Monzonaut AMA - Neal - Director of Machine Learning

Thanks @Ordog! You’re very kind. There are a ton of things that I’m really bad at too :slight_smile: – please don’t ask me about anything that is considered popular knowledge because I probably won’t know anything about it :joy:

To answer your questions:

a) Find time - I’m a strong believer in making the time instead of finding the time :smiley: and having a routine really helps.
b) Motivate myself - the strangest thing is that I’m often motivated after doing something, rather than before. Sometimes I don’t feel like working on a project so I’ll commit to just doing a very small bit of it. And as I do that, the motivation grows.
c) Come up with ideas - reading and talking to people really helps here. Often I’ve found that ideas that look “new” are actually two existing areas that have never been brought together.

7 Likes

If I were in a battle of the Dans I would immediately surrender as I’m clearly not the best Dan! I’ve worked in a squad that had two Dans (@danhughes was one of them!) and they are now both in the top-5 Dans I’ve ever worked with.

BUT if another Neal (and I mean a Neal with an “a” - not a Neil) shows up then we’ve got real trouble. I think I’ve only met one or two in my whole life? (Also TBT this community forum when someone said my name was fake)

5 Likes

What would anyone be able to do if they had toes for fingers? :face_with_raised_eyebrow: – it’s fingers for toes for me every time

2 Likes

Great questions @AlanDoe!

Looking back at your time with Monzo, what would you crown as your pride and joy in regards to a project you’ve worked on?

When I joined the Data Science team at Monzo, there were only 5 of us- so there wasn’t a machine learning team and everyone had to do a bit of everything. I’m really happy that a bunch of things I built back then are still useful to the new teams who have picked up those systems. The biggest technical project I was hands-on with recently was building a feature store. This is a system that now powers several of our machine learning models - so it’s great to see others build systems on top of the ones I’ve built!

How much has the full remote working impacted your workload during lockdown?

I’ve been grateful to not have to commute, but I think overall I’ve been working more - at least, it’s a bit harder to get away from the laptop in the evenings.

What’s your main downtime activity?

I am a purple belt in Brazilian Jiu Jitsu and absolutely love it :grinning_face_with_smiling_eyes: although it was near-impossible to train last year.

2 Likes

Thanks @SouthseaOne!

What is the biggest limit on what you can do with Machine Learning right now? Legislation? Time/effort? SQEP? Computing resource?

Our biggest limitation is people - we have more ideas that we need to say no to right now than ideas we can say yes to. That’s why we’re doing a lot of hiring!

2 Likes

What date does Skynet take control? :robot:

Fearing a rise of killer robots is like worrying about overpopulation on Mars :wink:

1 Like
  • The most accurate depiction of AI

If you haven’t seen them, check out Coded Bias and AlphaGo.

  • The most funny depiction of AI

I didn’t watch the film but Hitchhiker’s Guide to the Galaxy remains the only book that I’ve burst out laughing while reading (I was on the tube and got weird looks)

3 Likes

Great questions @BritishLibrary!

or things like deep fakes (which I’m not sure if that’s actually a machine learning thing).

Deep fakes use a recent(ish) development in machine learning called Generative adversarial networks, or GANs

What are your/Monzo’s goals when it comes to rolling out machine learning + AI in those areas, and what will that mean for the consumer, in terms of ‘adding more value’.

This is the type of question that I spend most of my time thinking about. Here’s one example from work last year: by using a combination of rules and machine learning, we built a system that could automatically route a chat to the right kind of specialist. How does that add more value? Customers, on average, would need to chat with fewer agents and so they would get to an answer faster. For Monzo, it’s more scalable to have one person work on a chat rather than two (so we can help more customers). So it’s a win-win all around!

3 Likes
  1. What approach do you take to finding the right balance of machine learning algorithms and the human element? Do you see machine learning as a means for replacing some human elements, or as something that can augment them?

For areas like customer service, machine learning does not replace a human touch – it helps to get you there faster. For other areas, like detecting and preventing financial crime, automated systems will work non-stop to protect you, and so will do a better job than if we had humans!

  1. What are your thoughts on how other banks approach Machine learning? What have others in the industry done that you think monzo has done, or could do, better thanks to machine learning?

I haven’t been very good at keeping abreast of what other banks are doing with machine learning- sorry! One bit that I do know is different is that our Machine Learning Scientists are empowered to work on machine learning systems end-to-end. In other companies, there is often one person/team who trains the model and another person/team that ships the model into production. We don’t do that: we upskill our ML Scientists to work on our production code safely. From this we’ve seen a huge speed up in the time it takes to test an idea.

  1. Is an AI chatbot something Monzo have explored, or planning to do, to serve as as a concierge between FAQs and human based support? How easy do you think these systems should be for users to bypass the AI and speak to a human? A shortcut (/human) like N26, or click buttons and hope for the best like RBS?

We experimented with an automated responses for some queries in 2019 – back then, the bot was called Monzo Helper. While it showed incredible promise, there was a decision to not pursue this direction any further in 2020. One of the motivating reasons for that is that there are many problems that customers should be able to resolve by simply tapping a button, rather than needing to chat to a bot. And so the teams refocused on improving our app on that front.

7 Likes

documentaries like Terminator

:joy_cat:

2 Likes

Thanks @AlanDoe! I’ve jumped in a little bit early and have answered some of the questions we’ve had so far. The themes so far have been

  • How we use machine learning at Monzo
  • Ethical machine learning
  • Getting motivated to work on side projects
  • Movies about AI

Looking forward to any more questions that folks have!

8 Likes

You can watch the latter in full on YouTube:

We watched it during the first lockdown and it is fascinating. It made one really leftfield stunner, but also an odd clunker as well

I would give the film a pass. Underwhelming

The TV series is ace though!

2 Likes

Do you like having a common name with an unusual spelling?

“It’s Neal. But it’s spelt enn-eee-ayy-ell”

1 Like
  1. How much Machine Learning mechanism is involved in seeking/predicting/finding fraud :monzo: accounts? :octopus:

  2. How much Machine Learning is important in taking final decision on customers :monzo: account activity close the account or it was just a system error/bug? :shark:

Yes! It’s made me a lot more aware about not making assumptions about people’s names - I know how it feels when I need to ask people (sometimes, several times) to spell my name correctly

3 Likes

We have our best and brightest working on this space! :eyes:

We don’t use any machine learning in this space at all - this is a great example of an area where defaulting to an expert human’s review is the better thing to do.

9 Likes

What’s the biggest AI/machine learning mistake that’s been made in your time at Monzo (or career more generally?)

(And by that I mean, when an AI has just made a mess in a “oh we didn’t expect it to work like this way”, rather than a “it was a mistake to do AI on this task” kind of way, if that makes sense?)

2 Likes

A few come to mind! The first is not really an AI fail, but I once broke our help article system and it would just… quietly not return any results. One of our QA testers picked this up quite quickly and my heart sank as I quickly rolled back the change I had made :scream:

Maybe more of what you are looking for: about a year or so ago, we ran an experiment on an old system where we used machine learning to try and find a better order for the chats that are our agents work on. The idea was that answering similar questions in quick succession (e.g., a bunch of customers who all want a bank statement) would be faster than getting agents to jump from one type of customer query to another and needing to context switch every single time.

The experiment showed some very promising results, but one thing that we had forgotten to factor in is that our expert agent look for patterns in the customer queries they are getting and report unusual trends to their managers - after all, it may mean that something has inadvertently broken! So we had panicked COps reporting in that there were unusually large volumes of requests for bank statements, and others reporting in unusually large volumes of lost cards – when in fact all that was happening was that our system was splitting these chats up and sending them to agents who had recently worked on similar questions. Oops!

We moved quickly to fix this - but in the end, we moved a different direction with this overall system.

14 Likes

Tonight and tomorrow will be your last chance to get questions in for @Neal who has been amazing and been jumping in all week to pick them up. Absolute hero!

6 Likes

Has he though? This could be his latest experiment! :thinking::thinking:

4 Likes