Mini-Workshop on Deep Learning and Artificial Intelligence Model Portability and Accessibility in Biomedical Applications

(Latest Draft)

Workshop Overview

This initial mini-workshop aims to bring together participants actively working in developing, creating and utilizing deep learning technologies in biomedical applications to foster greater collaborative efforts and exchange among the rapidly expanding community. As an initial mini-workshop, the participation in the workshop cannot be comprehensive, but strives to be at least partially representative to identify and begin taking steps to improve overall interoperability, portability and accessibility of advances in models and model technologies across the growing community.

Guiding elements for the workshop have been established using insights, input, and comments provided over the past many months as efforts involving biomedical applications of deep learning and artificial intelligence have grown exponentially. Significant guidance has been through collaborative interactions related to NCI-DOE collaborations such as the Joint Design of Advanced Computing Solutions for Cancer,  as well as insights and comments provided through various workshops, meetings and discussions.

The mini-workshop brings focus to two key challenge areas requiring effective interoperability of and access to deep learning and AI models. As a mini-workshop, only limited use cases can be considered.  The challenge areas are characterized by their respective use cases –

Use Case 1 – Use of deep learning models for purposes of transfer learning. In this use case, trained deep learning models are used as starting points for subsequent refinement with additional training data. The intended benefits of transfer learning include overcoming limitations in available data and/or available computing required to fully train a new model from inception.

Use Case 2- Use of deep learning models for inference. In this use case, the predictive/analytic capabilities of trained deep learning models are the focus. When used in an inference context, the deep learning models generate predictive/analytic outputs based upon provided input data.

The workshop aims to identify challenges, opportunities, resources and next steps that would enable greater collaborative interchange of deep learning models in support of the two framing use cases. While not fully enabling the breadth of possible use cases,  progress in these two driving use cases will have a broad impact across many necessary use cases.

Workshop Format

With driving use cases of inference and transfer learning serving to focus efforts of the mini-workshop, commonality has been identified in three key areas. 

Area 1 – Data Sources and Specifications. This area focuses on the inputs and outputs of the deep learning models and systems. Interoperability is enhanced greatly with increased coherence of data representations used when interacting with, using and developing these deep learning capabilities.

Area 2 – Model Descriptions and Representations. This area focuses on specifications for and depictions of the deep learning models and systems. Interoperability and interchange of created capabilities is enhanced greatly with increased opportunities to correctly utilize, transfer, implement, and execute the deep learning models and systems.

Area 3 – Interoperability, Portability and Accessibility. This area focuses on delivery of developed deep learning models and systems. With a focus on utilization, alignment of data representations and model definitions are not necessarily fully sufficient to support the effective use of deep learning models and model systems in transfer learning and inference. This area focuses on avenues to deliver the capabilities embodied in created deep learning models and systems for use in transfer learning and inference.

Presentations and Panel Discussions

Panel discussions and presentations help to establish the foundation for the ensuing breakout sessions. Experiences from different perspectives are shared with questions and answers to provide an increasingly common and informed understanding at a high-level leading into the breakout sessions. 

Breakout Sessions

The breakout sessions are intended to foster a rapid cross-sharing of insights, perspectives and ideas for taking next steps across the three areas described above.

An important element of both the Day 1 and Day 2 breakout sessions also includes identifying groups and efforts that are not included in the mini-workshop and strategies and approaches to engage and involve those efforts.

Day 1 Breakout Sessions

Attendees are arbitrarily broken into three assessment groups. The aims of the breakout sessions are to assess and identify across all three areas, contributing from their perspectives. Identify challenges, opportunities, issues, resources, existing capabilities, technologies, trends, good practices and other insights that would have potential to support and advance use of deep learning models and systems in the transfer learning and inference use cases.

Additional contributions are anticipated in the breakout sessions outside the scope of the two framing use cases, and outside the scope of the three areas. These contributions are meaningful and will be captured as supplemental perspectives and will likely serve as insights for future workshops.

Moderators of Day 1 Breakout Sessions are guided to gather the inputs shared during the breakout sessions and align the input into the respective areas – Data, Models, Access.

Between Day 1 and Day 2, inputs across the three breakout sessions will be coalesced into summaries in each of the respective areas.

Day 2 Breakout Sessions

Breakout sessions in Day 2 are Area focused. (Attendees are asked to sign-up in advance of Day 2 for the breakout session of greatest interest.)

The aim of the day 2 breakout sessions is to build upon the collective inputs provided in the three specific areas during the first day breakouts,  By reviewing the challenges, opportunities, issues, resources, existing capabilities, efforts, technologies and trends, the session would identify recommendations and next steps that can be taken, will be taken, and should be taken to make progress in the respective area for the session to better support and enable transfer learning and inference use cases among the community.

Expectedly, gaps will exist and differences of opinions will pervade the discussion.

Moderators aim to guide the discussion in terms of recommendations and next steps while capturing the insights provided, approximately reflective of the perspectives provided. 

Mini-workshop Deliverables and Outcomes

As the workshop is focused on fostering greater collaborations and exchange, major outcomes of the workshop include efforts across organizations to partner and work together to advance and enable the driving use cases for transfer learning and inference. 

A second major outcome of the mini-workshop is an identification of members of the community and efforts that are not represented among the community attending the mini-workshop, as well as avenues and actions to engage these members and efforts.

A third major outcome of the mini-workshop is a technical summary and report of the information provided and developed in the context of the workshop, to be used to inform the community, guide next steps and as valuable insights into additional issues and input to potential future workshops.