Source: Deep Learning on Medium
In this post we’ll go over some of the most interesting posts of the week. This is no intended to be a comprehensive summary of these papers, but a quick overview of why each of these are interesting and some additional resources to learn more about the topic.
A guide to deep learning in healthcare
This paper is worth reading since it gives a brief overview of how different Deep Learning techniques impact key areas of medicine. This paper goes over the use computer vision for medical imaging, Natural Language Processing (NLP) for electronic health record data, and Reinforcement Learning (RL) for robotic-assisted surgery. Why healthcare? Well, in the last few years there have been huge amounts of data in this field (150 exabytes, and growing!).
For Computer Vision, a typical example is the use of Convolutional Neural Networks (CNNs) to determine whether a patient’s radiograph has malignant tumors. But not only this, CV has promising results in complex image-based diagnostics. It is important to mention the limitations in the studies. When comparing human vs algorithmic performance, there has been a lack of clinical context, which is extremely useful to the humans.
For Natural Language Processing, Recurrent Neural Networks (RNNs) have been effective at processing sequential inputs. According to this paper, a single hospitalization can generate ~150,000 pieces of data, so there’s a huge potential for DL applications.
Suggestions to learn about the topic
- Transfer Learning. This topic is specially useful for CV methods. https://towardsdatascience.com/a-comprehensive-hands-on-guide-to-transfer-learning-with-real-world-applications-in-deep-learning-212bf3b2f27a
- Dermatologist-level classification of skin cancer. https://cs.stanford.edu/people/esteva/nature/
- Applying Deep Learning to Metastatic Breast Cancer Detection. https://ai.googleblog.com/2018/10/applying-deep-learning-to-metastatic.html
FIGR: Few-shot Image Generation with Reptile
This is an interesting paper that introduces an approach to GAN training for Few-shot image generation using Reptile (FIGR) and contributes with an interesting dataset for the same task. The application that inspired this paper is to create sketches based on few drawings. If an artist lacks inspiration, he/she can draw some sketches and FIGR can generate more.
The de facto dataset for few-image generation is called Omniglot. It contains 1,623 unique type of characters originated from 50 alphabets. Each of these characters was handwritten by 20 different individuals. FIGR-8 is the dataset invented in this paper. It has 1,548,944 images separated in 18,409 conceptually different classes, each class with at least 8 images. The images are of black-and-white icons.
Suggestions to learn about the topic
POET: Endlessly Generating Increasingly Complex and Diverse Learning Environments and their Solutions through the Paired Open-Ended Trailblazer
I’m including this since it is extremely interesting and opens doors for new problems. Open-endedness offers the potential to generate the own training data unlimitedly. POET was released this week. It creates new environments for OpenAI Gym. During training, POET generates more diverse and challenging environments, but, at the same time, optimizes agents to solve their paired environment.
Other interesting papers from this week
- Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context (https://arxiv.org/abs/1901.02860v1)
- Accurate, Data-Efficient, Unconstrained Text Recognition with Convolutional Neural Networks (https://arxiv.org/abs/1812.11894v1)
- A Theoretical Analysis of Deep Q-Learning (https://arxiv.org/abs/1901.00137v1)
- Deep Reinforcement Learning for Multi-Agent Systems: A Review of Challenges, Solutions and Applications (https://arxiv.org/abs/1812.11794v1)
Thank you for reading!