Fig.
1: Probability model of the sky and its observation by a radio
telescope. The arrows represent stochastic influences that are
ultimately imprinted on the data. When reconstructing the sky signal
from the data using NIFTy5, these correlations must be traced backwards
in order to deduce causes from the observed effects. © South African Radio Astronomy Observatory; MPA
The Information Field Theory Group at the Max Planck Institute for
Astrophysics has released a new version of the NIFTy software for
scientific imaging. NIFTy5 generates an optimal imaging algorithm from
the complex probability model of a measured signal. Such algorithms have
already proven themselves in a number of astronomical applications and
can now be used in other areas as well.
Each day, a large number of astronomical
telescopes scan the sky at different wavelengths, from radio to optical
to gamma rays. The images generated from these observations are usually
the result of a complex series of calculations developed specifically
for each telescope. But all these different telescopes observe the same
cosmos – possibly just different facets of it. Therefore, it makes sense
to standardize the imaging of all these instruments. Not only does this
save a lot of work in developing different imaging algorithms, it also
makes results from different telescopes easier to compare, allows
measurements from different sources to be combined into one common
image, and means that advances in software development will directly
benefit a larger number of instruments.
The research group on information field
theory at the Max Planck Institute for Astrophysics has taken a big step
towards achieving this goal of a uniform imaging algorithm by
developing and publishing the NIFTy5 software. The research topic of
this group, information field theory, is the mathematical theory on
which imaging processes are based. Information field theory uses methods
from quantum field theory for the optimal reconstruction of images. The
latest version, NIFTy5, now automates a large part of the necessary
mathematical operations.
To begin with, the user needs to program probability models of the image
signal (see Fig. 1) as well as the measurement. For this, (s)he can
rely on a number of prefabricated building blocks, which often simply
need to be combined or only slightly modified. These modules include
models for typical signals, such as point or diffuse radiation sources,
or for typical measurement situations, which may differ in terms of
noise statistics or instrument response. From such a 'forward' model of
the measurement, NIFTy5 creates an algorithm to 'backwards' calculate
the original signal, which results in computed image. However, since the
source signal can never be determined uniquely, the algorithm also
provides a quantification of the remaining uncertainties. This is
implemented by providing a set of possible images: the greater
uncertainties the greater the differences in each area.
NIFTy5 has already been used for a number of imaging problems, the
results of which are published simultaneously. These include the
three-dimensional reconstruction of galactic dust clouds in the vicinity
of the solar system (see Fig. 2, an animation can be found here), as well as a method to determine the dynamics of fields based only on their observation (see Fig. 3).
On the strength of past experience, NIFTy5 not only allows new, complex
imaging methods to be generated much more conveniently, this software
package also includes a number of algorithmic innovations. For example,
the "Metric Gaussian Variational Inference" (MGVI) was developed
specifically for NIFTy5, but can also be used for other machine learning
methods. In contrast to conventional methods of probability theory, the
implementation of this algorithm in NIFTy5 does not require the
explicit storage of so-called covariance matrices. As a result, the
memory requirement increases only linearly not quadratically with
problem size, so that also gigapixel images can be calculated without
problems.
NIFTy Download
NIFTY stands for Numerical Information Field Theory. The eponymous information
field theory was originally developed for the analysis of cosmological
data sets. Thanks to NIFTy5, it can now be used in other scientific and
technical fields as well, such as medical imaging.
Author
Enßlin, Torsten
Scientific Staff
Phone: 2243
Email: tensslin@mpa-garching.mpg.de
Room: 010
Links:
personal homepage (the institute is not responsible for the contents of personal homepages)
Original publications
1. Enßlin, Torsten A.
Information theory for fields
Annalen der Physik 2019, 1800127
DOI
2. Knollmüller, Jakob; Enßlin, Torsten A.
Encoding prior knowledge in the structure of the likelihood
submitted to Journal of Machine Learning Research
Source
3. Knollmüller, Jakob; Enßlin, Torsten
Metric Gaussian Variational Inference
submitted to Journal of Machine Learning Research
arXiv:1901.11033
Source
4. Leike, Reimar H.; Enßlin, Torsten A.
Charting nearby dust clouds using Gaia data only
submitted to Astronomy & Astrophysics
Source
5. Frank, Philipp; Leike, Reimar H.; Enßlin, Torsten A.
Field dynamics inference for local and causal interactions
submitted to Physical Review E
