Structure Inducing Pre-Training

Language model pre-training and derived methods are incredibly impactful in machine learning. However, there remains considerable uncertainty on exactly why pre-training helps improve performance for fine-tuning tasks. This is especially true when attempting to adapt language-model pre-training to domains outside of natural language. We analyze this problem by exploring how existing pre-training methods impose relational structure in their induced per-sample latent spaces—i.e., what constraints do pre-training methods impose on the distance or geometry between the pre-trained embeddings of two samples xi and xj.

Through a comprehensive review of existing pre-training methods, we find that this question remains open. This is true despite theoretical analyses demonstrating the importance of understanding this form of induced structure. Based on this review, we introduce a descriptive framework for pre-training that allows for a granular, comprehensive understanding of how relational structure can be induced.

We present a theoretical analysis of this framework from the first principles and establish a connection between the relational inductive bias of pre-training and fine-tuning performance. We also show how to use the framework to define new pre-training methods. Finally, we build upon these findings with empirical studies on benchmarks spanning 3 data modalities and ten fine-tuning tasks. These experiments validate our theoretical analyses, inform the design of novel pre-training methods, and establish consistent improvements over a compelling suite of baseline methods.

Existing Pre-training (PT) Methods

Our review summarized 71 existing natural language processing (NLP) and NLP-derived PT methods, which are categorized into clusters based on how they impose structural constraints over the PT per sample) latent space. Clusters are arranged on axes via manual judgments on whether the imposed constraint is shallow vs. deep and implicit vs. explicit. Clusters are sized such that the area corresponds to the number of citations methods included in that cluster have received on average per month since first publication, according to Google Scholar’s citation count. In the following figure, “None” captures models that leverage no pre-training loss over the per-sample embedding. “NSP” refers to “Next sentence Prediction,” the per-sample PT task introduced in BERT. “SOP” refers to “Sentence-order Prediction,” the per-sample PT task introduced in ALBERT. Note that over 90 studies in total were considered in our review, but only 71 met the inclusion criteria to be included in this figure. These methods are described in more detail in the manuscript.

Per-sample vs. Per-token Latent Space

Language model pre-training methods produce both per-sample and per-token latent spaces. Illustrated below via the RoBERTa model are traditional language modeling objectives, which use only a masked language model loss during pre-training) only constrain the per-token latent space.

Structure Inducing Pre-Training

We re-cast the PT formulation by taking a pre-training graph GPT as an auxiliary input. GPT is used to define a new structure-inducing objective LSI, which pushes a pre-training encoder fθ to embed samples such that samples are close in the latent space if and only if they are linked in GPT.

Publication

Structure Inducing Pre-Training
Matthew B.A. McDermott, Brendan Yap, Peter Szolovits and Marinka Zitnik
Nature Machine Intelligence 2023 [arXiv]

@article{mcdermott2023structure,
  title={Structure Inducing Pre-Training},
  author={McDermott, Matthew and Yap, Brendan and Szolovits, Peter and Zitnik, Marinka},
  journal={Nature Machine Intelligence},
  year={2023},
  publisher={Nature Springer Group}
}

Code

PyTorch implementation together with documentation and examples of usage is available in the GitHub repository.

Supplementary Information

Review of language model pre-training methods, Supplementary Figures 1–3 and Table 1 are available at Nature Machine Intelligence.

Authors

Latest News

Mar 2024:   Efficient ML Seminar Series

We started a Harvard University Efficient ML Seminar Series. Congrats to Jonathan for spearheading this initiative. Harvard Magazine covered the first meeting focusing on LLMs.

Mar 2024:   UniTS - Unified Time Series Model

UniTS is a unified time series model that can process classification, forecasting, anomaly detection and imputation tasks within a single model with no task-specific modules. UniTS has zero-shot, few-shot, and prompt learning capabilities. Project website.

Mar 2024:   Weintraub Graduate Student Award

Michelle receives the 2024 Harold M. Weintraub Graduate Student Award. The award recognizes exceptional achievement in graduate studies in biological sciences. News Story. Congratulations!

Mar 2024:   PocketGen - Generating Full-Atom Ligand-Binding Protein Pockets

PocketGen is a deep generative model that generates residue sequence and full-atom structure of protein pockets, maximizing binding to ligands. Project website.

Feb 2024:   SPECTRA - Generalizability of Molecular AI

Feb 2024:   Kaneb Fellowship Award

The lab receives the John and Virginia Kaneb Fellowship Award at Harvard Medical School to enhance research progress in the lab.

Feb 2024:   NSF CAREER Award

The lab receives the NSF CAREER Award for our research in geometric deep learning to facilitate algorithmic and scientific advances in therapeutics.

Feb 2024:   Dean’s Innovation Award in AI

Jan 2024:   AI's Prospects in Nature Machine Intelligence

We discussed AI’s 2024 prospects with Nature Machine Intelligence, covering LLM progress, multimodal AI, multi-task agents, and how to bridge the digital divide across communities and world regions.

Jan 2024:   Combinatorial Therapeutic Perturbations

New paper introducing PDGrapher for combinatorial prediction of chemical and genetic perturbations using causally-inspired neural networks.

Nov 2023:   Next Generation of Therapeutics Commons

Oct 2023:   Structure-Based Drug Design

Geometric deep learning has emerged as a valuable tool for structure-based drug design, to generate and refine biomolecules by leveraging detailed three-dimensional geometric and molecular interaction information.

Oct 2023:   Graph AI in Medicine

Graph AI models in medicine integrate diverse data modalities through pre-training, facilitate interactive feedback loops, and foster human-AI collaboration, paving the way to clinically meaningful predictions.

Sep 2023:   New papers accepted at NeurIPS

Sep 2023:   Future Directions in Network Biology

Excited to share our perspectives on current and future directions in network biology.

Aug 2023:   Scientific Discovery in the Age of AI

Jul 2023:   PINNACLE - Contextual AI protein model

PINNACLE is a contextual AI model for protein understanding that dynamically adjusts its outputs based on biological contexts in which it operates. Project website.

Jun 2023:   Our Group is Joining the Kempner Institute

Excited to join Kempner’s inaugural cohort of associate faculty to advance Kempner’s mission of studying the intersection of natural and artificial intelligence.

Jun 2023:   Welcoming a New Postdoctoral Fellow

An enthusiastic welcome to Shanghua Gao who is joining our group as a postdoctoral research fellow.

Jun 2023:   On Pretraining in Nature Machine Intelligence

Zitnik Lab  ·  Artificial Intelligence in Medicine and Science  ·  Harvard  ·  Department of Biomedical Informatics