深度学习金融峰会：看看Monzo, Barclays, Santander都有哪些亮点

“Go AI or die. I’ve died more than I’ve innovated so I’m speaking from experience. This is the only way to grow and move forward.” - Jackson Hull, GoCompare

This week in London, RE•WORK were joined by global experts in finance and deep learning to explore advances from the worlds leading innovators across industry, research and the financial sector. Speakers included experts from Monzo, Visa, Lloyds Banking Group, Santander UK, UBS and many more. Throughout the two days, we heard about fraud prevention, risk management, investment predictions, customer service amongst other emerging topics in deep learning and finance.

On Tuesday morning, attendees arrived in St. Paul’s, with the global reach of the summit immediately evident when we spoke to two of the first arrivals, one with a travel time of 5 minutes, to an attendee who had made a 15-hour journey, all the way from Singapore. The day was off to a busy start with coffee and breakfast kicking off the networking element of the summit -  something we really encourage by offering 12 hours of networking across the two days. 

Luigi Troiano, Professor of AI, Data Science and Machine Learning at the University of Sannio kicked off the morning’s sessions by welcoming the expert attendees and speakers alike, encouraging everyone to get involved in the conversation on the RE•WORK Twitter accounts #reworkFIN and @reworkfinance. Luigi introduced the first session, ‘The Changing Financial Landscape’, which featured speakers from GoCompare Group, Direct Line Group and UBS who honed in on the ways in which AI is transforming the financial space in areas such as customer service, insurance, and more. With the rise of chatbots and other online customer-facing products, both the demand and expectations of flawless systems from customers is growing quickly. This is driving companies to place an increased emphasis on making these systems as user-friendly as possible. Jackson Hull, COO and CTO at GoCompare group explained how they are using DL on customer, industry, and transactional data sets to provide customers with more value from financial services:

Our events welcome both new and returning guests, and we were delighted to hear from Huma Lodhi from Direct Line Group once again. Huma joined us on the Women in AI Podcast last year in London, and you can listen to the episode here, where she discusses deep learning from an insurance perspective. This year, Huma shared the developments in her work and she demonstrated the successful applications of DL for different tasks ranging from risk modelling to claims settlement. As the applications of Artificial Intelligence in insurance are increasing, we will be hosting the AI in Insurance Summit, alongside the next edition of the AI in Finance Summit in New York this coming September.

Trust and responsibility was a recurring theme, and Manuel Proissl, head of predictive analytics in Banking Products at UBS continued this conversation by stressing the importance of building trustworthy consumer-centric AI products in banking. He explained how the reliability and interpretability of AI assistants is paramount to build sustainable value for consumers. We need an ecosystem to build trust - we probe, train, test, deploy and measure. As you go deeper and deeper in this research you’ll find there is no bulletproof selection, there are trade-offs to be made.

With explainability and interpretability at the forefront of everyone’s minds, the morning continued with presentations from University of Oxford and Swedbank. The conversation turned to why we should be applying DL in finance and honed in on fraud prevention amongst other benefits. Katia Babbar from University of Oxford drew on the fact that many other talks in the morning’s session were about prediction ‘and we can use deep neural networks as very very powerful approximators. The reason why it works is that we’re saying we can approximate any function that is smooth enough.’ Although finance is an industry with an abundance of data, lots of it is sparse or unlabelled which presents challenges. Mehrdad Mamaghani from Swedbank said that although DL is hugely helping in many areas including fraud detection,  there is a ‘lack of access to recent cases because they’re still unfolding, so we don’t know exactly what we’re looking for. Once we know something deviates we know it could be fraudulent.’

One of the key focuses was for attendees and speakers to expand their network, and we hosted networking drinks, roundtable discussions with DL experts, panel discussions and more. We were able to chat to some of our guests to hear what they’d enjoyed most about the event:

Krista Caldwell, Mindle.ai Account Manager

“It’s been really great to understand what the hot topics are in the space, and this summit summarises them really well.”

William Rouse, Director, Costello & Reyes

“The breadth of experience is great, it’s not just limited to FinTech. Really excellent mix of technology and application!”

Merve Alanyali, LV=

“I’m really enjoying it! Seen some familiar faces as well as having some great new chats. I especially enjoyed the second half of this morning when Katia from UoX recommended some further reading - it was quite technical but I learned something!”

One of the afternoon’s panel discussions focused on the considerations for utilising AI to assist security efforts and enhance safety in finance, and panellist and a regular guest at our summits, Adam McMurchie from Barclays brought us back to the topic of the challenges with sparse data: from our point of view, one of the key challenges with applying DL for security starts at data collection, and the challenge here is labelling. You need some data massaging before you create the models. If you have an idea of what you want the models to look like and what the data will look like, you know the inputs or fields, and this is labelled by a human. The panellists explained that once you have the data, DL can be incredibly helpful in increasing safety, and Google Cloud’s Customer Engineer, Mason Edwardssaid that once people have the data labelled they need to be careful as he’s seen people go from not using their data at all to suddenly wanting to use a deep learning model. He explained that more simple AI methods should be used initially otherwise you’ll face legal issues amongst other challenges, but first things first, get your data in order and scope it down to the smallest viable project that will make a difference, and build on that. Adri Purkayastha from BNP Paribas Group explained that “raditionally fraud detection is an imbalanced qualification. You need something to clarify that the anomalies you find are actually anomalies.

Another recurring theme across the two days was bias in financial models. Some of the key challenges that were brought up focused on eliminating this before it’s built into the model. Rich Radley from Google Cloud shared their technique to tackle this, which has been open sourced and Rich said that they’ve written a technique that is calculating maximum mean discrepancy where we can find prototypes and criticisms in the data to increase the diversity in our data. What happens once you’ve downloaded a model where you don’t have access to the original training data? How do you inspect it for hidden bias? One of the techniques (in a paper we’ve published) is something called testing with concept activation vectors (TCAV). It learns to quantify the sensitivity of a model’s perceptions to an underlying high-level concept which is explainable to humans. E.g. I want to test whether an image is a doctor, and the concept I want to test for is whether gender has any bearing on which images are showing you.

Day two kicked off with another networking session and breakfast where we heard what everyone was looking forward to. Encouraging emerging companies to implement DL and share their work is something that we support in several ways, and day 2 highlighted some of the most promising startups working in DL and finance. Before introducing the  startups, the morning’s compére, Marcin Kacperczyk from Imperial College London told us that ‘Imperial is one of the research institutes in the world at the forefront of AI research, with more than 400 people involved in research in this space.’

Derek Burke, Senior Manager at WekaIO spoke about A Modern Data Lake Accelerating Deep Learning Workloads. WekaIO is a relatively new company but ‘won business intelligence AI intelligence Excellence award’ and is working to create ‘the world’s fastest, most scalable parallel file system for modern analytics and technical compute workloads on premises and in the public cloud. We provide the world’s fastest file system and speed up deep learning systems. We can also be fully deployed in the cloud.’

Nima Shahbazi, CEO at Mindle.ai then spoke about how ‘there is growing demand for ML stock market predictions, but to be valuable, algorithms must be highly accurate and adaptable.’ He explained how there is a risk of overfitting and he’s working to reduce this to get a better machine learning signal. This risk is one of the key bottlenecks in this area - ‘you might be doing really well on your backtest data, but when you put the model into production everything is changing - why? This is the talk of risk I’m talking about.’ Neural Networks are everywhere, but Nima’s advice is to not go deep! Deep models are very effective in highly structured featured models, ‘but financial data is highly noisy and there is a big difference between being super non-linear and super noisy.‘

With startups and new companies in mind, we spent the coffee break learning from leading VCs who shared their advice on securing funding for early and mid-stage AI companies, working specifically in the financial sector. Ash Puri from eNTIER CAPITAL explained that some of the most frequently made mistakes come with startups trying to sell to big corporates: ‘sometimes it’s better to have lots of small companies and you can scale from there. Also raising the right amount of capital is very important. People think it’s okay to raise 6-8 months of revenue but I advise raising 18 months because these things take time. Investment partners is like a marriage, you need to be willing to be flexible and adapt to each others needs.’ Back in the presentation room,  Angel Serrano from Santander continued the conversation of growing your business and seeking investment, and he advised that companies should ‘think big but start small. Request small investment, show value and then request more.’

Recently, you might have seen luminous orange-y pink debit cards popping up. This is Monzo. Neal Lathia, their Machine Learning Lead actually called their cards ‘hot coral’ and shared how Monzo is an app only bank, you get your card and the app which comes with a whole range of things from easy payments to money management. Neal linked back to several of yesterdays presentations by talking about customer operations. Focusing on NLP, Neal spoke about how they’re using embeddings at Monzo to help recommend relevant articles that can help with queries before you get passed to a human. He summarised something that has been recurring - AI is developing incredibly quicky: ‘We’re not trying to stay one step ahead, but stay 10 steps behind the latest research - it’s just not possible to keep up to date with everything that’s getting updated every day. With NLP we were heavily reliant on a great paper from 2017, but then in 2018 there were huge progressions. All the results looked super exciting, so we needed to stay on the ball.’ Looking at these new insights and technologies,   Javier Campos from Experian brought back the topic of bias and fairness and reminded us that although it’s an emerging topic in AI, it’s something that has been discussed for many years. Fairness is a complex issue and we have to think about the full AI framework: fairness, accountability, customer driven, transparency, ethics, safety and security. The fairness problem happens everywhere in machine learning, but even more so in deep learning, and Neal also mentioned that they are dealing with bias issues in Monzo with their language understanding models.

On the ethics panel, Daniel Drummer from J.P Morgan explained that in ethics, you have to do the opposite of Neal’s suggestion of staying 10 steps behind - in ethics, it’s important to always be thinking forward to ensure models are built with ethically sound data sets to eliminate bias. Javier, who was also on the panel, explained that ‘fairness isn’t something that can be created in a model and applied to everything - it’s contextual and needs continually changing frameworks.’

Wrapping up the presentations for the day were discussions on NLP, as well as solving problems in industry. Here are some of the highlights:

Eric Charton, Senior AI Sirector, National Bank of Canada

‘We’re seeing if we can improve some applications with AI models - we’re working on dialog systems, and we want to build it as a whole bank product, for the customers but also for our clients.’

‘The idea is to create an application that can generate a full operational bot with an easily generated model in 5-7 minutes.’ 

Dapeng Wang, Data Scientist, LV=

We want to break the problem down into smaller, managable tasks, but that’s very hard to do, and as a small team it’s not always the best strategy. So as an example, we have some images and we’re using a CNN to predict if something is damaged - so these are yes or no answers. We have the data (the image) we process the image, we train the model and then interpretability is reached. The most difficult part is data because getting the data defines the model you’ll build. 

Emiliano Martinez Sanchez, Senior Engineer, BBVA 

At the innovation labs we try to get the latest innovations into production in our three main disciplines which include security. Big data infrastructure must be reliable without interruption with a high speed network and of course security and data governance.

Brexit is a topic much discussed, so of course, it was only right to include it in the discussion. The closing panel looked at The Future of AI & FinTech in the U.K & Beyond post-Brexit featuring experts form UK Department for Business, Energy and Industrial Strategy, UK Department for International Trade,  Liberty Specialty Markets and more.

Jackek Wieclawski, Robobank/World Economic Forum: Brexit is a massive problem for us, we all suffer. Brexit isn’t slowing the process, and money is still there. The only problem is talent - we’re struggling to attract people coming from the European union and we need to attract this. The USA isn’t good at this either with Trump, but Europe may end up winning this game. The global fight for talent will only intensify and it might be quite brutal. It’s hard to find really talented people - banking isn’t as sexy as it used to be, and it’s not a wonder why people are picking big tech companies, but in finance it’s amazing because we sit on live data.

This panel will be covered in a stand-alone blog post later next week, so watch this space to hear what was discussed.

With the applications of AI in finance ever increasing, we’re excited to bring the AI in Finance Summit to New York this September 5 - 6 where we will be joined by experts from the likes of Prudential, Barclays, MIT-IBM Watson AI Lab and many more. Tickets are currently available at a discounted rate, so register here to learn more.