So your fintech product uses AI and ML…..But does it really?
A few weeks back I appeared on a panel of “dragons” at a fintech event where a number of firms pitched their products to us and the audience, and we in turn pitched back a series of questions and comments to dig into the detail of their product and its real or potential use cases. What struck me is the near ubiquitous claim, not just at this event but throughout fintech, that firms’ products incorporate some form of “AI+ML”. What the vendors are suggesting is that their product utilises both techniques encapsulating notions of artificial intelligence *and* of machine learning (*and* sometimes ‘DL’, Deep Learning). Given that it was a collegiate forum, and I hadn’t been warned that I would be on a judging panel until ten minutes before the event (so hadn’t done any homework on the vendors), I let it go. But it seems to be a theme right now across banking and fintech that needs to be challenged. What I really wanted to say was:
“Really?! Both? Prove it!”
But that’s a little incredulity-heavy, so I’ll propose a more gradual approach that can be employed to assess the density of the smoke, and the degree of distortion of the mirrors being deployed by vendors, so anyone can assess the salient elements of the product and how it applies to their business.
In the interest of full disclosure I’m not an expert on AI. While my PhD subject (and subsequent post-doc research) area was parallel automated reasoning it was the Wrong Kind Of Automated Reasoning. For brevity (and because it’s a sore point) I won’t go into that further here. Nonetheless I feel I have enough of a handle on the subject area to know when I find a proposal and a presenter cogent of the product area under discussion, or economical with the truth, or over-zealously describing their achievements.
To assess an “AI+ML” product we are trying to establish four fundamentals:
A. Does the vendor’s system really use AI, ML, both or neither?
- This is the baseline case – are we simply being conned – do they even know anything about the subject area?
B. Does the vendor understand the AI and/or ML that they’ve incorporated into their system?
- If they are not able to clearly articulate the reasoning process and the benefits of utilising it then my gut feel is that they don’t understand it themselves. Often the best way to determine whether you understand something is to explain it to someone else.
C. Does the product get any additional benefit from the AI techniques being employed?
- Just because a system doesn’t use AI doesn’t mean it’s intrinsically bad – I don’t need AI to knock nails into wood. Sometimes less is more, and additional complexity will inflate the price but not the functionality. It may actually make a product worse insofar as it is harder to support.
D. Are you (the prospective purchaser) able to explain why the system in general and AI/ML in particular are beneficial and legitimate for your use case?
- AI and ML are currently super-hot topics, but that doesn’t mean it will either make or save you any money.
So let’s consider a set of questions that could be employed to enable us to assert that we are comfortably able to answer these four fundamentals. The following questions are tailored towards firms that claim to use ML, but should give you a perspective to help tailor your own interrogation. Ultimately the following 8 questions are intended to enable you to assert A, B, C and D above.
Imagine you come across a firm, and in the interests of tradition, let’s call them “Mystery Incorporated” (with apologies to Hanna Barbera). They approach you and pitch their system as being AI+ML-enabled. The first question to ask is the AI101 challenge, and a precursor to being able to start any real interrogation; so adopt an innocent expression and ask:
- “What is the difference between AI and ML?”
[fundamental ‘A’]
Now Mystery Incorporated are a small operation, and the Account Execs assigned to deal with your firm are called Fred and Daphne. They’re smooth but hopefully smart enough to have paid attention when Velma, the Head of Engineering and Pre-Sales Support explained the difference to them. If not, this will result either in a great deal of bluster, or a pregnant pause accompanied by some shuffling of feet and a promise to find out Real Soon Now. However, in the event you get a reasoned response along the lines of traditional AI meaning that a system can exhibit indicators of intelligence but perhaps in a static, pre-programmed sense, whereas ML is a subset of AI that enables a system to reason (potentially) on its own and grow its understanding, then its worth continuing the discussion. If they use the terms “intuition” in the context of ML or “rule-based” for AI then bonus points, order another coffee. You could throw in “deep learning” (DL) as an additional variable and you should get the response that DL is a subset of ML, and involves developing multilayered NNs (neural nets with ‘hidden’ layers of nodes between the input and output layer) to reason about large quantities of data. If they can’t answer this question then they’re unlikely to be able to answer anything further so exit(1); /* report EXIT_FAILURE */.
The next question digs deeper:
- “Which elements of your systems utilise AI, and which use ML techniques?”
[fundamental ‘A’]
This will test not just their understanding of the subject area but also their own product. Typically Fred, Daphne and Velma arrive to do the pre-sales. Fred and Daphne are there to schmooze you, tell you what a great firm you have, what a great firm they have, and pay for the coffee. Don’t ask them this question, generally they won’t know the answer or at least won’t have an adequate depth of understanding to enable you to determine how to run a cogent POC or select an appropriate use case further down the line. The question is moving into Velma’s territory. She knows the product, the tech, plus she’s the SME so will understand your business. If she gives a coherent answer then you can continue, safe in the knowledge that fundamental question A has been answered. This is where we discover whether the product has static abilities to reason or if it can be trained further, with potential other use case opportunities.
There are a number of different algorithms used for AI and ML including decision tree methods (AI) or neural nets (NN for ML). NNs also have many forms including Now we want to really understand whether the product team understand their product so ask:
- “What kind of AI or ML model(s) does your system use?”
[fundamental ‘B’]
Bear in mind you need to get an understanding of how their product’s reasoning process works, and if they can’t tell you the pitch ends there. In order for you to rely on the decisions a product makes you will need to be able to explain the causality of the decision-making process. Imagine your next visit from the FCA when their representative asks how you managed to run up such a huge position. “The algorithm told me to do it” is explicitly not an acceptable defence. They will be looking for you to demonstrate that you have the core competencies in your team to support and understand the products you employ. You cannot rely on a black-box trading algorithm – you would generally backtest it to understand its behaviour. If a human research analyst told you to buy TWX.N they would generate a human-readable report to justify their decision. By the same rule if Mystery Inc’s system tells you to buy TWX.N based on the data it has consumed you would want an understanding as to how it came to the decision. In the days when I built Expert Systems for the gas processing industry I could render a linear path through the decision tree that had been traversed in order to reach the result. Although the weightings were essentially integer tables the “intelligence” was largely human-readable. Note that decision trees are AI not ML. They are essentially static even though I had the opportunity to tweak weightings as we went along. ML and especially DL are rather more difficult to explain, the whole idea of having hidden layers in a NN is that they are, well, hidden.
So since AI is now considered rather “Old School” I’m going to assume that Mystery Inc.’s system uses ML. Or they claim it does. If a vendor’s application is using NNs then they may be able to give an overview of its mechanics. The should be able to tell you why it’s a Feedforward, Radial basis, Kohonen, Recurrent, Convolutional or Modular NN for example – the Wikipedia page “Types of artificial neural networks” gives you enough detail to get a basic understanding, at least sufficient to see through the most obvious bluffs. Suffice to say they should be able to explain it to you. If it doesn’t make sense then there’s a good chance Velma doesn’t really understand it either.
You could ask:
- “How do I/did you train the NN?”
[fundamental ‘B’]
Hopefully they’ll start talking about training sets and one of the following training methods, these are four of the most commonly referenced:
Supervised: In this method inputs are labelled so that the algorithm has a known reference set of results. The algorithm is trained by a comparison of its calculated results to the pre-defined values and weightings adjusted until it gives a good prediction based on the target outputs. Using various methodologies such as regression or classification the algorithm is able to predict new outputs with accuracy based on its database of labelled data. This works best when historical data is strongly correlated to future data.
Unsupervised: Unsupervised training has no database of labelled results. The algorithm has to find patterns and structures within the data without target outputs to refer to. An example application might be classification whereby an algorithm could group objects according to various attributes such as colour or shape and therefore categorise them.
Semi- Supervised: This is an enhancement to Supervised learning where both a labelled and an unlabelled data set is used. Given that labelling data is hard work this algorithm primes the learning process with a small labelled dataset to speed up (for example) the classification process.
Reinforcement learning: This method is inspired by behavioural psychology, whereby the method aims to maximise its cumulative reward. Using trial and error it works out the how to achieve the highest cumulative reward as quickly as possible. So you don’t need a training set. Imagine you want to drive from London to Glasgow. Given a set of routes between cities around the UK it’s quickest to go by the shortest route, say London, Birmingham, Manchester, Glasgow, not London, Bristol, Birmingham, Manchester, Glasgow. The cumulative weightings (e.g. 1/mileage) on each intermediate route would score the shortest route highest.
At this point Velma is making solid progress to convincing you that she knows about AI+ML, but what about some of the snafus we might encounter? Where can the model go bad?
You can ask:
- “How do I avoid overfitting and underfitting?”
[fundamental ‘B’]
Overfitting is where your model fits the training dataset really well but not so well when you use a test set so it doesn’t generalise well. They could propose adding more data or more diverse data to the training dataset. An alternative is to augment the training set data with variants, so if you have an ANN that recognises cars you need to show them from all directions, not just from the front. They might also suggest reducing complexity to increase generalisation, so you could reduce the number of neurons in the layers or reduce some layers altogether. Another option might be to randomly “drop” or ignore nodes from the hidden layers (cunningly called “dropout”) for similar reasons.
Underfitting occurs when a model can’t classify data in the training set well (epic fail!). The problem can be ameliorated by increasing the complexity of the model – adding layers to the model, or adding nodes to the layers. Another option might be to increase the richness of the data provided. Finally, if we’ve implemented dropout we could reduce the number of nodes that get dropped (as discussed above).
If we’re still going well then it may be time to go for the jugular:
- “What benefit does <insert AI/ML model name here> give you and what metrics did you use to show it was superior to non-AI/ML methods or other AI/ML methods?
[fundamental ‘C’]
This is the point where we learn whether the entire development has been window dressing an otherwise conventional model. We have to be convinced that the development actually gives the product an edge, otherwise why bother?
Finally, we want to make sure that the vendors is honest, trustworthy and won’t knowingly sell us their product for an inappropriate use case. If you can describe a use case that takes client data as input including names and addresses then ask:
- “What impact does GDPR have on your modelling? Do you have to make any concessions with the way you handle the data?”
[fundamental ‘D’]
Any vendor that deals with data will know the implications here. Interestingly I recently asked this question of a presenter (who shall remain nameless to protect the guilty, so let’s call him “Shaggy”) in a reasonably crowded public seminar. Shaggy had started his talk by telling us he had received a prize the previous evening and spent a great deal of his time working with, and for banks. He then proceeded to show us some data mining that involved personal data including names, etc. so I figured that question 7 was a fair one. His response to it was fascinating. He began blustering about how banks didn’t care about GDPR, in fact they didn’t care about legislation at all, and that bank X and bank Y (not the names he used)
“are money laundering billions through the Cayman Islands!!!”
I had hit a nerve. Clearly Shaggy either didn’t know anything about GDPR or didn’t want to admit that the use case was wholly inappropriate and, being unable to admit his lack of omniscience resorted to theatrical tactics. I responded with another question:
“Seriously?! I can assure you that banks are extremely concerned about legislation and I hope there are no representatives from bank X or bank Y present today!”
Shaggy carried on blustering until I said perhaps we should leave it till later. The audience were amused and I chatted to a number of them afterwards who certainly did know about GDPR and confirmed my understanding that neither bank X nor bank Y, both being high street names, had ever been accused of any misdeeds associated with the Cayman Islands. We beat a hasty retreat from the venue. It is important that you and the vendor’s pre-sales team understand your use case and the ramifications of any new technology. Sometimes it could get you into trouble.
The final question is not for Mystery Incorporated but for you, the prospective purchaser and provider of service to your business:
- “Can you prepare a presentation for your boss/Board/Exec that explains why your firm should adopt the product, confidently answering each of the previous questions when they ask you them?”
[fundamental ‘D’]
If you get through all 8 questions and ultimately decide to transact, super-smooth Fred and Daphne will seal the deal and promise you the world, but make sure you contract Velma in to oversee the implementation. Otherwise the post-sales team of Shaggy and Scooby Doo will turn up and wreak unabated havoc on your environment. Worse still, if Mystery Inc. purchased their tech by buying another firm then firing the expensive resources (i.e. Velma), retaining only the maintenance team (Scooby and Shaggy) under the direction of their own Sales Execs (like Fred and Daphne but in shiny suits with no domain or product knowledge) then draw the meeting to a close at the earliest opportunity. Remember - "they could've gotten away with it if it wasn't for those pesky kids!"
You have been warned.
Postscript
As an extension to the fundamental questions I said we need to ask ourselves before purchasing AI/ML/DL products there are some ancillary or correlated ones we need to be comfortable that we understand before progressing with a major undertaking:
- Could I do this better with existing (simpler) tech?
- Is AI/ML worth the premium to my business?
- Does the vendor have an adequate support model post-sales?
- Do I have sufficient inhouse skills to use the product effectively? If not, can I acquire them?
- Does the vendor provide training?
- Will I need to engage the vendor each time I want to update the system or address a new use case?
- How much extra development or config will I need to do to make the product reusable after a POC?
- Can I self-support this product and how much ongoing support will it require?
- If I don’t get the answer I was expecting can I get an understanding of the reasoning process that was undertaken?
Embedding AI into a product doesn’t make it truly autonomous, self-reliant or self-supporting; at least until we get a bit closer to the singularity. You need to be able to explain a system’s reasoning processes, know its strengths and weaknesses, and be able to update it as requirements and environments (datasets) change, or at least know if it will still be successful under changed circumstances.
Engineered AI is a nascent subject area, especially where fintech is concerned. It is changing rapidly so any code you write today may be out of fashion next month. Consequently you need to convince yourself that you’re not just playing with the tech, there’s a genuine benefit to your business, and the product is both cost-effective and sustainable.
Vendors doing potentially cool stuff
I’ve seen presentations from a fair slice of the market covering the areas of AI and data science recently. I should note that this collection of firms in no way resemble Mystery Incorporated and may be worth a look. You might like to try answering the fundamentals for yourself when discussing their product set.
Thoughtonomy (www.thoughtonomy.com) are masters of Robotic Process Automation (RPA) on Azure. RPA is not something I’ve discussed previously, so well worth mentioning here. Given the volume of tasks that are automatable there is considerable opportunity to make your business more efficient using RPA and they have plenty of experience in real-world scenarios. Thoughtonomy claim to utilise “cognitive AI”. My understanding of this is that they use NLP to answer real-world questions and that this is all wrapped around the Blue Prism product (which you can't buy direct so have to speak to a partner such as Thoughtonomy). Anything is welcome if it improves my interaction with those irritating automated response systems that leave you waiting for “Press ‘9’ to speak to an operator” because they are essentially just lazy routing tools. One example use case could be to “create a new user named Finbar Saunders”. If this request is sent to a platform-specific mailbox then much of it could be autocompleted, or at least partially so. In the interests of completeness Blue Prism also partner with a range of other consultancies including the goliaths like EY etc so you have some choices.
fraXses (www.fraxses.com) have built a data federation framework that enables you to do data ingestion without transformation i.e. while leaving it in situ. It creates metadata and has an interface to every common repository of data at rest on the market (that I can think of offhand). Interestingly one of their use case articles details a range of public data sources such as https://data.gov.uk/, http://www.census.gov/data.html, http://aws.amazon.com/datasets/, etc which could prove invaluable although beware of trying to download the 300Tb archive from CERN’s repository.
TAMR (www.tamr.com) are concerned by the state of financial institution data and observe that there is a large data debt in most firms. Organisations want structured data but actually their data is distributed and in many disparate structures so analysts spend 60-80% of time on data wrangling rather than analysis. TAMR reckon they are using supervised machine learning to look at the profile of data. They ask 500 questions to create a profile, then push a million records through it, so data unification and classification is their angle, but you’ll need a different tool to render the BI.
iManage (www.imanage.com) automatically read, parse, extract key data, label it and categorise the data. They claim to use AI too. It’s document data only, so no video or phone calls and at present it is a service rather than a product you can play with. They’ve not ruled out opening it up to users in the long run though.
Ayasdi (www.ayasdi.com) are focused on Compliance. From their own website “Once the data set is understood…supervised approaches are applied to predict what will happen in the future. These types of problems include classification, regression and ranking. For this…Ayasdi uses a standard set of supervised machine learning algorithms including random forests, gradient boosting, and linear/sparse learners. The discovery capabilities of our technology are highly useful in that they: 1) generate relevant features for use in prediction tasks and 2) find local patches of data where supervised algorithms may struggle.” They presented a case study for HSBC that focused on Anti-Money Laundering (AML) where they reduced the false positive rate by 20%.
IO Tahoe (www.io-tahoe.com) do data discovery and relationship identification across heterogeneous data types, although that excludes video, pictures or voice. They claim to be technology-agnostic and will scrape any data lake. As you might expect this is “AI-driven”.
Privitar (www.privitar.com) like other firms (refreshingly) don’t claim to provide all the answers insofar as they don’t do data discovery. What they are concerned with is privacy. They seek to remove identity but preserve data useful for analytics, so you don’t have to throw away all the data you’re collecting provided you can demonstrate that you are using it in a legitimate fashion. They employ lots of cool tricks to this end such as homomorphic encryption, they ring fence subsets so they can’t be recombined (and have no common keys), sandboxing, and by adding a watermark (meta-data) to data so if it’s found where it shouldn’t be then you can prove its origin by extracting the watermark. Maybe Privitar are not using AI+ML in its strictest sense, but without legally cleansed data your reasoning system will be crippled, especially by GDPR if you use personal data, so it’s worthy of inclusion here.
Finally, if you have a large enterprise that needs a guiding vision those nice folk at Thoughtworks will be able to point you in the right direction and give you something to show your Board that you mean business (or maybe Business); and if you need infrastructure support to get your vision kickstarted (in the UK at least but on cloud) try www.behindeverycloud.co.uk.
As you may have realised, they don’t all provide a “soup to nuts” solution but you may be able to tessellate their services to solve your business problem. Using AI. And Machine Learning. Probably.
[Photo by Franck V. on Unsplash]