Let’s talk Anti-Money Laundering (AML)!
When you Co-Founded MinervaAI with Jennifer Arnold and Victor Tay, you were working at the McMaster Stem Cell and Research Institute using deep learning to discover more effective pharmaceuticals to treat leukemia.
How are the applications of AI for cancer research like using AI to manage AML?
When I was working on cancer drug discovery research, we were looking at ways to make sense of all the structured and unstructured data out there around different potential treatments. When MinervaAI searches enhanced due diligence (EDD) information on an individual or an organization, she looks at over 50,000 sources – 4.5 billion pieces of data in 55 different languages – so they’re remarkably similar tasks. They’re both automating the process of an investigator trying to connect the dots, to automate what is tedious and to make sense of what they are seeing.
“When you translate [cancer treatment research] over to what we’re tackling in the anti-money laundering space…it’s kind of a similar problem. With AML focused enhanced due diligence you’re also still trying to dive into these massive sources of structured and unstructured data…and make sense of them very quickly.”
In leukemia cancer research, the investigative problem is trying to discover evidence from clinical data and the literature and trying to differentiate what genes or treatments are important to influence a better outcome. These are usually varied sources, such as published research articles from medical journals, human genome expression matrixes, and longitudinal patient studies. You’re trying to read and interpret this huge basket of information all at once trying to tie the dots together to understand where the risks are in cancer. The information must be relevant, valid and stand up to peer review in the form of a manuscript.
When you translate that over to what we’re tackling in the anti-money laundering space to conduct Know Your Customer (KYC) and EDD, to get a full view of the customer risk profile, it’s kind of a similar problem. With AML you’re also still trying to dive into these massive sources of structured and unstructured data trying to read and understand vast volumes of diverse social media sources, news information, published registries, legal information, sanctions, and politically exposed persons (PEP) lists. You’re trying to marry that all together for a cogent narrative to make better business decisions, faster. All the information you collect from the various sources also needs to be tied together in the form of an auditable AML relevant report that is consistent among each investigator and can stand up to regulator scrutiny.
In addition to removing the manual work from the investigations team, one of the things that we hear from our customers is how impressed they are with MinervaAI’s ability to look at all this data, in less than 30 seconds. What makes MinervaAI so different?
Our financial systems and the users of those systems are global in nature, so there are many sources to consider during an enhanced due diligence review or AML investigation. That data also varies greatly in structure depending on the source or the jurisdiction so it’s critical to be able to understand the context behind the information to establish the level of risk. That’s where machine learning and deep learning techniques really come in handy because they can pull these points together and make sense of them very quickly. This helps our customers reduce the noise and eliminate false positives – which at its core is irrelevant or low value data that clogs up the decision-making process.
“…think of a traditional keyword matching algorithm like a single-layer, non-scalable solution – a kind of “What you see is what you get” scenario; but with deep learning, you’re able to actually automate much deeper levels of inference that can be really useful, especially when you apply that across the amount of data you need to consider as part of onboarding a client or a deeper AML investigation.”
With MinervaAI, we’ve created name entity recognition and classification models that apply a more “human-like” intuition to the information the models are reading. Human intuition applies a much more complex set of vetting criteria than simple keywords. For example, if a person reads an article about an expert giving their opinion about money laundering or who is involved in special interest groups that combat human trafficking, the person reading the information can comprehend that this person is not a criminal element. In traditional keyword matching, the logic would scan for money laundering and human trafficking without being able to distinguish the criminal piece from what is altruistic. Deep learning algorithms can consider the whole body of text and understand deeper levels of information, how those words relate together. So even though the words “money laundering” are in the article, MinervaAI can also understand how those words relate to all the other words in the body of the text to understand those deeper patterns and relationships.
Deep neural networks have many different layers. You can think of a traditional keyword matching algorithm like a single-layer solution – a kind of “What you see is what you get” scenario; but with deep learning, you’re able to actually automate much deeper levels of inference that can be really useful, especially when you apply that across the amount of data you need to consider as part of onboarding a client or a deeper AML investigation.
MinervaAI’s adverse media model leverages natural language processing to rank and score sentiment to reduce false positives.
Keyword matching is a static, non-scalable, one-time algorithm – but with deep learning, you can take a lot of information in and have that model learn from that information. As you continue to pipe new information into the model, it continues to learn, and it continues to innovate. There’s a lot of opportunity for these models to improve or “self-correct” to reduce the false positives, and that’s what our customers are really noticing as our differentiator. It helps the team focus on what matters; to remove the bottle-necks in customer onboarding created by traditional processes and systems.
Embracing true AI can be daunting for some investigations units because it’s a departure from what they’re used to. What advice would you give leaders who need to modernize their AML solutions?
This is where the concept of Explainable AI (XAI) is becoming more and more critical in modern applications and industry. It’s natural for people to have a fear of what they feel they can’t control or hesitancy to embrace a new process because it can be difficult to change existing habits. By introducing concepts of Explainable AI, where you can trace the lineage of the model predictions, you can understand certain insights more clearly. It gives people a little bit more confidence when they can see the explanations behind those insights. Traditionally, a lot of machine learning algorithms were treated as a black box, but human nature being what it is, we need to understand why those decisions were made. Explainable AI is really about trying to get into the deeper layers of those models, to break up the models instead of having one big black box model so it’s easier to manage, trace and understand.
Our financial systems and the users of those systems are global in nature, so there are many sources to consider during an enhanced due diligence review. That data also varies greatly in structure depending on the source or the jurisdiction so it’s critical to be able to understand the context behind the information to establish the level of risk. That’s where machine learning and deep learning techniques really come in handy because they can pull these points together and make sense of them very quickly. This helps our customers reduce the noise and eliminate false positives – which at its core is irrelevant or low value data that clogs up the decision-making process.
“This helps our customers reduce the noise and eliminate false positives – which at its core is irrelevant or low value data that clogs up the decision-making process.”
That’s what we do at MinervaAI, we break logic up into components, instead of having one gigantic model that makes a single broad inference about risk and it spits it out. We tie several smaller models in a chain through engineering where we can trace where the data is coming from and how it’s being transformed. When our customer generates a final report, they can see all the decisions and probabilities, the data lineage, and changes to that data, presented in one place. It just gives people a lot more confidence because they can see how the sausage is being made, and that it’s being made right – in addition to greater efficiency over the manual approach that many groups use today. Transparency is an important thing to continue to phase into the deep learning solutions industry – in fact, regulators are demanding it.
Having used AI in support of innovations to treat cancer and to fight financial crime, what do you think is next for AI?
I think it’s important for industry to not be afraid to experiment with taking new directions for AI. We’ve seen over the last decade a lot of interesting innovations that have come out of organizations who are willing to take that step forward, to take on the risk of implementing new technology, and we’ve seen those risks pay off. New technologies that are bringing forward AI, deep learning, and automation are boosting efficiencies to a large degree so that our work force can focus on things that really matter.
An area that we need to explore further is activity within the dark web and tying together what we can see in the public domain with what’s happening below the line in the deep web. By leveraging tools like MinervaAI that maintain the entire data lineage, it could change the game for investigations into human trafficking, the drug trade and ransomware attacks.