Model Exchange
& Convergence Initiative
ModECI (Model Exchange and Convergence Initiative) is a multi-investigator collaboration that aims to develop a standardized format for exchanging computational models across diverse software platforms and domains of scientific research and technology development, with a particular focus on neuroscience, Machine Learning and Artificial Intelligence.

Why is this necessary?

As the complexity of the phenomena addressed by most scientific disciplines has increased, theory is relying increasingly on the use of computational modeling. This has produced a proliferation of software tools for modeling that, in turn, has led to a “Tower of Babel” problem, producing barriers to the exchange of theoretical and algorithmic advances across disciplines, that is making it harder to replicate and validate existing results, impeding the creation of more sophisticated models that integrate components developed by researchers across disciplines, and introducing barriers to portability of models across software platforms and development environments. The goal of ModECI is to overcome this balkanization of computational research and technology by developing an MDF that will permit transparent exchange of models.

Currently, there are no existing standards that fully meet the identified need. The closest standard, ONNX, is focused more on providing an intermediate representation (IR) for compilation and optimization rather than exchange. NeuroML is a model description format for biologically inspired neuronal networks, but is not currently used for more abstract networks and graphs used cognitive neuroscience, cognitive science or in machine learning applications.

mdf-logo

ModECI will provide a standardized Model Description Format (MDF) that can be used to exchange models between different software modeling environments and scientific domains and will be presented as an exchange standard, not a programming language. ONNX and NeuroML will provide guidance for the scope of MDF, and will be potential input/output formats. Such an MDF would have numerous benefits, both scientific and technological, including:

  • Ease of dissemination of models and validation of model reproducibility
  • Migration of models across domains (e.g., use of models of brain function in machine learning applications)
  • Integration of models at different levels of analysis (e.g. biophysically-realistic neural models into models of cognitive function, cognitive models as agents in population level models)
  • Exploitation of complementary strengths of existing packages (e.g. design in a familiar environment but execute in one with better tools for parameter tuning and/or data-fitting)
  • More efficient development of new tools, by providing developers with a representative diversity of models, all in a common format
mdf-logo

The purpose of the MDF is to provide a standard

format for describing computational models across a wide range of scientific disciplines

The ModECI MDF is in early stages of development.

See details on GitHub
The goal is to provide a common exchange format that allows models created in one environment that supports the standard to be expressed in a form – and in sufficient detail – that it can be imported into another modeling environment that supports the standard, and then executed in that environment with identical results, and/or integrated with other models in that environment.

Supported elements

The goal of the MDF is to provide a common exchange format that allows models created in one environment that supports the standard to be expressed in a form – and in sufficient detail – that it can be imported into another modeling environment that supports the standard, and then executed in that environment with identical results, and/or integrated with other models in that environment.

More info icon

ONNX (Open Neural Network Exchange) is a well-established and actively supported standard in the machine learning community.
We have leveraged this, by developing a proof of concept interface between MDF and ONNX, showing not only that MDF can complement ONNX as an exchange-oriented format among machine learning environments, but also bridge to ONNX from environments in other domains.

More info icon

NeuroML is a widely accepted standard for describing biophysically-realistic models of neurons, neuronal subcomponents, and neural circuits in machine readable form, that is supported by a large number of existing programming and execution environments.
We have shown that MDF can be used to encode neuroscientific models created in NeuroML, allowing such models to be run on a wider range of simulation platforms.

More info icon

We are planning to extend MDF support for mapping to and from the SONATA format for large scale neuronal modeling.

More info icon

PyTorch is a widely used framework in machine learning and we have developed an initial mapping between MDF models and PyTorch

More info icon

We are planning to extend MDF support to TensorFlow in the next phase of this project.

More info icon

PsyNeuLink is an open-source, software environment written in Python, and designed for the needs of neuroscientists, psychologists, computational psychiatrists and others interested in learning about and building models of the relationship between brain function, mental processes and behavior.
We have developed mappings between a number of model types in PsyNeulink and MDF, as a proof-of-concept for use of MDF in cognitive modelling.

More info icon

We are planning to implement MDF support for the Emergent simulator in the next phase of this project.

More info icon

We have developed an initial mapping to MDF from the ACT-R framework for modeling cognitive processes.

More info icon

We are planning to implement MDF support for the Nengo simulator in the next phase of this project.

More info icon

We are planning to extend MDF support to The Virtual Brain simulator in the next phase of this project.

More info icon

Computational models of neuronal processes inspired by the biological brain are a key target of the MDF format.

More info icon

Supporting models of cognitive processes will be a fundamentally important part of the design of the MDF format.

More info icon

Network models generated by machine learning frameworks such as PyTorch and TensorFlow, and serialised in ONNX format, are an important target of MDF.

More info icon

MDF's support for multiple export formats mean that the models can be executed across a range of conventional hardware platforms.

More info icon

We are actively investigating mapping of models in MDF format to quantum hardware with our collaborators in Track C of the NSF Convergence Accelerator program

More info icon

The goal of the MDF is to provide a common exchange format that allows models created in one environment that supports the standard to be expressed in a form – and in sufficient detail – that it can be imported into another modeling environment that supports the standard, and then executed in that environment with identical results, and/or integrated with other models in that environment.

More info icon

ONNX (Open Neural Network Exchange) is a well-established and actively supported standard in the machine learning community.
We have leveraged this, by developing a proof of concept interface between MDF and ONNX, showing not only that MDF can complement ONNX as an exchange-oriented format among machine learning environments, but also bridge to ONNX from environments in other domains.

More info icon

NeuroML is a widely accepted standard for describing biophysically-realistic models of neurons, neuronal subcomponents, and neural circuits in machine readable form, that is supported by a large number of existing programming and execution environments.
We have shown that MDF can be used to encode neuroscientific models created in NeuroML, allowing such models to be run on a wider range of simulation platforms.

More info icon

We are planning to extend MDF support for mapping to and from the SONATA format for large scale neuronal modeling.

More info icon

PyTorch is a widely used framework in machine learning and we have developed an initial mapping between MDF models and PyTorch

More info icon

We are planning to extend MDF support to TensorFlow in the next phase of this project.

More info icon

PsyNeuLink is an open-source, software environment written in Python, and designed for the needs of neuroscientists, psychologists, computational psychiatrists and others interested in learning about and building models of the relationship between brain function, mental processes and behavior.
We have developed mappings between a number of model types in PsyNeulink and MDF, as a proof-of-concept for use of MDF in cognitive modelling.

More info icon

We are planning to implement MDF support for the Emergent simulator in the next phase of this project.

More info icon

We have developed an initial mapping to MDF from the ACT-R framework for modeling cognitive processes.

More info icon

We are planning to implement MDF support for the Nengo simulator in the next phase of this project.

More info icon

We are planning to extend MDF support to The Virtual Brain simulator in the next phase of this project.

More info icon

Computational models of neuronal processes inspired by the biological brain are a key target of the MDF format.

More info icon

Supporting models of cognitive processes will be a fundamentally important part of the design of the MDF format.

More info icon

Network models generated by machine learning frameworks such as PyTorch and TensorFlow, and serialised in ONNX format, are an important target of MDF.

More info icon

MDF's support for multiple export formats mean that the models can be executed across a range of conventional hardware platforms.

More info icon

We are actively investigating mapping of models in MDF format to quantum hardware with our collaborators in Track C of the NSF Convergence Accelerator program

More info icon

Interaction with the wider computational community is essential to the development of the Model Description Format

We have hosted a number of workshops as part of this process to engage the wider community, and developed partnerships with other teams in the NSF Convergence Accelerator Program.

Community

Experts in areas of cognitive science/psychology, computational neuroscience, neuroinformatics, brain imaging and cognitive/symbolic modeling, systems and AI architecture, and Machine Learning participated in discussions on the feasibility of a Model Description Format (MDF) as posited by the ModECI team. The overall consensus - there is a real need for a format like MDF and each participant saw a practical benefit of MDF furthering convergence science in their respective domains.

Attendees:

  • Anton Arkhipov, (Allen Institute)
  • Trevor Bekolay (Applied Brain Research)
  • John Griffiths (Griffslabs)
  • Anurag Khandelwal (Yale University)
  • Christian Lebierre (Carnegie Mellon University)
  • Maryann Martone (University of California, San Diego)
  • Randy McIntosh (Rotman Research Institute)
  • Brian Nosek (Center for Open Science)
  • Jim Pivarski (Princeton University)
  • Russ Poldrack (Stanford University)
  • Fred Rothganger (Sandia labs)
  • Carole-Jean Wu (Facebook’s AI Infrastructure Research)
  • Yuan Yu (Microsoft)

Domain experts gathered to deliberate the impact an MDF could have on the cognitive science process-modeling community. Rich discussions solicited insightful feedback and cautionary tales from participants’ experiences, thus leading to a better-informed prototype for the ModECI MDF.

Attendees:

  • Christian Lebierre, Carnegie Mellon University (Act-R)
  • John Laird, University of Michigan (Soar/Common Model)
  • Pat Langley, Stanford University (ICARUS)
  • Rick Lewis, University of Michigan (ACT-R/Soar/CORE)
  • Frank Ritter, Pennsylvania State University (ACT-R/Soar/Herbal)
  • Paul Rosenbloom, University of Southern California (Sigma graphical model/Common Model)
  • Terry Stewart, National Research Council Canada (ACT-R/Nengo)
  • Andrea Stocco, Institute for Learning and Brain Sciences (ACT-R/Common Model)

The ModECI team hosted their second focused group workshop, working with the Artificial Intelligence and Machine Learning/Deep Learning communities to identify user-needs and areas of growth for the project.
Experts in several areas including Python, PyTorch, and IR discussed the potential for the MDF while also providing valuable insight into future pathways.

Attendees:

  • Ryan Adams, Princeton
  • Jim Pivarski, Princeton
  • Carole Jean Wu, FAIR
  • Michael Suo, FAIR

The ModECI team hosted their third Domain-specific focus group workshop, working with the Neuroscience international community. The objective was to identify end-user needs and areas of growth for the project. Several topics were discussed including the importance of identifying use-cases, precision, benchmarks, environments, and procedural coding. The lively discussion and debates provided much feedback for the ModECI team to consider in building the prototype and for future launching.

Attendees:

  • Anton Arkhipov, Allen Institute (Sonata)
  • Kael Dai, Allen Institute (Sonata)
  • Nora Abi Akar, Swiss National Supercomputing Centre (Arbor Project)
  • Ben Cumming, Swiss National Supercomputing Centre (Arbor Project)
  • Hananel Hazan, Tufts University (Arbor Project, BINDSNET)
  • Brent Huisman, Institute for Advanced Simulations (Arbor Project)
  • Sam Yates, Swiss National Supercomputing Centre (Arbor Project)
  • Marmaduke Woodman, Institut de Neurosciences des Systèmes, Aix Marseille Université (The Virtual Brain)
  • Spase Petkoski, Institut de Neurosciences des Systèmes, Aix Marseille Université (The Virtual Brain)
  • Michael Schirner, Charité University Medicine Berlin (The Virtual Brain)
  • Petra Ritter, Charité University Medicine Berlin (The Virtual Brain)
  • Robert McDougal, Yale University (Neuron)
  • Ted Carnevale, Yale University (Neuron)
  • Sharon Crook, Arizona State University (NeuroML)
  • Marcel Stimberg, Sorbonne Université (Brian)
  • James Knight, University of Sussex (GeNN simulator)
  • Thomas Nowotny, University of Sussex (GeNN simulator)
  • Terry Stewart, National Research Council Canada (ACT-R/Nengo)
  • Charl Linssen, Institute for Advanced Simulations (NESTML)
  • Jerry Skefos, Metacell

The ModECI team hosted their fourth domain-specific focus group workshop, working with the quantum computing community. The objective of this workshop was to explore the capacity for MDF to use quantum technology. The group identified several areas for convergence and collaborative efforts were solidified.

Attendees:

  • Lev Bishop, IBM
  • Ali Javadi-Abhari, IBM
  • Tom Marshall, Bloomberg
  • Andras Bako, Bloomberg
  • Pranav Gokhale, Super.tech
  • Victory Omole, Super.tech
  • Teague Tomesh, Super.tech and Princeton University
  • Kaitlyn Smith, University of Chicago
  • Gokul Ravi, University of Chicago
  • Rich Rines, University of Chicago
  • Yongshan Ding, University of Chicago
  • Sebastian Will, Columbia University
  • Minho Kwon, Columbia University
  • Max Aalto, Columbia University
  • Mahmut Kandemir, Penn State University
  • Swaroop Ghosh, Penn State University
  • Mike Pozmantier, NSF
  • Chaitanya Baru, NSF

The ModECI team hosted their final focus group workshop. The goal was to examine options for an open-source exposure. Suggestions from the participants included: joining professional group forums such as the International Neuroinformatics Coordinating Facility (INCF), development of strong governance model for the initiative, and hosting activities such as hackathons to solicit community engagement while also stirring community excitement.

Attendees:

  • Maryann Martone, University of California, San Diego
  • Russ Poldrack, Stanford University
  • Mike Pozmantier, NSF

Sponsor: University of California-San Diego

  • Jingbo Shang (Principal Investigator)
  • Rajesh Gupta (Co-Principal Investigator)
  • Lucila Ohno-Machado (Co-Principal Investigator)
  • Arun Kumar (Co-Principal Investigator)
  • Giorgio Quer (Co-Principal Investigator)

Sponsor: Vanderbilt University

  • Janos Sztipanovits (Principal Investigator)
  • James Pipas (Co-Principal Investigator)
  • Douglas Norris (Co-Principal Investigator)
  • David Smith (Co-Principal Investigator)
  • Ethan Jackson (Co-Principal Investigator)

Sponsor: Howard University

  • Claudia Marin (Principal Investigator)
  • Qiang (Shawn) Cheng (Co-Principal Investigator)
  • Jale Tezcan (Co-Principal Investigator)

Sponsor: University of Iowa

  • Stephen Baek (Principal Investigator)
  • Daniel Rubin (Co-Principal Investigator)
  • William Street (Co-Principal Investigator)
  • Xiaodong Wu (Co-Principal Investigator)
  • Paul Chang (Co-Principal Investigator)

Sponsor: Columbia University

  • Sebastian Will (Principal Investigator)
  • Alexander Gaeta (Co-Principal Investigator)
  • Nanfang Yu (Co-Principal Investigator)
  • Layla Hormozi (Co-Principal Investigator)

Sponsor: Pennsylvania State University

  • Swaroop Ghosh (Principal Investigator)
  • Nitin Samarth (Co-Principal Investigator)
  • Mahmut Kandemir (Co-Principal Investigator)
  • Sean Hallgren (Co-Principal Investigator)

About the ModECI project

MDF has been generously supported by the NSF Convergence Accelerator Program.

Learn More
ModECI (Model Exchange and Convergence Initiative) is a multi-investigator collaboration and currently involves:
PI: Jonathan D. Cohen

Jonathan Cohen is a cognitive neuroscientist with over three decades of expertise in computational modeling of brain and cognitive function, and contributions that range from models of the neural mechanisms underlying cognitive function (e.g., ServanSchreiber et al., 1990; Miller & Cohen, 2001; Shenhav et al., 2013) to models of higher level cognitive functions such as multitasking, planning, and relational reasoning that span from cognitive science to ML (e.g., Musslick et al., 2016; Segert et al., 2020; Ho et al., 2020). He has also been lead or co-lead developer for several large software development efforts, including: PsyScope, the first graphically-oriented tool for the design and implementation of computer based cognitive experiments (Cohen et al., 1993); BrainIAK, an open-source toolbox for the application of advanced machine learning methods to the analysis of brain imaging data (co-lead with Ted Willke, Intel); PsyNeuLink, an open-source, graph-based environment for developing models of systems-level brain and cognitive function (collaboratively with Bhattacharjee); and SweetPea, for the declarative description of factorial design and unbiased sampling of trials in empirical and modeling studies. Cohen will take primary responsibility for overall coordination of the project, as well as interactions with the cognitive science and cognitive neuroscience communities.

Tal Yarkoni

Tal Yarkoni is an informaticist with expertise in psychology, cognitive neuroscience and computational social science applications. He is the creator of the Neurosynth framework for large-scale automated meta-analysis of functional neuroimaging data (Yarkoni et al., 2011) and a number of other measurement and modeling techniques. He is a core contributor to the Brain Imaging Data Standard (BIDS) ecosystem, including principal contributions to the BIDS-StatsModels standard for machine-readable representation of fMRI statistical models, and co-developed the prototype BIDS-MDF spec with Cohen and Gleeson. Yarkoni has extensive expertise in formal standard development and automation technologies more broadly in the social and behavioral sciences (Yarkoni, et al.; Yarkoni, 2010). In addition to core contributions to feature prioritization and prototype development, Yarkoni will take the lead on identifying potential use cases in the social sciences, and developing and analyzing the surveys used to solicit community input.

Padraig Gleeson

Padraig Gleeson is a domain expert in biophysically detailed neural modeling, and is a lead developer of the NeuroML model description format (Cannon et al., 2014; Gleeson et al., 2010), as well as a long-standing editor of the standard. He is also involved in a number of neuroinformatics initiatives, including Open Source Brain (Gleeson et al., 2019), where standardized models in computational neuroscience can be shared with the community. He has collaborated with Prof. Cohen on the original development of BIDS-MDF, which included creating a functioning prototype of NeuroML to PsyNeuLink interoperability. He will be a main point of contact with the computational neuroscience community.

Abhishek Bhattacharjee

Abhishek Bhattacharjee is an experimental computer scientist with experience in heterogeneous hardware platforms and software environments. He will add expertise on the hardware/software portability and efficiency goals of this work. Bhattacharjee and Cohen have been working actively over the last three years on relevant software development projects, including compilation and LLVM IR representation of models in PsyNeuLink.

Ted Willke

Ted Willke is the director of the Brain Inspired Computing Lab, focused on the development of state-of-the-art machine learning algorithms that are inspired by neuroscience. He works on techniques that advance deep-learning based imaging analysis (Vyas et al., 2018), reinforcement learning (Haj-Ali et al., 2020), similarity learning (Ma et al., 2019), and language modeling (Turek et al., 2020), with a focus on hardware/software co-design problems and applications to systems research. He will help coordinate interactions with the machine learning community, focusing on exchange requirements among deep learning models, applications in machine learning, and their relationship to neuroscientific and cognitive models, and help oversee overall software design.

Sharon Crook

Sharon Crook is the co-director of the Informatics and Computation in Open Neuroscience Lab at Arizona State University. Along with Gleeson, she is a leader of the NeuroML model description format initiative (Cannon et al., 2014; Gleeson et al., 2010), and she has developed and maintains the NeuroML-DB portal, where models and their characterizations are shared for efficient selection of models and re-use on multiple simulation platforms. She is involved with several other neuroinformatics initiatives including SciUnit, a framework for validating computational models against experimental data or testing model/model agreement. She will contribute to the further development of the MDF and associated tools that provide bridges between computational neuroscience, machine learning, and other scientific communities.

Pranav Gokhale

Pranav Gokhale is the Co-founder and CEO of Super.tech, a quantum software company. His research focuses on bridging the gap from quantum hardware to practical applications. Pranav's work in quantum computing has led to over a dozen peer-reviewed publications, three best paper awards, three patent applications, and several software packages. Pranav will collaborate with ModECI to link MDF—particularly for graph problems involving Ising models—to quantum solvers. This work will bring the quantum computing community into the MDF ecosystem.

Terrence Stewart

Terrence Stewart is a Research Officer for the National Research Council Canada, as part of the University of Waterloo Collaboration Centre. His research is on how cognitive architectures can be implemented in the physical brain, emphasizing how dynamics, timing, and energy consumption put important constraints on the types of algorithms that are suitable for cognition. While his initial work was in ACT-R, for the last ten years he has worked with Chris Eliasmith at the University of Waterloo to adapt such algorithms to be implementable with neurons, using the Neural Engineering Framework to generate large-scale spiking models capable of performing basic cognitive tasks.

Get in contact

We are very keen to collaborate with members of the broader community. To contact the ModECI investigators directly, please mail: info -at- modeci.org.

Additionally, discussions around the standard and related software are taking place on our GitHub Discussions forum, where you can ask your own questions about the initiative.