2021 Fiscal Year Research-status Report
Scaling up CNN computations for data-intensive scientific applications
| Project/Area Number |
20K19823
|
|---|
| Research Institution | Kobe University |
|---|
Principal Investigator |
|
|---|
| Project Period (FY) |
2020-04-01 – 2023-03-31
|
| Keywords | Computation optimization / Computer Vision / Deep Learning / Non-Convex Optimization |
|---|
| Outline of Annual Research Achievements |
On the mathematical optimization side, the original analysis has been slightly redirected to applications surrounding neural implicit representations (NIR). Indeed, NIR have recently found important applications in signal compression, where it has achieved state of the art (SOTA) on 3D shapes, near SOTA results on video and extremely interesting compression rates on climate data (most notably on the ERA5 dataset). Unfortunately, these models are slow to converge. We thus focused our effort on applying the curvature surgeon to understand the convergence properties of these new models that promise several impactful implications to the world of science. Preliminary work on applying these models currently being submitted to the NeurIPS2022 conference.
On the computational side, we have focused our activity on exploring the limits of high-speed deep learning processing on both edge and server side. In particular, we have focused on the problem of extreme latency reduction with application to high speed camera denoising and on the limit of very large data processing using climate data. On the high-speed camera front, we have found CNN to be able to denoise under-exposed images coming from high-speed imaging in limited lightning conditions. This application is however limited in practice by the ability of software to process these frames with reasonable latency. Several works for this topic are currently awaiting reviews. On the climate data part, we have contributed to several hydrological studies that required application of models at large scale.
|
| Current Status of Research Progress |
Current Status of Research Progress
2: Research has progressed on the whole more than it was originally planned.
Reason
The project is progressing at what I would consider a relatively fast pace. However, original goals have been redefined in terms of the applications targeting. While the original applications we proposed are still progressing, we have found additional applications for the mathematical and computational optimizations we have been working one.
In addition, slowed down journal activities and travel restrictions have additionally added apparent slow down to our progress as several important work’s publications have been delayed due to the current sanitary situation.
|
| Strategy for Future Research Activity |
Recent successes in NIR research, as well as new found applications in earth science have opened new doors and slightly shifted our attention towards these more impactful applications. In the following year, we plan on summarizing our achievements on high-speed imaging, Neural Implicit Representation of Videos and climate studies into impactful publications.
|
| Causes of Carryover |
In pandemic times, travels have been very restricted and the proper use of the original budget was not possible. It would have been impossible to spend the allocated amount without wild deviation from the original plan. We have thus opted towards saving the money for the next year.
In addition, disturbances in journal organization have led several publications to be importantly delayed. I have good hope these contributions will be ready for publication before the end of the next year.
|
