I was fortunate to be able to attend NeurIPS 2018, the largest artificial intelligence conference in the world! Since the tickets were sold in 11 minutes, I applied to be a volunteer during the event with a letter of recommendation, as requested by the organizers. Even with that in hand, I was refused as a volunteer!
You can understand the joy that I experienced when Element AI contacted me, and offered to give me an all access ticket to attend with them. Since they were sponsors this year, they also offered a ticket to my friend Kristof!
This article is sponsored by Element AI, but all the content and opinions mentioned are my own!
Monday, December 3, 2018
The event began December 2nd. For me, Monday was my first day at the conference. As the week of the event coincided with my last week of classes at the university before the final exams, we could say that I was very busy! So I preferred doing homework on the first day of the event.
8:30am – 10:30 am : Visualization for Machine Learning
Visualization for machine learning was the first session that I attended! Presented by Fernanda Viégas and Martin Wattenberg, I learned a lot from this conference. I was quite surprised that I knew much of the first half of the presentation, having had a class at the university about chart making, and different ways of expressing information (size, color, shape, etc.)
On the other hand, the second part of their presentation focused more on the visualization linked more specifically to machine learning, and I learned a lot!
For example, both speakers mentioned that it might be convenient to graph the input data of the model. This may make the dataset more accessible, and one can then understand, in the case of animal images, that a frog has been classified by humans as a cat, while our system thinks that is not a cat! This example comes from Quick, draw! from Google.
Another example that struck me in the Quick, Draw! Dataset is that according to people’s cultures, they do not draw the same objects in the same way. Even when it comes to something like a chair! It becomes apparent when you look at the drawings that are drawn according to the country of the players:
As you can see, identified by “TR” in the image above, Turkey is the only country that draws a fish with the tail on the left. Other countries do it either right or both sides! Therefore, visualizing the data may allow us to notice these kinds of differences in our data.
They also mentioned a resource for teaching artificial intelligence to children. You can play Quick, Draw! and make the 6 drawings as requested. Then you can choose a particular design and the site will explain the reasoning used in order to draw a conclusion on what was drawn. Try it, you will understand!
Watch the whole presentation on Facebook
Resources mentioned during the presentation:
- Slide with all their recommended resources
- Example of data visualization related to stock exchange
- NY Times publishes outstanding examples
After this presentation, I spent some time networking at the Women in Machine Learning workshop. I also spent several hours visiting the kiosks of the sponsors of the event! I was able to meet Catherine and Camille as well as my friend Hannah, who presented her poster as part of the workshop!
6 pm – 11 pm – Women in Machine Learning – Amazon Party
I was able to go to the Amazon party! It was a beautiful evening, where I met Merel and Michelle, two people who follow me on social networks! It was an open bar party, and with great food, so we had a good time with friends!
Tuesday, December 4, 2018
At 10:30 am, I had the chance to meet with a member of the Element AI team. We had lunch together, and it was lovely! I had a good time learning more about the company and discussing with her!
Then I went for a tour of the posters at the conference. On the other hand, it was challenging, as it was jam-packed. I was looking for Element AI’s “Towards Text Generation with Adversarially Learned Neural Outlines” poster, but I never found it! There were 172 posters I think, so it was hard to find.
After that, I went to work a little. I was lucky enough to choose a place where people were sitting next to me to eat their lunch, so I had the chance to chat and network with 3 people in a very short amount of time! 🙂
Diversity and Inclusion Town Hall Meeting
After lunch, I, unfortunately, missed the talk about inclusion. Fortunately, though, I was able to listen to it via their videos on the Facebook page of the conference! It was a panel with people from Queer In Ai, Black in AI, Latin in AI (WiML), and two members of the NeurIPS Foundation board.
Why is diversity important in AI?
Machine learning algorithms have biases for minorities. People can be hurt physically if we design products that have not been tested or designed by a group of people representing the diverse population. We conceive products that affect different groups of people. We have to make sure that everyone is represented so as not to harm them.
Why are groups like Women in Machine Learning, Queer in AI, Black in AI important?
The representative of Queer in AI mentioned that before the creation of the group, he knew no queer researcher who did not commit suicide several years ago … it touched me! I have a friend who told me the same thing recently, and I confess that there is no one that comes to my mind … I met some people from Queer in AI, and it’s a beautiful group!
Why do we need these groups with women or queer exclusively, etc.?
There is a sense of empowerment that comes with being together in a room with people who have the same problems as us. It helps to know that we can be ourselves without overthinking. Also, seeing individuals like us that are further along in their career is inspiring as they act as a role model for others.
This means that sexism, racism or homophobia cannot be studied separately. They cannot be fully explained alone. For example, a black woman has to face the fact of being a woman in a man’s world, but also that she is black. Both are hand in hand and can not be separated. For example, people of color in the United States do not face the same realities as people in Africa when they come here, said the representative of Black in AI.
For example, in the United States, people of color face more criminal problems with the police, such as being falsely accused. While Africans, when they come to North America, have more problems with their visas and poverty.
We all have unconscious prejudices. We can not be perfect, however, to be an ally to the minorities, to be aware of these prejudices and bias is the first step. That does not make us a monster.
Track 2: Algorithms, Theory, Optimization
Then, I attended track 2 on “Algorithms, Theory, Optimization”. This is 14 talks that are 5 minutes each. I really like this formula where we go on a surface subject only! It’s easier to absorb! It’s not so long that I’ve been interested in machine learning and artificial intelligence, so I do not understand everything that’s being said since I’m just getting started and those who are presenting are experts in their fields.
However, I am happy to see that I understand more and more, every time I attend a new event, I see my progress! That’s what matters!
Wednesday, December 5, 2018
Reproducible, Reusable and Robust Reinforcement learning
Guest talk by Joelle Pineau on “Reproducible, Reusable and Robust Reinforcement Learning.” I was interested in listening to a lecture by such an inspiring woman. Also, I do not know much about reinforcement learning since I have not yet had a course that has addressed this type of learning.
Reinforcement learning is about rewarding good decisions, much like a child. The intelligent agent, therefore, aims to accumulate the rewards and thus make as many choices as possible.
She showed this interesting graph that illustrates the number of papers on reinforcement learning published each year; it’s increasing enormously!
Also, as her talk is about experiences that can be replicated, these results were surprising to me:
To be able to reproduce the machine learning experiences that have been made in the papers, it is important to be able to learn and confirm the conclusions of the authors of the paper. It must be shown that it is not just luck and that the figures obtained are true. I myself almost reproduced the results of a paper published, but the hyperparameters were not all shared and it lacked details so that I could do it to learn.
She suggests a “reproducibility checklist” that can be used when sending papers to conferences to ensure that her article is reproducible. I think the checklist is not yet officially available, but it’s around the 30th minute of the video.
Investigations into the Human-AI Trust Phenomenon, par Ayanna Howard
Lesson # 1: Humans trust socially-interactive (emotional) robots
One hundred fifty million children have disabilities. 15% of the world’s population lives with a disability. Being such a large part of the population, this is the researcher’s area of focus.
In the 20th minute of the video, Ayanna mentions that through her research with robots and their interactions with children, she discovered that if the robot is expressing an emotion of disappointment, the concerned human will change its behavior to make the robot happy. It’s quite scary to think that an intelligent agent (the robot) can influence the behavior by expressing an emotion that can very well be intentionally generated to manipulate the human… A reasonably certain conclusion was reached as they experienced it repeatedly.
Through this emotional connection, they can increase people’s trust in robots. They intentionally change the behavior of the human for “their well-being” …
In the 33rd minute, she mentions that in 8 weeks, with 2 sessions a week, if I am not mistaken, they improved the movements of children with mild to moderate disabilities.
Humans even say they trust robots more than human therapists.
Lesson # 2: Humans trust robot in emotionally-driven scenarios
Source: Robinette, et al, “Overtrust of Robots in Emergency Evacuation Scenarios,” ACM/IEEE HRI, 2016.
They did a study where the study participants entered a building. They are guided by a robot into a room, where they come in and their task is to read a scientific paper. As humans know they are participating in a study, they try to remove their bias by making them believe that their primary task is to read the paper in the room.
First, the researchers fill the room with smoke. The participants then have to leave the room and move to evacuate the building. The robot is placed in an area that is far from the exit that the participants used to enter. The results surprised the researchers : 100% of the participants decided to follow the robot.
Then, the researchers tried to lower the number of people that would follow the robot. They inserted errors in the robot, for example, the robot that rotates in 360 *, goes to random places, etc. Once again, 100% of humans continued to follow the robot in an emergency.
What is the problem? Humans trust and humans have biases, and machines are influenced by their creators.
They experimented with a white robot and a black robot. When they evaluated the trust of the people who interact with them, they discovered that people do not trust the black robot. They did the same thing with female robots and with different nationalities. Although it is a machine, it has the same programming as white, but humans do not have the same interactions according to the physical appearance.
Ayanna Howard has a podcast called The Interaction Hour in collaboration with the Georgia Tech School of Interactive Computing to explain to people what can terrify them about technology: facial recognition algorithms, autonomous cars, and so on.
Element AI Evening
Element AI invited me to their Wednesday night party and Thursday night networking event. I really spent two beautiful evenings networking with researchers that are very inspiring in machine learning! I am so grateful to have had these great opportunities to network!
Lessons learned for my first year at NeurIPS
I have participated in several technological events and conferences in recent years, but NeurIPS is undoubtedly on another level. It’s so big and more scientific than my previous experiences. This led me to make some mistakes. I would like to share them with you so that you do not do the same!
Make a plan of the talks that interest you
With NeurIPS being such a big conference and having so many posters and talks going on at the same time; I realized that I should have had a better look at the talk schedule. For example, I missed all the talks about diversity. If you know me, you know very well that this is probably my favorite subject!
I would also say that it is essential to plan the talks you want to attend beforehand. I should have established a more precise schedule at the beginning of the conference to identify the different meetings that are important to me. What I did wasn’t enough. I told myself that I would look at the schedule regularly during the day. However, there is so much going on at the same time that it’s not the right time to go over the talks and posters.
Visit companies during a conference that interests you less
Networking with the companies present is a good way to expand your network of contacts. It also allows you to discover exciting companies that you didn’t know before. On the other hand, during the coffee breaks, all the participants go to the kiosks of the companies. There are so many people; it’s not as easy to get to know the companies during these moments!
It is possible to create much larger connections by having a little more time and fewer people all around!
Every day is exhausting: we cannot do everything
It is essential to identify some conferences that interest us the most and to conserve our energy and our concentration for those! I know that the fear of missing something, or “fear of missing out” in English can be very present when attending such a big conference. Especially when we know how lucky we are to manage to get tickets.
By cons, I think it is better to take a break. Maybe not to visit the conference for a whole day in the middle of the week if it can allow us to be back with 100% good energy the next day. (The conference was taking place from December 2 to 8, which is quite long considering that it starts at 8:30 am and ends at 11 pm with company events)
Keeping energy for the parties as it is the perfect opportunity to network
I enjoyed going to the different parties organized by the companies! Google X, Apple, Intel, Nvidia, Element AI, and more organized parties on different days! This is the perfect opportunity to decompress and network with the other participants, but also, and especially, with the companies present!
Meet participants according to our common interests
NeurIPS uses the “Whova” application to allow participants to view the event calendar. There is also a forum where it is possible to create discussions and activities with others to network according to our common interests! I have not had the opportunity to use this feature this year, but it’s something to use in the future! It is much easier to connect with people who have common interests.
For example, my friend Kristof went to dinner with a group of researchers working on referral systems in AI. He came back from his meal with lots of ideas for his research!
I had a lovely week. I am so grateful to Element AI for allowing me to attend this significant conference on artificial intelligence. It allowed me to learn a lot and met a lot of people. It was a very positive experience!
I cross my fingers for the chance to return to NeurIPS in the following years. I hope this article has been useful for you to learn what happened at the conference!