france is AI

I participated the conference “france is AI” in this Thursdayand Friday. In this conference, lots of companies and institutes talked abouttheir thinking of AI, like Google, Microsoft, INS, INSA Rou...

I participated the conference “france is AI” in this Thursday and Friday. In this conference, lots of companies and institutes talked about their thinking of AI, like Google, Microsoft, INS, INSA Rouen, ENSAE. Thanks to them, I know more about Data Science and AI, and I’d like to share with you some interesting points.

What is AI?

Rather than the natural intelligence (NI) of humans and other animals, artificial intelligence (AI, also machine intelligence, MI) is apparently intelligent behaviour by machines. In computer science AI research is defined as the study of “intelligent agents”: any device that perceives its environment and takes actions that maximize its chance of success at some goal. Colloquially, the term “artificial intelligence” is applied when a machine mimics “cognitive” functions that humans associate with other human minds, such as “learning” and “problem solving”. (source: Wikipedia)

In the conference, Microsoft thinks that AI is using big compute, powerful algo and massive data to attain more general intelligence, so that to master human-AI collaboration and to provide insights on AI, people and society.

Data Science Ecosystem

Data science is an interdisciplinary domain. Besides the usual challenges involving experts of two distinct domains, a successful data science project also need to include a third pole of software and system enfineers, which can implement the methods developed by data scientists, and maintain the tools. CNRS’s researcher showed us a complete data science ecosystem with the following figure.

data science ecosystem

This figure presents the roles of each actor. We don’t mean to isolate different actor, real profile can even overlap several roles at the same time. Due to limited space, I won’t go through in detail. You can read this blog to understand more about different actors.

Challenges

Since AI is a new technical domain, there are many challenges to address. Microsoft mentioned several keys challenges: learning about the world is an unsupervised manner, so it’s difficult to endow AI systems with commonsense reasoning that provide them with core understandings about the world; besides, we also need to imbue systems with key sets of social and emotional skills that people rely on in communicating, coordinating and collaborating with one another, and find method for endowing systems with knowledge about the surrounding physical world and interacting with people and objects in the real world.

Some other companies and institutes also talked about the challenges. PriceMinister faced attribute test normalisation as a machine translation problem. CNRS’s researcher talked about the data challenges with code submission. He pointed out the importance of coding prototype and of code review among team member. Because code review can help us learn from each other and collaborate with teammates. For this point, I’m totally agree with him.

Solution

Chain of data value

france is AI

By the figure above, DreamQuark mentioned the chain of data value. Since it’s easy to read, I won’t explain it here.

Deep Learning in timelines ranking

Twitter uses deep learning in timelines ranking. As we can imagine, Twitter has billions of data. In oder to ensure the functionality, Twitter customised code CPU/GPU to speed up, used the max normalization and MDL to make better ranking, and for fast training, Twitter used partitioned HDFS streamer and customised data deserializer.

How to improve data team’s productivity?

Twitter talked about the capacities that so-called “Machine Learning experts” should have: they should master pure research (model exploration, training and flexibility), applied research (dataset/feature exploration, flexibility and robustness) and production (feature engineering, data addition, robust training, deployment and A/B test). In order to increase experiments of ML expert, we should keep the following points in mind: automation and visualisation, which facilitate the reproducibility and refreshing of program, result testing, comparing and analyzing; creating an AI environment, which can help us to explore data, train fast and manipulate data.

Serena Capital, ISAI and Partech mentioned some important elements of a data team: technical background(IT + statistic) is necessary, full stack talent (data correction, integration of business), understanding application of algorithms, the problem to be solved, and the set of data.

Golden rules for AI-first product design

Before applying algorithm into your business or product, we should firstly well understand them. Tinyclues presented us their golden rule for AI product design: understanding your algorithms’ failure modes. That makes sense, only if you know its failure modes you can analyse the result in your case. Another rule is keeping an eye on long-term/ayatemic stability. This is important, many people don’t realise that we should not only think for this moment or short-term, but also should make a plan for long-term, so that we can regconise which points are more important for us.

AI rethinking Retail: Walmart Use Case

As a data scientist of retail business, it’s really inspiring to hear the speaking of the largest grocery retailer in the world, Walmart. They used AI to discovery digital relationships: how people buy? how products stock in the stores?

According to catalog of data, they observed the preference of clients, which contribute to recommendation and personalization program, and can ensure the stockage in fullfillment center. During the holidays, it’s a challenge for scanning capacity, because holiday traffic is 10 times more than usual. Moreover, in this intelligent world, many merchants sell their products online, so does Walmart. It’s really important to ensure the safety of clients’ information. In order to let clients shop safely, Walmart concentrates their efforts on web security.

Future of AI

As rising of AI, we have to face many challenges and try to find solution. Google’s AI researcher François Chollet said next frontier of AI will be abstraction, reasoning and planning, and should expand the model search space as well.

Thanks to this fantastic conference, it opens the door of AI to me, it makes me to realise the amazing technique of AI.