Kshitij Gupta

Ethan Caballero, Kshitij Gupta, Irina Rish et al.

We present a smoothly broken power law functional form (that we refer to as a Broken Neural Scaling Law (BNSL)) that accurately models & extrapolates the scaling behaviors of deep neural networks (i.e. how the evaluation metric of interest varies as amount of compute used for training (or inference), number of model parameters, training dataset size, model input size, number of training steps, or upstream performance varies) for various architectures & for each of various tasks within a large & diverse set of upstream & downstream tasks, in zero-shot, prompted, & finetuned settings. This set includes large-scale vision, language, audio, video, diffusion, generative modeling, multimodal learning, contrastive learning, AI alignment, AI capabilities, robotics, out-of-distribution (OOD) generalization, continual learning, transfer learning, uncertainty estimation / calibration, OOD detection, adversarial robustness, distillation, sparsity, retrieval, quantization, pruning, fairness, molecules, computer programming/coding, math word problems, "emergent phase transitions", arithmetic, supervised learning, unsupervised/self-supervised learning, & reinforcement learning (single agent & multi-agent). When compared to other functional forms for neural scaling, this functional form yields extrapolations of scaling behavior that are considerably more accurate on this set. Moreover, this functional form accurately models & extrapolates scaling behavior that other functional forms are incapable of expressing such as the nonmonotonic transitions present in the scaling behavior of phenomena such as double descent & the delayed, sharp inflection points present in the scaling behavior of tasks such as arithmetic. Lastly, we use this functional form to glean insights about the limit of the predictability of scaling behavior. Code is available at https://github.com/ethancaballero/broken_neural_scaling_laws

Aniket Didolkar, Kshitij Gupta, Anirudh Goyal et al. · mila

Recurrent neural networks have a strong inductive bias towards learning temporally compressed representations, as the entire history of a sequence is represented by a single vector. By contrast, Transformers have little inductive bias towards learning temporally compressed representations, as they allow for attention over all previously computed elements in a sequence. Having a more compressed representation of a sequence may be beneficial for generalization, as a high-level representation may be more easily re-used and re-purposed and will contain fewer irrelevant details. At the same time, excessive compression of representations comes at the cost of expressiveness. We propose a solution which divides computation into two streams. A slow stream that is recurrent in nature aims to learn a specialized and compressed representation, by forcing chunks of $K$ time steps into a single representation which is divided into multiple vectors. At the same time, a fast stream is parameterized as a Transformer to process chunks consisting of $K$ time-steps conditioned on the information in the slow-stream. In the proposed approach we hope to gain the expressiveness of the Transformer, while encouraging better compression and structuring of representations in the slow stream. We show the benefits of the proposed method in terms of improved sample efficiency and generalization performance as compared to various competitive baselines for visual perception and sequential decision making tasks.

Kshitij Gupta, Benjamin Thérien, Adam Ibrahim et al.

Large language models (LLMs) are routinely pre-trained on billions of tokens, only to restart the process over again once new data becomes available. A much cheaper and more efficient solution would be to enable the continual pre-training of these models, i.e. updating pre-trained models with new data instead of re-training them from scratch. However, the distribution shift induced by novel data typically results in degraded performance on past data. Taking a step towards efficient continual pre-training, in this work, we examine the effect of different warm-up strategies. Our hypothesis is that the learning rate must be re-increased to improve compute efficiency when training on a new dataset. We study the warmup phase of models pre-trained on the Pile (upstream data, 300B tokens) as we continue to pre-train on SlimPajama (downstream data, 297B tokens), following a linear warmup and cosine decay schedule. We conduct all experiments on the Pythia 410M language model architecture and evaluate performance through validation perplexity. We experiment with different pre-training checkpoints, various maximum learning rates, and various warmup lengths. Our results show that while rewarming models first increases the loss on upstream and downstream data, in the longer run it improves the downstream performance, outperforming models trained from scratch$\unicode{x2013}$even for a large downstream dataset.

Tomohiro Sawada, Daniel Paleka, Alexander Havrilla et al. · gatech

Large Language Models (LLMs) have demonstrated remarkable performance on various quantitative reasoning and knowledge benchmarks. However, many of these benchmarks are losing utility as LLMs get increasingly high scores, despite not yet reaching expert performance in these domains. We introduce ARB, a novel benchmark composed of advanced reasoning problems in multiple fields. ARB presents a more challenging test than prior benchmarks, featuring problems in mathematics, physics, biology, chemistry, and law. As a subset of ARB, we introduce a challenging set of math and physics problems which require advanced symbolic reasoning and domain knowledge. We evaluate recent models such as GPT-4 and Claude on ARB and demonstrate that current models score well below 50% on more demanding tasks. In order to improve both automatic and assisted evaluation capabilities, we introduce a rubric-based evaluation approach, allowing GPT-4 to score its own intermediate reasoning steps. Further, we conduct a human evaluation of the symbolic subset of ARB, finding promising agreement between annotators and GPT-4 rubric evaluation scores.

Kshitij Gupta, Devansh Gautam, Radhika Mamidi

Vision-and-language tasks are gaining popularity in the research community, but the focus is still mainly on English. We propose a pipeline that utilizes English-only vision-language models to train a monolingual model for a target language. We propose to extend OSCAR+, a model which leverages object tags as anchor points for learning image-text alignments, to train on visual question answering datasets in different languages. We propose a novel approach to knowledge distillation to train the model in other languages using parallel sentences. Compared to other models that use the target language in the pretraining corpora, we can leverage an existing English model to transfer the knowledge to the target language using significantly lesser resources. We also release a large-scale visual question answering dataset in Japanese and Hindi language. Though we restrict our work to visual question answering, our model can be extended to any sequence-level classification task, and it can be extended to other languages as well. This paper focuses on two languages for the visual question answering task - Japanese and Hindi. Our pipeline outperforms the current state-of-the-art models by a relative increase of 4.4% and 13.4% respectively in accuracy.

Vinodh Kumar Sunkara, Satheeshkumar Karuppusamy, Hangjun Xu et al.

Traditional ads recommendation systems have primarily focused on optimizing for prediction accuracy of click or conversion events using canonical metrics such as recall or normalized discounted cumulative gain (NDCG). With the hyper-growth of ads inventory and liquidity with generative AI technologies, the prediction stability and predictability is becoming increasingly critical. Intuitively, prediction stability and predictability can be defined to quantify system robustness with respect to minor/noisy input (ads, creatives) perturbations, the lack of which could lead to advertiser perceivable problems such as repeatability, cold start and under-exploration. In this paper, we introduce a new evaluation framework for quantifying stability and predictability of an ads recommender system, and present an online validated semantic candidate generation framework powered by fine-tuned Large Language Models (LLMs) that showed significant improvement along these metrics by fundamentally improving the semantic-awareness of the system. The approach extracts hierarchical semantic attributes from ad creatives to obtain LLM representations, which serve as the foundation for graph-based expansion, ensuring the retrieved candidates encapsulate semantic variants of an ad, guaranteeing that small creative variants from the advertiser yield consistent and explainable delivery results to the user. We tested this LLM ads retrieval framework in a large-scale industrial ads recommendation system, demonstrating significant improvements across offline and online A/B experiments, showcasing gains in both predictability and traditional performance metrics. Although evaluated in the ads stack, this is a general framework that can be applied broadly to any large-scale recommendation and retrieval systems facing similar scaling and predictability challenges.

Taishi Nakamura, Mayank Mishra, Simone Tedeschi et al. · ibm-research, stanford

Pretrained language models are an integral part of AI applications, but their high computational cost for training limits accessibility. Initiatives such as Bloom and StarCoder aim to democratize access to pretrained models for collaborative community development. Despite these efforts, such models encounter challenges such as limited multilingual capabilities, risks of catastrophic forgetting during continual pretraining, and the high costs of training models from scratch, alongside the need to align with AI safety standards and regulatory frameworks. This paper presents Aurora-M, a 15B parameter multilingual open-source model trained on English, Finnish, Hindi, Japanese, Vietnamese, and code. Continually pretrained from StarCoderPlus on 435B additional tokens, Aurora-M surpasses 2T tokens in total training token count. It is the first open-source multilingual model fine-tuned on human-reviewed safety instructions, thus aligning its development not only with conventional red-teaming considerations, but also with the specific concerns articulated in the Biden-Harris Executive Order on the Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence. We evaluate Aurora-M across a wide range of tasks and languages, showcasing its robustness against catastrophic forgetting and its superior performance in multilingual settings, particularly in safety evaluations. We open-source Aurora-M and its variants to encourage responsible open-source development of large language models at https://huggingface.co/aurora-m.

Kshitij Gupta

Automated essay scoring is one of the most important problem in Natural Language Processing. It has been explored for a number of years, and it remains partially solved. In addition to its economic and educational usefulness, it presents research problems. Transfer learning has proved to be beneficial in NLP. Data augmentation techniques have also helped build state-of-the-art models for automated essay scoring. Many works in the past have attempted to solve this problem by using RNNs, LSTMs, etc. This work examines the transformer models like BERT, RoBERTa, etc. We empirically demonstrate the effectiveness of transformer models and data augmentation for automated essay grading across many topics using a single model.

Adam Ibrahim, Benjamin Thérien, Kshitij Gupta et al.

Large language models (LLMs) are routinely pre-trained on billions of tokens, only to start the process over again once new data becomes available. A much more efficient solution is to continually pre-train these models, saving significant compute compared to re-training. However, the distribution shift induced by new data typically results in degraded performance on previous data or poor adaptation to the new data. In this work, we show that a simple and scalable combination of learning rate (LR) re-warming, LR re-decaying, and replay of previous data is sufficient to match the performance of fully re-training from scratch on all available data, as measured by the final loss and the average score on several language model (LM) evaluation benchmarks. Specifically, we show this for a weak but realistic distribution shift between two commonly used LLM pre-training datasets (English$\rightarrow$English) and a stronger distribution shift (English$\rightarrow$German) at the $405$M parameter model scale with large dataset sizes (hundreds of billions of tokens). Selecting the weak but realistic shift for larger-scale experiments, we also find that our continual learning strategies match the re-training baseline for a 10B parameter LLM. Our results demonstrate that LLMs can be successfully updated via simple and scalable continual learning strategies, matching the re-training baseline using only a fraction of the compute. Finally, inspired by previous work, we propose alternatives to the cosine learning rate schedule that help circumvent forgetting induced by LR re-warming and that are not bound to a fixed token budget.

Kshitij Gupta

Large pre-trained language models have brought remarkable progress in NLP. Pre-training and Fine-tuning have given state-of-art performance across tasks in text processing. Data Augmentation techniques have also helped build state-of-art models on low or zero resource tasks. Many works in the past have attempted at learning a single massively-multilingual machine translation model for zero-shot translation. Although those translation models are producing correct translations, the main challenge is those models are producing the wrong languages for zero-shot translation. This work and its results indicate that prompt conditioned large models do not suffer from off-target language errors i.e. errors arising due to translation to wrong languages. We empirically demonstrate the effectiveness of self-supervised pre-training and data augmentation for zero-shot multi-lingual machine translation.

SIMA Team, Maria Abi Raad, Arun Ahuja et al. · deepmind, stanford

Building embodied AI systems that can follow arbitrary language instructions in any 3D environment is a key challenge for creating general AI. Accomplishing this goal requires learning to ground language in perception and embodied actions, in order to accomplish complex tasks. The Scalable, Instructable, Multiworld Agent (SIMA) project tackles this by training agents to follow free-form instructions across a diverse range of virtual 3D environments, including curated research environments as well as open-ended, commercial video games. Our goal is to develop an instructable agent that can accomplish anything a human can do in any simulated 3D environment. Our approach focuses on language-driven generality while imposing minimal assumptions. Our agents interact with environments in real-time using a generic, human-like interface: the inputs are image observations and language instructions and the outputs are keyboard-and-mouse actions. This general approach is challenging, but it allows agents to ground language across many visually complex and semantically rich environments while also allowing us to readily run agents in new environments. In this paper we describe our motivation and goal, the initial progress we have made, and promising preliminary results on several diverse research environments and a variety of commercial video games.

Kshitij Gupta

We present a novel data set, WhoDunIt, to assess the deductive reasoning capabilities of large language models (LLM) within narrative contexts. Constructed from open domain mystery novels and short stories, the dataset challenges LLMs to identify the perpetrator after reading and comprehending the story. To evaluate model robustness, we apply a range of character-level name augmentations, including original names, name swaps, and substitutions with well-known real and/or fictional entities from popular discourse. We further use various prompting styles to investigate the influence of prompting on deductive reasoning accuracy. We conduct evaluation study with state-of-the-art models, specifically GPT-4o, GPT-4-turbo, and GPT-4o-mini, evaluated through multiple trials with majority response selection to ensure reliability. The results demonstrate that while LLMs perform reliably on unaltered texts, accuracy diminishes with certain name substitutions, particularly those with wide recognition. This dataset is publicly available here.

Alexis Roger, Prateek Humane, Daniel Z. Kaplan et al.

The proliferation of Vision-Language Models (VLMs) in the past several years calls for rigorous and comprehensive evaluation methods and benchmarks. This work analyzes existing VLM evaluation techniques, including automated metrics, AI-based assessments, and human evaluations across diverse tasks. We first introduce Robin - a novel suite of VLMs that we built by combining Large Language Models (LLMs) and Vision Encoders (VEs) at multiple scales, and use Robin to identify shortcomings of current evaluation approaches across scales. Next, to overcome the identified limitations, we introduce CHIRP - a new long form response benchmark we developed for more robust and complete VLM evaluation. We provide open access to the Robin training code, model suite, and CHIRP benchmark to promote reproducibility and advance VLM research.

Kshitij Gupta, Devansh Gautam, Radhika Mamidi

Multimodal Machine Translation (MMT) enriches the source text with visual information for translation. It has gained popularity in recent years, and several pipelines have been proposed in the same direction. Yet, the task lacks quality datasets to illustrate the contribution of visual modality in the translation systems. In this paper, we propose our system under the team name Volta for the Multimodal Translation Task of WAT 2021 from English to Hindi. We also participate in the textual-only subtask of the same language pair for which we use mBART, a pretrained multilingual sequence-to-sequence model. For multimodal translation, we propose to enhance the textual input by bringing the visual information to a textual domain by extracting object tags from the image. We also explore the robustness of our system by systematically degrading the source text. Finally, we achieve a BLEU score of 44.6 and 51.6 on the test set and challenge set of the multimodal task.

Devansh Gautam, Kshitij Gupta, Manish Shrivastava

Tables are widely used in various kinds of documents to present information concisely. Understanding tables is a challenging problem that requires an understanding of language and table structure, along with numerical and logical reasoning. In this paper, we present our systems to solve Task 9 of SemEval-2021: Statement Verification and Evidence Finding with Tables (SEM-TAB-FACTS). The task consists of two subtasks: (A) Given a table and a statement, predicting whether the table supports the statement and (B) Predicting which cells in the table provide evidence for/against the statement. We fine-tune TAPAS (a model which extends BERT's architecture to capture tabular structure) for both the subtasks as it has shown state-of-the-art performance in various table understanding tasks. In subtask A, we evaluate how transfer learning and standardizing tables to have a single header row improves TAPAS' performance. In subtask B, we evaluate how different fine-tuning strategies can improve TAPAS' performance. Our systems achieve an F1 score of 67.34 in subtask A three-way classification, 72.89 in subtask A two-way classification, and 62.95 in subtask B.

Kshitij Gupta, Devansh Gautam, Radhika Mamidi

Memes are one of the most popular types of content used to spread information online. They can influence a large number of people through rhetorical and psychological techniques. The task, Detection of Persuasion Techniques in Texts and Images, is to detect these persuasive techniques in memes. It consists of three subtasks: (A) Multi-label classification using textual content, (B) Multi-label classification and span identification using textual content, and (C) Multi-label classification using visual and textual content. In this paper, we propose a transfer learning approach to fine-tune BERT-based models in different modalities. We also explore the effectiveness of ensembles of models trained in different modalities. We achieve an F1-score of 57.0, 48.2, and 52.1 in the corresponding subtasks.