tokens are automatically given by a state-of-the-art human parser. Experimental results demonstrate that our BiLMa framework with SemMIM achieves state-of-the-art Rank@1 and mAP scores on three benchmarks.