Imbedding Deep Neural Networks
Authors:
Andrew Corbett,
Dmitry Kangin
Abstract:
Continuous-depth neural networks, such as Neural ODEs, have refashioned the understanding of residual neural networks in terms of non-linear vector-valued optimal control problems. The common solution is to use the adjoint sensitivity method to replicate a forward-backward pass optimisation problem. We propose a new approach which explicates the network's `depth' as a fundamental variable, thus re…
▽ More
Continuous-depth neural networks, such as Neural ODEs, have refashioned the understanding of residual neural networks in terms of non-linear vector-valued optimal control problems. The common solution is to use the adjoint sensitivity method to replicate a forward-backward pass optimisation problem. We propose a new approach which explicates the network's `depth' as a fundamental variable, thus reducing the problem to a system of forward-facing initial value problems. This new method is based on the principle of `Invariant Imbedding' for which we prove a general solution, applicable to all non-linear, vector-valued optimal control problems with both running and terminal loss. Our new architectures provide a tangible tool for inspecting the theoretical--and to a great extent unexplained--properties of network depth. They also constitute a resource of discrete implementations of Neural ODEs comparable to classes of imbedded residual neural networks. Through a series of experiments, we show the competitive performance of the proposed architectures for supervised learning and time series prediction.
△ Less
Submitted 15 February, 2022; v1 submitted 31 January, 2022;
originally announced February 2022.
A perspective on Microscopy Metadata: data provenance and quality control
Authors:
Maximiliaan Huisman,
Mathias Hammer,
Alex Rigano,
Ulrike Boehm,
James J. Chambers,
Nathalie Gaudreault,
Alison J. North,
Jaime A. Pimentel,
Damir Sudar,
Peter Bajcsy,
Claire M. Brown,
Alexander D. Corbett,
Orestis Faklaris,
Judith Lacoste,
Alex Laude,
Glyn Nelson,
Roland Nitschke,
David Grunwald,
Caterina Strambio-De-Castillia
Abstract:
The application of microscopy in biomedical research has come a long way since Antonie van Leeuwenhoek discovered unicellular organisms. Countless innovations have positioned light microscopy as a cornerstone of modern biology and a method of choice for connecting omics datasets to their biological and clinical correlates. Still, regardless of how convincing published imaging data looks, it does n…
▽ More
The application of microscopy in biomedical research has come a long way since Antonie van Leeuwenhoek discovered unicellular organisms. Countless innovations have positioned light microscopy as a cornerstone of modern biology and a method of choice for connecting omics datasets to their biological and clinical correlates. Still, regardless of how convincing published imaging data looks, it does not always convey meaningful information about the conditions in which it was acquired, processed, and analyzed. Adequate record-keeping, reporting, and quality control are therefore essential to ensure experimental rigor and data fidelity, allow experiments to be reproducibly repeated, and promote the proper evaluation, interpretation, comparison, and re-use. To this end, microscopy images should be accompanied by complete descriptions detailing experimental procedures, biological samples, microscope hardware specifications, image acquisition parameters, and image analysis procedures, as well as metrics accounting for instrument performance and calibration. However, universal, community-accepted Microscopy Metadata standards and reporting specifications that would result in Findable Accessible Interoperable and Reproducible (FAIR) microscopy data have not yet been established. To understand this shortcoming and to propose a way forward, here we provide an overview of the nature of microscopy metadata and its importance for fostering data quality, reproducibility, scientific rigor, and sharing value in light microscopy. The proposal for tiered Microscopy Metadata Specifications that extend the OME Data Model put forth by the 4D Nucleome Initiative and by Bioimaging North America [1-3] as well as a suite of three complementary and interoperable tools are being developed to facilitate the process of image data documentation and are presented in related manuscripts [4-6].
△ Less
Submitted 31 May, 2021; v1 submitted 24 October, 2019;
originally announced October 2019.
