Akira’s ML News #Week46, 2020
Here are some of the papers and articles that I found particularly interesting I read in week 46 of 2020 (8 November~). I’ve tried to introduce the most recent ones as much as possible, but the date of the paper submission may not be the same as the week.
Topics
- Machine Learning Papers
- Technical Articles
- Examples of Machine Learning use cases
- Other topics
— Past Articles
Week 45⇦ Week 46(this post) ⇨ Week 47
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
1. Machine Learning Papers
— —
Stabilize the GAN by discretizing the middle layer representation
Feature Quantization Improves GAN Training
https://arxiv.org/abs/2004.02088v2
A study to improve the performance of a GAN by discretizing the feature map of the Discriminator intermediate layer output by replacing the nearest neighboring feature among the features stored in a dynamic dictionary. It can be easily incorporated into existing GAN frameworks and improves the performance with a slight increase in computational cost.
The underspecification of the solution leads to performance degradation in actual operation
Underspecification Presents Challenges for Credibility in Modern Machine Learning
https://arxiv.org/abs/2011.03395
They show that the problem of performance degradation of ML models deployed in real-world is related to underspecification, in which there are multiple solutions (combinations of model parameters) with the same predictive performance. This underspecification appears in all fields such as NLP, medical imaging, and computer vision, and should be tested with these parameters in mind.
Domain adaptation with a constraint that the nearest neighbor patches across domains are the same class.
Pixel-Level Cycle Association: A New Perspective for Domain Adaptive Semantic Segmentation
https://arxiv.org/abs/2011.00147
A study of domain adaptation in semantic segmentation task by constraining the labels of randomly extracted source patches -> its nearest neighbor target patches -> its nearest neighbor source patches to be the same. The results greatly exceed those of previous studies.
Detecting Deep Fake with heartbeat
DeepRhythm: Exposing DeepFakes with Attentional Visual Heartbeat Rhythms
https://arxiv.org/abs/2006.07634
This research detects Deep Fake by detecting changes in skin color due to heartbeat. The current generative model can make realistic movies, but it cannot reproduce the rhythm of subtle changes in skin color due to blood.
Use ODE to stabilize the GAN
Training Generative Adversarial Networks by Solving Ordinary Differential Equations
https://arxiv.org/abs/2010.15040
The instabilities in training GANs arise from the integration error in discretising the continuous dynamics and they experimentally verify that well-known ODE solvers (such as Runge-Kutta) can stabilise training. Even without SpectralNorm, the learning was stable and produced excellent results.
Consistent frame-by-frame image processing
Blind Video Temporal Consistency via Deep Video Prior
https://arxiv.org/abs/2010.11838
To solve the problem that frame-by-frame image processing produces inconsistent videos, they proposed a simple method of learning image processing by CNN. Their method eliminates the need for regularization and large amount of data to correct for the inconsistencies in previous studies. At first glance, it’s not clear why it works, but they think of the turbulence in the video as noise, and the prior distribution without noise (the main part of the video) is reproduced before the noise in the trained network, resulting in no turbulence.
Permutating the weights to improve network compression performance
Permute, Quantize, and Fine-tune: Efficient Compression of Neural Network
https://arxiv.org/abs/2010.15703
The study of quantization by permutating the weights to make them more compressible. It divides the permutated weights into blocks and replaces them with a nearest neighbor in set of vectors stored without using them directly. The ResNet50 can be compressed to 1/31th of its original size while maintaining accuracy.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
2. Technical Articles
— — — —
Computer Vision Notable Research 2020
It picks 10 computer vision papers published in 2020 that are likely to be of high importance and summarizes them in terms of overviews, the heart of the technology, the reaction of the machine learning community, and their use in business.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
3. Examples of Machine Learning use cases
— — — —
Using AI to manage the detailed progress of a construction site
Buildots, a UK-Israeli startup, can monitor the condition of about 150,000 components (at three to four levels, such as installed) from a 360° camera mounted on its head. It’s already being installed on small construction sites and is expected to allow human managers to focus on more important tasks rather than monotonous tasks such as checking.
US government subsidizes medical AI usage fees
CMS, an organization affiliated with the U.S. Department of Health and Human Services, announced that it will pay for the use of two systems that diagnose complications of diabetes that can cause blindness, and a system that warns patients of strokes through brain scans. The decision could help promote the use of AI more broadly in healthcare.
Machine learning to detect small earthquakes
Detecting small earthquakes from sensor data is a task that takes experts months to analyze, but with the use of machine learning, it can be done in 20 minutes. By analyzing such small earthquakes, the three-dimensional structure of the fault can be determined, which will lead to countermeasures against large earthquakes.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
4. Other Topics
— — — —
Pre-training on image data sets without copyright/discriminatory elements
A project to create datasets using fractal structures that are frequently found in nature. Large datasets on the Web may be subject to copyright and image rights issues. Instead of such natural images, we can construct datasets with fractal structures that frequently appear in nature and use them for pre-training to avoid such problems and learn important structures.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
— Past Articles
Week 45⇦ Week 46(this post) ⇨ Week 47
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
