Warmstarting for Scaling Language Models

Scaling model sizes to scale performance has worked remarkably well for the current large language models paradigm. The research and empirical findings of various scaling studies led to novel scaling results and laws that guides subsequent research. High training costs for contemporary scales of data and models result in a lack of thorough understanding of how to tune and arrive at such training setups. One direction to ameliorate the cost of pretraining large models is to warmstart the large-scale training from smaller models that are cheaper to tune. In this work, we attempt to understand if the behavior of optimal hyperparameters can be retained under warmstarting for scaling. We explore simple operations that allow the application of theoretically motivated methods of zero-shot transfer of optimal hyperparameters using {\mu}Transfer. We investigate the aspects that contribute to the speedup in convergence and the preservation of stable training dynamics under warmstarting with {\mu}Transfer. We find that shrinking smaller model weights, zero-padding, and perturbing the resulting larger model with scaled initialization from {\mu}P enables effective warmstarting of $\mut{}$.

Transfer Learning for Finetuning Large Language Models

As the landscape of large language models expands, efficiently finetuning for specific tasks becomes increasingly crucial. At the same time, the landscape of parameter-efficient finetuning methods rapidly expands. Consequently, practitioners face a multitude of complex choices when searching for an optimal finetuning pipeline for large language models. To reduce the complexity for practitioners, we investigate transfer learning for finetuning large language models and aim to transfer knowledge about configurations from related finetuning tasks to a new task. In this work, we transfer learn finetuning by meta-learning performance and cost surrogate models for grey-box meta-optimization from a new meta-dataset. Counter-intuitively, we propose to rely only on transfer learning for new datasets. Thus, we do not use task-specific Bayesian optimization but prioritize knowledge transferred from related tasks over task-specific feedback. We evaluate our method on eight synthetic question-answer datasets and a meta-dataset consisting of 1,800 runs of finetuning Microsoft's Phi-3. Our transfer learning is superior to zero-shot, default finetuning, and meta-optimization baselines. Our results demonstrate the transferability of finetuning to adapt large language models more effectively.

Ensembling Finetuned Language Models for Text Classification

Finetuning is a common practice widespread across different communities to adapt pretrained models to particular tasks. Text classification is one of these tasks for which many pretrained models are available. On the other hand, ensembles of neural networks are typically used to boost performance and provide reliable uncertainty estimates. However, ensembling pretrained models for text classification is not a well-studied avenue. In this paper, we present a metadataset with predictions from five large finetuned models on six datasets, and report results of different ensembling strategies from these predictions. Our results shed light on how ensembling can improve the performance of finetuned text classifiers and incentivize future adoption of ensembles in such tasks.

Large Language Models Engineer Too Many Simple Features For Tabular Data

Tabular machine learning problems often require time-consuming and labor-intensive feature engineering. Recent efforts have focused on using large language models (LLMs) to capitalize on their potential domain knowledge. At the same time, researchers have observed ethically concerning negative biases in other LLM-related use cases, such as text generation. These developments motivated us to investigate whether LLMs exhibit a bias that negatively impacts the performance of feature engineering. While not ethically concerning, such a bias could hinder practitioners from fully utilizing LLMs for automated data science. Therefore, we propose a method to detect potential biases by detecting anomalies in the frequency of operators (e.g., adding two features) suggested by LLMs when engineering new features. Our experiments evaluate the bias of four LLMs, two big frontier and two small open-source models, across 27 tabular datasets. Our results indicate that LLMs are biased toward simple operators, such as addition, and can fail to utilize more complex operators, such as grouping followed by aggregations. Furthermore, the bias can negatively impact the predictive performance when using LLM-generated features. Our results call for mitigating bias when using LLMs for feature engineering.

A Human-in-the-Loop Fairness-Aware Model Selection Framework for Complex Fairness Objective Landscapes

Fairness-aware Machine Learning (FairML) applications are often characterized by complex social objectives and legal requirements, frequently involving multiple, potentially conflicting notions of fairness. Despite the well-known Impossibility Theorem of Fairness and extensive theoretical research on the statistical and socio-technical trade-offs between fairness metrics, many FairML tools still optimize or constrain for a single fairness objective. However, this one-sided optimization can inadvertently lead to violations of other relevant notions of fairness. In this socio-technical and empirical study, we frame fairness as a many-objective (MaO) problem by treating fairness metrics as conflicting objectives. We introduce ManyFairHPO, a human-in-the-loop, fairness-aware model selection framework that enables practitioners to effectively navigate complex and nuanced fairness objective landscapes. ManyFairHPO aids in the identification, evaluation, and balancing of fairness metric conflicts and their related social consequences, leading to more informed and socially responsible model-selection decisions. Through a comprehensive empirical evaluation and a case study on the Law School Admissions problem, we demonstrate the effectiveness of ManyFairHPO in balancing multiple fairness objectives, mitigating risks such as self-fulfilling prophecies, and providing interpretable insights to guide stakeholders in making fairness-aware modeling decisions.

Mamba4Cast: Efficient Zero-Shot Time Series Forecasting with State Space Models

This paper introduces Mamba4Cast, a zero-shot foundation model for time series forecasting. Based on the Mamba architecture and inspired by Prior-data Fitted Networks (PFNs), Mamba4Cast generalizes robustly across diverse time series tasks without the need for dataset specific fine-tuning. Mamba4Cast's key innovation lies in its ability to achieve strong zero-shot performance on real-world datasets while having much lower inference times than time series foundation models based on the transformer architecture. Trained solely on synthetic data, the model generates forecasts for entire horizons in a single pass, outpacing traditional auto-regressive approaches. Our experiments show that Mamba4Cast performs competitively against other state-of-the-art foundation models in various data sets while scaling significantly better with the prediction length. The source code can be accessed at https://github.com/automl/Mamba4Cast.

LLM Compression with Neural Architecture Search

Large language models (LLMs) exhibit remarkable reasoning abilities, allowing them to generalize across a wide range of downstream tasks, such as commonsense reasoning or instruction following. However, as LLMs scale, inference costs become increasingly prohibitive, accumulating significantly over their life cycle. This poses the question: Can we compress pre-trained LLMs to meet diverse size and latency requirements? We leverage Neural Architecture Search (NAS) to compress LLMs by pruning structural components, such as attention heads, neurons, and layers, aiming to achieve a Pareto-optimal balance between performance and efficiency. While NAS already achieved promising results on small language models in previous work, in this paper we propose various extensions that allow us to scale to LLMs. Compared to structural pruning baselines, we show that NAS improves performance up to 3.4% on MMLU with an on-device latency speedup.

Dynamic Post-Hoc Neural Ensemblers

Ensemble methods are known for enhancing the accuracy and robustness of machine learning models by combining multiple base learners. However, standard approaches like greedy or random ensembles often fall short, as they assume a constant weight across samples for the ensemble members. This can limit expressiveness and hinder performance when aggregating the ensemble predictions. In this study, we explore employing neural networks as ensemble methods, emphasizing the significance of dynamic ensembling to leverage diverse model predictions adaptively. Motivated by the risk of learning low-diversity ensembles, we propose regularizing the model by randomly dropping base model predictions during the training. We demonstrate this approach lower bounds the diversity within the ensemble, reducing overfitting and improving generalization capabilities. Our experiments showcase that the dynamic neural ensemblers yield competitive results compared to strong baselines in computer vision, natural language processing, and tabular data.

GAMformer: In-Context Learning for Generalized Additive Models

Generalized Additive Models (GAMs) are widely recognized for their ability to create fully interpretable machine learning models for tabular data. Traditionally, training GAMs involves iterative learning algorithms, such as splines, boosted trees, or neural networks, which refine the additive components through repeated error reduction. In this paper, we introduce GAMformer, the first method to leverage in-context learning to estimate shape functions of a GAM in a single forward pass, representing a significant departure from the conventional iterative approaches to GAM fitting. Building on previous research applying in-context learning to tabular data, we exclusively use complex, synthetic data to train GAMformer, yet find it extrapolates well to real-world data. Our experiments show that GAMformer performs on par with other leading GAMs across various classification benchmarks while generating highly interpretable shape functions.

Bayes' Power for Explaining In-Context Learning Generalizations

Traditionally, neural network training has been primarily viewed as an approximation of maximum likelihood estimation (MLE). This interpretation originated in a time when training for multiple epochs on small datasets was common and performance was data bound; but it falls short in the era of large-scale single-epoch trainings ushered in by large self-supervised setups, like language models. In this new setup, performance is compute-bound, but data is readily available. As models became more powerful, in-context learning (ICL), i.e., learning in a single forward-pass based on the context, emerged as one of the dominant paradigms. In this paper, we argue that a more useful interpretation of neural network behavior in this era is as an approximation of the true posterior, as defined by the data-generating process. We demonstrate this interpretations' power for ICL and its usefulness to predict generalizations to previously unseen tasks. We show how models become robust in-context learners by effectively composing knowledge from their training data. We illustrate this with experiments that reveal surprising generalizations, all explicable through the exact posterior. Finally, we show the inherent constraints of the generalization capabilities of posteriors and the limitations of neural networks in approximating these posteriors.