Welcome to the September 2018 edition of our best and favorite articles in AI that were published this month. We are a Paris-based company that does Agile data development. This month, we spotted articles about human posture estimation and generation and language. Let’s kick off with the comic of the month:
The first big piece of news this month is the new search tool made available by Google: Google Dataset Search. Despite being still in Beta, the new search tool indexes and gives access to a broad range of datasets. Coming from sources as diverse as Kaggle, the Nasa or just usual public data, if you are looking for some non-usual data, now you can just Google it.
September also saw the rise of a new machine learning library. With the release of TransmogrifAI, Salesforce brings us what is maybe the easiest to use library of the kind. When Tensorflow seems like an overkill to build a small predictive model, I will definitely give it a try. Oh! and it runs natively on top of Apache Spark.
On the side of academic research, this month has been prolific on papers generating human poses. The first two articles we spotted are about a really cool usage of Generative Adversarial Networks (GANs): performing “do as I do” motion transfer between two images. I sit, someone else dances, and the picture of me sitting becomes a picture of me dancing! Besides being an entertaining usage of neural networks, I find it instructive to read these articles. Compared side by side, they make you realize the versatility that artificial neural networks bring to solve machine learning problems.
Another impressive article about machine learning on human position that stood out this month came out of the MIT laboratories. And it claims no less than enabling computers to see through walls! Exploiting the readily available human position converters (which extract a human position from a picture of a person) and radio signals which can see through walls, they trained a model to extract a position estimation from the radio signals. The trick was to train the model converting the radio signal to a position first without a wall. Under that condition, the traditional image-to-position models can supervise the training of the radio-signal-to-position model.
Read the paper published in Computer Vision and Pattern Recognition (CVPR), 2018
Jumping to the machine translation research field, a major paper came out from Facebook labs with a work on unsupervised machine translation. Translating Spanish into English is easy because there are a lot of bilingual texts to train on. However what about translating some Italian text into Japanese? There are enough text datasets in both languages to build a lexical or grammatical model for each language separately but finding some bilingual texts may be more challenging. The idea from Facebook is to exploit mono-lingual language models to perform translation between languages with no bilingual data available, and they are very clever about it. Introducing various machine translation concepts such as back translations or word by word vs. language models, the article is worth a read.
Transitioning from language to voice, researchers from the Federal University of the State of Rio de Janeiro implemented an attention-based neural network model for speech command recognition. Attention-based model are recurrent neural networks which focus on a specific part of their input at each iteration. As they naturally divide a task into smaller and simpler tasks by focusing sequentially on different aspects of the data, I have a strong belief that they will bring some decisive improvements in the AI field. Despite being already widespread in image recognition, this article barely introduces the concept to the speech recognition research sphere and yet obtains some surprisingly good results.
Read the paper published by the Laboratory of Voice, Speech and Singing, from the Federal University of the State of Rio de Janeiro
Talking about speech, we also spotted a very interesting research paper coming out from Facebook labs on making chat-bots more engaging. Indeed, current chat-bots tend to be monotonous and have no strategy to re-engage conversation. According to the paper, the problem comes from the data used and they introduce a new more comprehensive dataset. They show how it improves the performance of the chat-bots and prove us that sometimes we tend to forget that the solution is maybe not in the model used but in the data.
Read the paper.
Uber’s articles are known for their quality and pedagogic value, and their latest one is no exception. Introducing the reader to the goals and principles of their forecasting, they explain how the number and place of customers can be predicted to adapt the supply of their driver-partners. However, forecasting at Uber doesn’t stop here. First article in a series to be continued, it is definitely a must-read.
Finally, September is the month of the Ignobels: a distinction given to research works which make you laugh, then think for a while. And I just couldn’t resist concluding this Best of AI by sharing this post presenting research by Facebook aiming to combine natural language processing and image recognition to detect hate speech in memes. At first building a “meme recognizer” may seem like a funny topic, but given the impact Facebook can have on the propagation of hate speech the topic is definitely more serious than it looks.
3 Steps to Improve the Data Quality of a Data lake
From Customising Logs in the Code to Monitoring in Kibana
Automate AWS Tasks Thanks to Airflow Hooks
This article is a step-by-step tutorial that will show you how to upload a file to an S3 bucket thanks to an Airflow ETL (Extract Transform Load) pipeline
Few-Shot Image Classification with Meta-Learning
Here is how you can teach your model to learn quickly from a few examples.