Earlier this week at the Women in AI Reception in New York, RE•WORK brought together leading experts working in AI for an evening of networking and keynote presentations over a drinks reception and canapés.
When our guests arrived at the reception, they were welcomed with fruitful conversations about the world of artificial intelligence and its impact in both research and industry. Throughout the evening, key discussions included the importance of bringing diversity into the field of AI and technology more generally, as well as potential future insights into the development of AI.
After plenty of networking and some really exciting conversations, our guests were seated to listen to our 3 keynote speakers. Teal Willingham, General Manager at NYU Future Labs, our compere for the evening, began our evening on her take on diversity in the field.
“It’s not true that we’re missing women in AI - they’re definitely here but we just need to get together more at events like these to support each other and share achievements”
Our first speaker was Catie Edwards, Machine Learning Engineer at Spotify. Catie works personalising content on Home and she also co-founded Code Squad, a non-profit organisation dedicated to bringing computer science education to underrepresented middle schools across Washington, DC.
Catie began by discussing Making Music Magic: Personalizing Explainable Recommendations on Home.
“What is magic? Magic at Spotify is making a recommendation system that makes it feel like we can read user’s minds”
Spotify has been looking into “How to infer information to the user?” This is a problem they have been looking into fixing as Spotify needs to provide explanations for their recommendations as “Transparency gives confidence to the user”. Catie discussed the issue of how to get users to trust Spotify with recommendations and to try something new. Spotify recognised that popularity is actually the weakest driver for users, whereas recent listening and genre is highest for user engagement, showing that the relationship with the user is very personal.
Catie explained that Spotify is employing the bandit approach when introducing exploration into the user’s experience. By doing this, Spotify can successfully generate feedback to new options presented to the user, therefore gathering more data and then making the experience more personalised to the individual.
Next was Julia Kroll, Data Engineer at Amazon Alexa. Julia is on the Applied Modelling and Data Science team at Amazon Alexa in Boston. She develops novel solutions related to sourcing and distributing multilingual language data applied to internationalising Alexa.
Julia began by sharing the process of internationalising Alexa and the natural language processing (NLP) features that power Alexa, which aid Alexa to be able to understand country dialects.
“Machine translation doesn’t fully cover text to speech due to things like dialect. It can misunderstand semantic context.”
She gave us insight into the special skill that Amazon developed for Alexa, which is Cleo. Cleo allows Alexa users to teach Alexa local languages, and through this, users can teach Alexa not only their own language but also culture. Cleo uses the crowdsourcing method for collecting language data. Julia explained that catalogue data is very important to Alexa as this adds more personalisation and localism to the user experience. Catalogue data weighted on localised offerings is very useful, as for example, Alexa can then differentiate between the U.S Office (TV Show) as opposed to the U.K Office.
To end the evening, Lucy X Wang, Senior Data Scientist at BuzzFeed shared with us the process of matchmaking audiences to the BuzzFeed content available. Lucy is working on machine learning tools for optimising audience reach and engagement.
Lucy explained that one of the obstacles they are facing at BuzzFeed is curation scaleability.. How do they scale their efforts to seed content in social media to reach new audiences? This is especially difficult for BuzzFeed as they are working across so many social media channels.
“Human created data is a treasure trove of opportunity. We can then try to build a model that mimics human behaviour”
BuzzFeed trains models to find out which social channels does a particular piece of content have the best match with. They are working to solve the problem of what social channel they should direct content to. Lucy spoke about a common problem in machine learning: “Underlying data is not a source of truth. It is very messy”. She spoke about identifying the need to mimic human creation behaviour, therefore they had created models off of past human-generated data. Lucy explained that a lot of quality assessment is required for results. Therefore they created a slack bot to ask users if it was a good recommendation and the answer is fed back into the model to retrain.
The evening drew to a close with our guests and speakers exchanging business cards and discussions around the presentations of the evening. Thank you to Catie, Julia, Lucy and Teal for the great evening and we look forward to seeing some of our guests again at the AI in Finance Summit, New York on the 6 & 7 of September!