Welcome to the July 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 autonomous driving, artificial general intelligence, deepfakes and much more. Let’s kick off with the comic of the month:
After conquering board games such as Chess and Go, video games such as Dota, computers now aim to beat card games. If your ethics allow it, it is now possible to become rich by letting a program play Poker for you.
Programs that could win against professionals in a 1 vs 1 setting have been there for a while. This article presents Pluribus, the first AI able to beat human experts in six-player no-limit Hold’em, the most popular Poker format in the world. It was developed by Carnegie Mellon University (CMU) in collaboration with Facebook. If the bets were real, it would have won around $1000/hour. Does it motivate you to learn more about it?
We have just seen a way of cheating in Poker, with a built-in program that can even beat professionals. Poker players, do not worry, as AI can also be used to fight fraud.
Classic techniques to fight fraud consist of rule-based engines and simple predictive models. However, security breaches nowadays are getting more and more complex and the previous systems cannot compete. The countermeasure to the adaptation of scammers is to adapt the defense systems to them, using AI. But what does it mean exactly?
With AI, fraud prevention can use both past experiences and current trends. Moreover, it can do so in real-time, compared to previous methods which could take 6 to 8 weeks to get the same results. Do you want to read more about the benefits of AI against fraud?
The pinnacle of AI, the one we see in every futuristic movie with intelligent robots, is the concept of Artificial General Intelligence (AGI). It represents a machine that has the same learning abilities as humans do. It could even surpass us by making connections across disciplines humans never saw.
AGI is a controversial term for many researchers. How can we define “general intelligence”? Is human intelligence even that “general”? To answer these questions and advance on this domain, OpenAI has received a $1 billion investment from Microsoft. They aim to tackle the technical difficulties of building an AGI and ensure a safe and secure deployment in society.
This partnership also implies that a lot of current OpenAI techs will become licensed. However, I believe that combining resources and intellect with some of the best researchers in AI can lead to many exciting discoveries.
Over the past decade, we have gone from looking at computer screens to interacting with them thanks to touch-screen technology. Children often believe that they can interact with every screen they see, because they grew up with it. Well, for tomorrow’s children, they may grow used to talking to the electronic devices around them.
Uber just developed Plato Research Dialogue System to build, train and deploy conversational AI agents. It helps deal with the different steps required for these agents to interact with humans. It is designed to be modular and flexible. Now, everyone, experienced or not with this domain, can play around and create their own conversational agents.
Conferences are wonderful to discover the latest development in AI. You might know some famous ones such as CVPR or ICLR. This month, the Conference on Uncertainty in Artificial Intelligence (UAI) took place in Tel Aviv.
Studies on uncertainty have increased a lot in the past years. Indeed, we want our models to be able to deal with intrinsic ambiguity. For example, let’s say I trained a neural network to classify different breeds of dogs. If I input an image of a cat, the model will output a dog’s breed, while I‘d rather have the network refusing to give a prediction.
This was only an example of how we could use uncertainty and probability in AI. Researchers at the University of Helsinki gave a method to sample from Directed Acyclic Graphs (DAG). They won the best student paper award. If you are interested in their method and familiar with DAGs, you can read their paper: Exact Sampling of Directed Acyclic Graphs from Modular Distributions. You can also read the winner of the best paper award: General Identifiability with Arbitrary Surrogate Experiments.
If you are interested in probability theory and AI, then you can dive deeper into the different papers and talks given at this conference. See their website: The Conference on Uncertainty in AI.
Almost all AI programs are specialized. They can see with Computer Vision, they can hear with Speech Recognition, understand language with Natural Language Processing. Technologies as vocal assistants have appeared to combine all these specialties into one unique agent.
Like every milestone, we can climb step by step and first solve an easier problem. To approach this goal of global assistant AI, Facebook researchers first conquered the famous video game Minecraft. They released CraftAssist, a platform to create bots able to assist the users by performing various difficult tasks. We can interact with the bot via an in-game chat. They receive orders in natural language, understand them and execute them.
This has tremendous applications in a future where the bot would be a moving robot, the chat would be our voice, and the video game would be the real-world. CraftAssist is an open-source platform. You can even play with the framework to build your own assistants.
Another way of pushing innovation in AI across the world, besides the release of public datasets, is to motivate people through challenges. On the other hand, unsupervised learning has been less studied than supervised learning, even though it is considered by many as the future of AI.
To balance out the equation, Facebook just announced the first session of the Facebook AI Self-Supervision Challenge, with a prize pool going up to $8000. Participants will compete in four different tasks: three tasks of Image Segmentation and one of Object Detection. A bonus of $500 is offered to winners who open-source their code. I believe it is another step in the right direction for research in AI. Sharing people’s work and research is always beneficial for science and knowledge.
Do you want to participate for the thrills of the competition, the price, or just follow the event? The competition started on July 25, 2019, and will close on October 15, 2019. Check out the challenge’s homepage: Facebook AI Self-Supervision Challenge.
As we have seen at the beginning of this best of with Poker, AI can also be used for malicious intents. As it grows more effective and more dangerous, this raises the recurrent question, nonetheless important, of ethics in AI.
This article talks about different uses of Deepfake. This is the name we give to the different techniques of faking content using AI, like videos or recordings. A famous example is the video of Mark Zuckerberg released in June where he talks about controlling the future with data. This realistic video is indeed a total fake and raises awareness of the efficiency of the technology. It is becoming harder and harder to spot the real from the fake.
It is an important debate to have: should we still be pushing AI researches while such malicious applications exist? What do you think?
Autonomous driving is still a milestone to overcome. The real world impact of driverless vehicles is limitless: from safety to energy consumption optimization, passing by comfort and time-saving. However, data has been lacking for self-driving cars, until now.
Lyft just released its brand-new dataset on autonomous driving. This is a major game-changer as it will help people from all around the world to move forward in this challenging task. Every time a big company shares its technology, it is a step in the right direction in the improvement of global human knowledge.
The dataset includes all types of information, from LiDAR raw inputs to high quality labeled bounding boxes of traffic agents. If you want to be the first to design a level 5 autonomous vehicle, or if, like me, you are just interested in the field, check out the lyft level 5 dataset.
GANs (Generative Adversarial Networks) have proven to be an effective generation paradigm. Researchers have delivered consistently on new results and experimentation with GAN, with landscape and face generation or even making photographs move.
Nvidia began thinking about more applications to GAN. Stanislas Chaillou believes we can use AI to help architects. In this article, he presents an AI he built upon GAN and Pix2Pix, a model to perform image segmentation. His program can convert a simple sketch to plans that contain building footprints (basically where we can walk), the emplacement of doors, windows and walls, and can even give the furniture layout of the floor.
This model can help architects solve highly constrained problems. It is also a proof of concepts of what could be done in the future to help architects using AI.
3 Steps to Improve the Data Quality of a Data lake
From Customising Logs in the Code to Monitoring in Kibana
Basics in R Programming
You are about to begin a project on R? Before you watch any tutorial, read these basic standards.
Git Branch Control when Deploying on AWS with Serverless Framework
Use a Serverless Plugin to check your Git branch before deploying on AWS.