This evening at the Women in Machine Intelligence Dinner in San Francisco, RE•WORK were joined by guests keen to support women working in AI for a night of networking, drinks, a three course meal, and of course keynote presentations from leading minds in the field.
The evening, kindly sponsored by Facebook, kicked off with a champagne reception with attendees discussing their work in the field. Key discussions centralised around the importance of the rapid expansion of Machine Intelligence and its impact on multiple sectors including retail, manufacturing, transport, healthcare & security.
I’m most looking forward to the open collaboration of this evening. The event is bringing together different people from different places which you don’t get from a lot of events.
Richard Castro, MD FIXR
Once guests were seated, Nina D’Amato, Chief of Staff from The San Francisco Department of Technology, our compère for the night began by introducing the evening:
It’s fantastic to bring together experts in AI and Machine Intelligence, and there's an exciting collective women power in the room! I’ll say a special thanks to Facebook for making it possible, and I’m looking forward to hearing groundbreaking updates on their current research.
Nina’s role focuses on the strategy side of planning and development, implementation, and performance. She oversees the Project Management Team, Geographic Information Systems group, and the Salesforce team.
Between courses, guests were invited to rotate seats to maximise networking, and we heard some great discussions going on, as well as the areas of people's interests.
First up to present, Annie Liu, Research Scientist from Facebook started by exploring how Facebook approach the difficulty of understanding newsfeed content. Annie studied pre-med in college before changing to computer science in 2005, when computer science wasn't flashy or a big deal, but times have changed and AI is now so woven into society that 'experts in the area are rock stars!' Annie explained that at Facebook 'we have a pretty important and tough job, to bring people together rather than to alienate them.'
The newsfeed started over 10 years ago with a simple algorithm that has become more and more complicated over the years. It now considers clicks, likes, comments and shares. Earlier last year, Facebook realised they had a problem - clickbait content! Annie asked the attendees, 'Who’s seen things like ’10 ways to lose weight, point A will blow your mind!’? Most people in the room raised their hands. This is a super effective way for companies and individuals to get engagement on Facebook, but inadvertantly it's become a publishing platform which was unforseen. This is a problem as it exploits ranking to make money which provides little value and is a problem for users! Facebook started building a model to capture the problematic sentence structure. Annie said that they 'measured a decrease in the content but it also got us thinking ‘are all engagements created equal? Are we optimising all the right things?’ Facebook is now shifting back to focus on friends and family as well as focusing accordingly to create healthy conversations. It’s a big change! Facebook haven't had a big shift for many years but believe it’s the right things to do - 'what’s good for the business is good for the user and vice versa.' This means that personalisation is even more important. The second approach is to classify the contents into interests e.g. sports, ML, fashion, DL etc. Facebook use a hierarchical approach, starting at a high level and then categorising into sub topics. 'We're investingand it will be important in the long run. Usually when there are problems. the problem doesn't lie in the tech, but the metrics we’re optimising the technology for. If we’re building more AI into our products we have to me mindful that there’s an effect by optimising on a small set of metrics that’s changing and evolving and should be evolving with the world.'
We next heard from Cristina Scheau, one of the top ranked engineers at Facebook. She's currently working to help people find out what discussions are going on around things that matter to them. As users, we know this as Facebook search. Cristian started her work in AI 7 years ago when there was little media attention, but the journey has been exciting with the latest investments paired with older methods making AI more successful. People want to connect with friends through videos and photos to share with friends, or camera live streaming on Facebook. Social network data is very different to web data. There are different intents, and Facebook are doing their best to fulfil them. The main components of search are understanding and relevance, as well as understanding intent, specifically the type of intent. What’s the user after? Cristina used an exmaple: if you have a bunch of lady gaga photos and you need to find her and to know who she is and identify her in each image - how do you do this? Not long ago approaches in NLP were different and used a sequence level. In this, the lady gaga image would use conditional random yields, as thesewere the go to example. Is theis sentence beginning with a particular field? This has shifted into designing effective neural networks. NLP is still being used but, with neural networks incorporated into the sequence label.
Been Kim from Google Brain was next to present and began her presentation by expressing her thoughts on the event:
I'm humbled and happy to be here with an incredible group of women and men in MI in SF and in the world! Been explained that you really need to care about what you do otherwise you won’t finish it! Machine Intelligence is a very powerful tool that could be used to predict adverts, manage stock exchanges, decide how long you have to spend in jail, and predict and manage medical diagnosis. These industries have the potential to have life changing consequences, so it's important to ensure that you know exactly what you're doing. At Google Brain, it's important for them to democratise Machine Learning: 'We’re one of the special sets of people educated & privileged enough to come and have a great dinner. Lots of other people in the world don’t have that opportunity but everyone in the world should be able to find ML accessible. For example, if someone spent their life savings on setting up a store, I’d love them to leverage ML to decide the best coffee bean to sell - I want them to be able to leverage this powerful tool just like we do in research!
Been went on to open the discussion up to the attendees by asking if anyone could specify what we mean by safe autonomous car? Discussions centralised around the moral decisions of autonomous vehicles and and how we could build a unique set of tests: if you pass them you’ve built a car that’s safe -what would these tests be? It's really hard to decide what a classifier would be as we don't want machines to be biased, sexist or racist, and it's hard to define what we mean by these tests. That’s underspecification of the problem. Just because I talk about how important it is, you don’t need it all the time - if you just want prediction and there’s not much serious consequences you might not need interprebility at all. I don’t understand how to fly a plane, but I”m happy to ride in one most of the time because I trust the engineers. The systems that are verified like this don’t need interprebility.'
The floor was opened for questions, and we heard an attendee ask 'how do you define interprability?' Been explained that if your explanation helps achieve your final answer correctly you’ve achieved it. You need to use humans in A/B testing to see e.g. in a medical example, how many more people did you save?
A couple of us at Brain are working on a pretty different way of explaining NN, called TCAV (testing with concept activation vector). Instead of using input features, we use higher level 'concepts' to produce explanations, all without retraining the model or changing architecture of it. The goal is to connect granular complex language that an ML model speaks to a human understandable, high-level language.
Finally we heard from Shivani Rao from LinkedIn Learning Relevance Group. Since completing her undergraduate degree in computer science, AI has come on a long way! Shivani learned C then C+++ and python. By the time she was graduating people were using spark - the progressions are so hard to keep up with, with so many languages in a short time! Similar to this, learning and online learning platforms is something that’s expanding constantly. There are various reasons people may want online courses: careers and roles change, or you might just be wanting to keep up with the times and learn to be relevant. Shivani explained that it’s so rewarding to help people keep up to date!
I went to a shop in Mumbai near my house back in the 90s and bought a book called 'C+++ in 21 days'. This is all that was available to me at the time. Years later when I had internet in my house I could do it online: this is what I knew of learning! Most of my colleagues used correspondence learning (when you got books and quizzes in the mail). In 1995 Lynda.com started with Lynda - she wanted someone to make a video to help education. She started recoding for other likeminded people to learn, and in 2002 they started recording the videos and in 2016 this became LinkedIn learning.
LinkedIn is one of the largest professional networks worldwide providing a wealth of knowledge on individuals skills & educational backgrounds. The acquisition of Lynda brought learning to LinkedIn users helping everyone to learn. The courses are sorted by recomednadtion and search. Think about looking for a course or qualification online - It can get overwhelming so recommendations analysis narrow it down and help organise the results. This also helps in job search, profile views, LinkedIn news feed etc.
The evening drew to a close with exchanges of contact numbers and business cards, discussions around the presentations of the evening, and an air of excitement for the upcoming Deep Learning Summit and AI Assistant Summit this Thursday 25 and Friday 26 January. We are looking forward to hearing from leading minds in AI such as Ian Goodfellow from Google Brain, Daphne Koller from Calico, Yves Raimond & Justin Basillico from Netflix, Keith Adams from Slack and many more. There are only 30 tickets remaining, so register now for the last chance to join RE•WORK and learn about the most cutting edge advancements in AI and deep learning this week.