Towards Foundation Models for Science

Siavash Golkar

on behalf of Shirley Ho and the Polymathic AI team

What are foundation models?

Large models pre-trained
on massive datasets

They can perform a variety of  downstream tasks (zero-shot generalization)

They are good starting points for fine-tuning on data poor domains (carry useful inductive bias)

The Key Ideas

  • Large models pre-trained on task-agnostic objectives on massive & diverse datasets
     
  • They can perform a variety of  downstream tasks (zero-shot generalization)
     
  • They are good starting points for fine-tuning on data poor domains (carry useful inductive bias)

Can we translate these innovations into a paradigm shift in machine learning for scientific applications?

Polymathic

 

Advancing Science through Multi‑Disciplinary AI

Our mission: to usher in a new class of machine learning for scientific data, building models that can leverage shared concepts across disciplines."

Meet the Polymathic AI Team

Colm-Cille
Caulfield
University of Cambridge
Leslie
Greengard

Flatiron Institute
New York University
David
Ha

Sakana AI
Yann
LeCun

Meta AI
New York University
Stephane
Mallat

École Normale Supérieure
Collège de France
Flatiron Institute
David
Spergel

Simons Foundation
 
Olga
Troyanskaya

Flatiron Institute
Princeton University
Laure
Zanna
New York University

Our Resources

SCIENTIFIC ADVISORY GROUP

COMPUTING RESOURCES

  • Internal H100 GPUs resources at the Flatiron Institute 
  • External 500k GPU hours (V100 and A100)
  • In the process of securing additional O(10^2) dedicated H100 GPUs

Thanks Ian!

The Foundation Model Spectrum

Scientific Reasoning 

Multi-Modality

Generalization to Data-Limited Domains

How can we build foundation models that can have these properties?

  • How many assumptions should we make about the structure of our inputs?
  • What kind of embedding schemes should we use?

    Different choices lead to foundation models with different strengths and weaknesses.

more efficient, less general

more general, less efficient

Language-like/less structured

Structured-data

The Foundation Model Spectrum

Language-like/less structured

Structured-data

xVal
 A Continuous Number Encoding for LLMs

AstroCLIP
Cross-Modal Pretraining for Astronomical data

MPP
Multiple Physics Pretraining for Physical Surrogate Models

Scientific Reasoning 

Multi-Modality

Generalization to Data-Limited Domains

more efficient, less general

more general, less efficient

MPP

Multiple Physics Pretraining for
Physical Surrogate Models

Project led by Michael McCabe,  Bruno Régaldo, Liam Parker, Ruben Ohana, Miles Cranmer
Oral presentation at the NeurIPS 2023 AI4Science Workshop

Context

  • Previous works on large domain-specific pretrained models: chemistry, medicine, astrophysics, climate, ...
    → extension on surrogate modeling of spatiotemporal physical systems
  • Spatiotemporal predictions motivated by faster surrogates for PDE solvers, systems that are hard to simulate with current models/hardware
  • Situations where data is expensive…

Time

Ex: N-body simulation

Springel et al. 2005

The Foundation Model Spectrum

Language-like/less structured

Structured-data

xVal
 A Continuous Number Encoding for LLMs

AstroCLIP
Cross-Modal Pretraining for Astronomical data

MPP
Multiple Physics Pretraining for Physical Surrogate Models

Scientific Reasoning 

Multi-Modality

Generalization to Data-Limited Domains

Can we  build a domain specific foundation model that would be finetunable on a few training examples?

Background

  • Main pretraining strategies
     - autoregressive prediction
     - masked reconstruction
     - contrastive learning
  • No conditioning on physical parameters
  • Spatiotemporal physics: PDEs from a physical system with typically conservations and symmetries…
         → Suggests that there are some learnable shared features

Natural choice for physical surrogate modeling

Physical Systems from PDEBench

Navier-Stokes

Incompressible

Compressible

Shallow Water

Diffusion-Reaction

Takamoto et al. 2022

Compositionality and Pretraining

Balancing objectives during training

Normalized MSE:

Architecture for MPP

Experiment 1: Performance on Pretraining Tasks

Context size: 16 frames

Experiment 2: Transfer

Compressible Navier-Stokes

M = 0.1

M = 1.0

Fun Fact: Finetuning with VideoMAE works quite well…

Tube masking

ViT

ViT

ViT

Trained on reconstructing masked pixels on natural videos (SSV2 and K400)

Tong et al. 2022

Experiment 3: Broader Downstream Tasks

Regression Problems on Incompressible Navier-Stokes

Mixed results

Long time predictions on Navier-Stokes?

Conclusion

  • A single pre-trained transformer model matches/outperforms specialized baselines for each in-distribution task.
  • Demonstrated transfer capabilities were (both near and far).​
  • Early days. We are in the process of scaling up to more diverse data.
     
  • Open code and pretrained models:                      https://github.com/PolymathicAI/multiple_physics_pretraining

xVal

 A Continuous Number Encoding for LLMs

Project led by Siavash Golkar, Mariel Pettee, Michael Eickenberg, Alberto Bietti
Accepted contribution at the NeurIPS 2023 AI4Science Workshop

The Foundation Model Spectrum

Language-like/less structured

Structured-data

xVal
 A Continuous Number Encoding for LLMs

AstroCLIP
Cross-Modal Pretraining for Astronomical data

MPP
Multiple Physics Pretraining for Physical Surrogate Models

Scientific Reasoning 

Multi-Modality

Generalization to Data-Limited Domains

The problem: existing LLMs are not suitable for reliable zero-shot numerical operations.

They make erratic, discontinuous predictions.

Even fine-tuning exhaustively does not grant out-of-distribution generalization abilities.

Can we  build an LLM better suited for numerical data analysis?

xVal in a Nutshell: a continuous numerical encoding for language models

xVal in a Nutshell: a continuous numerical encoding for language models

This encoding strategy has 3 main benefits:

  • Continuity

    • The model is now end-to-end continuous by construction.
      (Standard LLMs are discontinuous both at the input and output stage.)

  • Interpolation

    • It makes better out-of-distribution predictions than other numerical encodings.

  • Efficiency

    • By using just a single token to represent any number, it requires less memory, compute resources, and training time to achieve strong results.

 

xVal in a Nutshell: a continuous numerical encoding for language models

xVal shows improved predictions for out-of-distribution values.

xVal in a Nutshell: a continuous numerical encoding for language models

When evaluated on multi-digit multiplication tasks, xVal performs comparably well, and is less prone to large outliers:

And when evaluated on compound operations of basic arithmetic, xVal shows the strongest performance:

xVal in a Nutshell: a continuous numerical encoding for language models

Future directions: improving the dynamic range of the embedded values.

Conclusion

  • xVal improves numerical analysis performance of transformer models.​
  • Good fit for numerical-data-rich texts.
  • Not great for general purpose LLMs
    (limited dynamic range, can't easily handle 13, 666, 1337, ...).
  • Not suitable for heavy numerical computation.
     
  • Code available:        https://github.com/PolymathicAI/xVal

AstroCLIP

Cross-Modal Pre-Training for Astronomical Foundation Models

Project led by Francois Lanusse, Liam Parker, Siavash Golkar, Miles Cranmer
Accepted contribution at the NeurIPS 2023 AI4Science Workshop

The Data Diversity Challenge

  • Scientific data is extremely multimodal.
    • Observations from different instruments are not interchangeable.
    • Metadata matters

Credit:DESI collaboration/DESI Legacy Imaging Surveys/LBNL/DOE & KPNO/CTIO/NOIRLab/NSF/AURA/unWISE

The Foundation Model Spectrum

Language-like/less structured

Structured-data

xVal
 A Continuous Number Encoding for LLMs

AstroCLIP
Cross-Modal Pretraining for Astronomical data

MPP
Multiple Physics Pretraining for Physical Surrogate Models

Scientific Reasoning 

Multi-Modality

Generalization to Data-Limited Domains

Towards Large Multi-Modal Observational Models

Most General

Most Specific

Independent models for every type of observation

Single model capable of processing all types of observations 

Towards Large Multi-Modal Observational Models

Most General

Most Specific

Independent models for every type of observation

Single model capable of processing all types of observations 

Bytes Are All You Need (Horton et al. 2023)

Towards Large Multi-Modal Observational Models

Most General

Most Specific

Independent models for every type of observation

Single model capable of processing all types of observations 

Bytes Are All You Need (Horton et al. 2023)

AstroCLIP

Can we build a baseline multimodal by bringing together specialized models trained on each modality?

Contrastive Learning in Astrophysics

Self-Supervised similarity search for large scientific datasets (Stein et al. 2021)


How do we add in other modalities?
(e.g. spectral information?)





 

Example of Science Application: Identifying Galaxy Tidal Features

What is CLIP?

Contrastive Language Image Pretraining (CLIP)
(Radford et al. 2021)

One model, many downstream applications!

Flamingo: a Visual Language Model for Few-Shot Learning (Alayrac et al. 2022)

Hierarchical Text-Conditional Image Generation with CLIP Latents (Ramesh et al. 2022)

The AstroCLIP approach

  • We use spectra and multi-band images as our two different views for the same underlying object.
     
  • DESI Legacy Surveys (g,r,z) images, and DESI EDR galaxy spectra.
     
  • Once trained, we can do example retrieval by nearest neighbor search.

How we do it

We take a two steps approach:

  • Build self-supervised model separately for images and spectra
    • Images: Start from pre-trained ResNet 50 from Stein et al. (2021)
    • Spectra: Pretrain by mask modeling a GPT-2 like transformer on spectra
  • Train an embedding module on top of each backbone under InfoNCE loss
    • Images: Simple MLP
    • Spectra: Cross-attention module 

More examples of retrieval

Image Similarity

Spectral Similarity

Image-Spectral Similarity

Visualizing properties of the embedding space

UMAP representation of spectra embeddings

Testing Structure of Embedding Space by k-NN Regression

Comparison to image-only SSL (from Stein et al. 2021)

The Information Point of View

  • The InfoNCE loss is a lower bound on the Mutual Information between modalities

Shared physical information about galaxies between images and spectra

Thinking about data from a hierarchical Bayesian model point of view

y \sim \underbrace{p_{obs}( y | x )}_{\text{instrument specific}} \qquad \underbrace{p(x | \theta)}_{\text{generative process}} \qquad \underbrace{p(\theta)}_{\text{physical parameters}}

=> We are building summary statistics for the physical parameters describing an object in a completely data driven way

What comes next!

  • We want to build a shared embedding space for all astrophysics observation modalities.
    • Next steps would be, embedded data from different instruments, different filters, etc... to build a universal embedding for types of galaxy observations
       
    • Deeper interaction between the modalities beyond alignment.

Teaser for what comes next

We are just getting started!

Thank you for listening!

Grand Challenges for Foundation Models in Science

  • Scientific data is massively multi-modal (not finite). Training a specialist  model for each modality is not feasible as we grow the scope of our models.
     
  • Transformers have good inductive bias for many tasks (vision, language, etc.)
    But they are not necessarily the best models for scientific foundation models.
    (Provably inefficient in specific cases.)
     
  • Many fundamental challenges remain (especially in physics).
    (Uncertainty estimates, interpretation, concept discovery, etc)
    "large data + large transformer" does not solve everything.

Towards Foundation Models for Science

By Siavash Golkar

Towards Foundation Models for Science

Overview talk of the Polymathic AI Initiative

  • 122