Akira’s ML News #Week48, 2020
Here are some of the papers and articles that I found particularly interesting I read in week 47 of 2020 (15 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
- Examples of Machine Learning use cases
- Other topics
— Weekly Editor’s pickup
— Past Articles
Week 47⇦ Week 48(this post) ⇨ Week 49
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
1. Machine Learning Papers
— —
GAN which generates high-resolution images with a small amount of data, small computation time
TOWARDS FASTER AND STABILIZED GAN TRAINING FOR HIGH-FIDELITY FEW-SHOT IMAGE SYNTHESIS
https://openreview.net/forum?id=1Fqg133qRaI
GANs that can train/generate high resolutions (256²~1024²) from scratch with small amounts of data (100~1000) and small computational complexity (1 GPUx10hours~). The technical crux are an SLE module that combines information at each resolution and a constraint that allows reconstruction from Discriminator’s intermediate feature maps.
Curriculum learning in many tasks is possible if you use a loss function to automatically determine the threshold of what is reliable
SuperLoss: A Generic Loss for Robust Curriculum Learning
https://papers.nips.cc/paper/2020/hash/2cfa8f9e50e0f510ede9d12338a5f564-Abstract.html
They proposed a loss function, SuperLoss, which can dynamically determine the confidence level of each sample and automatically determine the threshold of trustworthiness of the loss function for curriculum training. Because it can be added to existing loss functions, it can be used in various tasks such as classification, object detection, and regression, and was effective in tasks with label noise.
Leveraging the pre-trained model allows transformation by GAN with a small amount of data.
DeepI2I: Enabling Deep Hierarchical Image-to-Image Translation by Transferring from GANs
https://arxiv.org/abs/2011.05867
In Image-to-Image Translation using GAN, they proposed to utilize the existing trained models for Generator, Discriminator and Encoders (Encoders to read the source of translation are trained pre-trained Discriminator weights). Prior knowledge can be used, so less data is needed to make the conversion possible.
End-to-end framework for anomaly detection
A Transfer Learning Framework for Anomaly Detection Using Model of Normality
https://arxiv.org/abs/2011.06210
An end-to-end abnormality detection method. The abnormality is measured by the distance between the inspected image and feature distributions embedded in the normal image dataset with a training model. The distance between the distribution and the inspected image is measured using SSIM and so on.
Perform a style transformation by swapping layers
Resolution Dependent GAN Interpolation for Controllable Image Synthesis Between Domains
https://arxiv.org/abs/2010.05334
A study of style transfer using trained StyleGANs. A GAN that performs a style transformation by swapping the shallow part of a transfer-learned StyleGAN with the deep part of a regular trained StyleGAN on a new dataset, allowing for a Disney-like transformation of photos.
Solving physical puzzles in deep learning
Solving Physics Puzzles by Reasoning about Paths
https://arxiv.org/abs/2011.07357
A study of solving a puzzle in which one sphere is made to touch in target place using another sphere. Solve the task by having the user learn four models in supervised learning that predict the trajectory of the spheres with no action, the ideal trajectory of the spheres, etc..
Improving the generative model with the perceived loss function
A Loss Function for Generative Neural Networks Based on Watson’s Perceptual Model
https://arxiv.org/abs/2006.15057
Research to improve the quality of the generated images by using an improved version of the Watson model’s loss function that more closely resembles human perception in the generation model; confirmed that applying it to VAE produces high quality images with less blurriness.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
2. Technical Articles
— — — —
Careful explanation of machine learning model interpretability
The article explains why interpretability is important, and functions such as model transparency, with careful explanations and diagrams from the definitions. For example, with regard to transparency, the article explains factors such as whether a human being can reason through the same steps as the model, and whether each step is interpretable.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
3. Examples of Machine Learning use cases
— — — —
Planting trees with the help of AI.
The heat island effect becomes a public health concern, but it can be prevented by planting trees in cities. Google’s TreeCanopy Lab can use aerial images and machine learning to create a map showing the density of tree cover in a city. This will eliminate the need for manual tree surveys. Tree Canopy Lab has a short-term goal of planting and maintaining 90,000 trees by 2021 and continuing to plant 20,000 trees per year in a city of more than 503 square miles
Protecting Animals from Poaching Using Machine Learning
Google and the international conservation charity ZSL have built a machine learning model that uses machine learning to identify gunshots. Acoustic sensors can detect gunshots from up to 1 km away, thereby assisting wildlife conservationists in their work.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
3. Other Topics
— — — —
Machine Learning Discrimination and the Law
Machine learning models can unintentionally be potentially discriminatory depending on the training data. A simple countermeasure would be to adjust for this, such as scoring minorities, but the current law makes this difficult.
Generating a Monster GAN
A GAN for automatic generation of game monsters. As existing datasets of illustrations could not be used for diversity and copyright reasons, so they created a dataset using a 3D model, and the number of creatures, their structure, modeling, and proportions for each part of the creature are clearly indicated by the mask.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
— Past Articles
Week 46⇦ Week 47(this post) ⇨ Week 48
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —
