Women in AI Dinner, London: Highlights

While the nation was watching England in the FIFA World Cup, we hosted an evening of networking and presentations at the Women in AI Dinner, London. Joined by guests from industry and academia keen to support women working in AI for a three-course dinner accompanied by three fascinating presentations from leading minds in the field.

The evening, kindly sponsored by Prowler.io, kicked off with a champagne reception with attendees networking and discussing interests and backgrounds. Key discussions centralised around the importance of the rapid expansion of Artificial Intelligence in multiple sectors including manufacturing, finance, healthcare & security.

“Meeting skilled and like-minded professionals sharing their experiences about work is inspiring.” Manju Rangam, Data Scientist, Mclaren Applied Technologies

Once guests were seated, our compère Elena Sinel, began by introducing the speakers and topics of discussion. Elena is a Social Entrepreneur, and the Founder of Acorn Aspirations and Teens in AI. “Acorn Aspirations is educating the next generation of AI/VR/AR/Blockchain technologists, thinkers, and leaders globally”.

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. Darko Matovski, CEO, causaLens told us “I was impressed by the conversation. You normally get ‘fluffy’ conversation at dinner events but at the Women in AI dinner all of my conversations were very “meaty” and full of meaning - much like the main course!”

First up to present, Aleksandra Piktus, Software Engineer at Facebook, began by explaining why Facebook is working on understanding malformed text with word embeddings. Aleksandra asked the room if anyone had tried to learn a new language. “I think you all agree that we know the challenges of understanding a word we have never seen before. Attempts to capture the meaning of words in a computational sense is important” was the response after almost everyone raised their hands.

The inherent complexity of language makes training models to recognise vocabulary in human-generated text very difficult. Aleksandra emphasised the importance of preserving misspelled words in text to ensure models identify meaning in all words, no matter how they’ve been spelled. One particular example of this is its application in offensive language detection. Users will misspell words in order to deceive the system and share content that violates Facebook’s community rules. Aleksandra hopes to leverage word embeddings to decipher misspelled texts, she explains: “We assign vectors of a fixed dimension and use neural networks, specifically free layer neural nets. The model can then learn to predict context given a target word and also learn to predict the word given its context” It is the hidden layer of the neural network specifically that controls the dimensionality of the vector. Aleksandra suggests fastText as the solution; it relies on observation to leverage the morphological structure of words, including subwords and full words. For example, the model can learn ‘Great’ and ‘ness’ separately and eventually a meaning will be derived for the meaning of ‘greatness’.

So, what's the ultimate goal of their work in word embeddings? They hope that eventually misspellings will not be lost in human-generated text and that the semantic relationship between words and misspelled words will be preserved. They are doing this by utilising a multi-task learning algorithm which is trained on misspelled variants of the same word and the correct spelling in a supervised fashion. At the moment they are seeing encouraging results. There are also plans to internationalise models and have the ability to apply them to all languages.

Our second presentation began with a question: “How can we align the goal of the agent with the goal of the designer? And how can we ultimately get agents to make better decisions?” Sofia Ceppi from our partner, Prowler.io, introduced the topic of ‘Leading Self-Interested Agents Towards Desired Outcomes’. Sofia is a Senior Machine Learning Researcher and is part of the multi-agent systems team. Her work aims to combine the advantages of mechanism design with reinforcement learning techniques.

Classic mechanism design shows how agents that have private information can be led by appropriately designed incentives towards desired outcomes. Inspired by this, Prowler.io aim to tackle the problem of designing incentives to guide learning agents towards desired outcomes and away from naturally suboptimal ones. But can we induce them to change their behaviour and thus lead them toward the desired equilibrium?

It’s worth remembering that Sofia’s work is not intended to instruct or give orders to agents; “What we are trying to do is NOT tell you what to do... Instead, we are trying to design an incentive with your preferences in mind to warrant a change in your behaviour in order to end up with the preferred solution”. The challenge, however, is “How can we try to get agents to make better decisions? How can we align the goal of the agent with the goal of the designer?” In simple mechanism design, the main concern is that there is an assumption of the agents' behaviour and if this does indeed represent a ‘realistic environment’.

Sofia proposes a solution; “To achieve this, we consider stochastic games in which an agent – called central decision maker – is tasked with designing incentives for other agents. We propose a framework that models the central decision maker's problem as a Markov decision process where its action affects the other agents' environment, and its reward function depends on the decisions they take within it. This framework – together with mechanism design, reinforcement learning, and deep neural networks – can be used to tackle a broad set of problems in multi-agent systems with stochastic environments”.

Our third and final presentation ‘Drug Discovery Disrupted: Quantum Physics Meets Machine Learning’ was brought to us by Professor Noor Shaker, Co-Founder, and CEO of GTN. Before starting GTN, Noor was an Assistant Professor at Aalborg University in Copenhagen working on different aspects of machine learning with a special interest in generative models. At GTN, she is working with leading researchers on a novel, patent-pending, technology in drug discovery bringing ideas from quantum physics, machine learning, and biochemistry.

Question: What takes 15 years and $3bn to create? This isn’t the start of a bad joke, this is how long, and how much money it takes to bring a new drug to market. In real life terms, this shows the immeasurable loss of life undoubtedly caused by a challenging and ineffective process. The journey to drug discovery begins with the filtering of millions of molecules to identify the promising hundreds with high potential to become medicines. Around 99% of selected leads fail later in the process due to inaccurate prediction of behaviour and the limited pool from which they were sampled. Noor and the team at GTN are addressing the main bottlenecks in drug development with new innovations marrying ideas from machine learning, quantum physics, and biochemistry. They are pioneers in their approach, others tend to be limited in what they can feed their machine learning tools. “We are trying to be creative by breaking boundaries and going beyond what we simply know” Using their unique methods they are “able to sample new chemicals from outside the chemical space”.

Fundamentally, drug development is challenging, inefficient, long and expensive. Furthermore, return on an investment in drug development is currently expected to drop by 50% every 9 years. GTN’s work will provide the capability to terminate dead-end projects before further funding is wasted “you want to predict what chemical series you want to keep investing in”. The team is leading the way in overcoming this issue with their combination of quantum physics, ML, and Biochemistry. By using DL models in supervised and unsupervised settings with the help of generative adversarial networks, GTN hopes to close the gap between disease identification, and the successful launch of a new drug, and ultimately cure challenging diseases.

The evening drew to a close over coffee, with exchanges of business cards, discussions around the presentations of the evening, and plans being made to attend the Deep Learning and Deep Learning in Healthcare Summit in London this September. Early Bird tickets are available to purchase until Friday 20th July and the event is expected to sell out. Early registration for this event is highly recommended. Did you know we recently released episode 30 of our Women in AI Podcast? Listen and subscribe here to learn from the leading female minds in AI, Machine Learning, and Deep Learning.

Finally, hear what some of our guests thought about the evening:

“I enjoyed learning about imaginative uses of computer science/machine learning/ mathematical algorithms in very different real-life areas.” Maria Hernandez-Fuentes, Head of Translational Biology, UCB

“The dinner has been very varied which is very important to see.” Chantal Hernandez-Fuentes, Cross-Functional Project Manager, Orange Spain