RegTech: Changing Regulation and Compliance
MIT researchers argue that new Artificial Intelligence “RUM” algorithms can analyze complicated text and produce easy to understand summaries.
Introduction and Summary
In what began as work on more advanced Artificial Intelligence (AI) to solve complicated physics computations, scientists from the Massachusetts of Technology (MIT) have produced new algorithms that surpass existing natural language capabilities. Much research has gone into the possibilities of AI algorithms to analyse text, whether it be the statutory instruments (Zubek, et. al) or financial regulation (Banwo) but these new algorithms go further by producing crisp, logical and brief summaries of complicated scientific topics. We already know AI algorithms can tell us whether a specific text is relevant and important or unimportant for particular purposes but now we know it can accurately summarize this text as well. If these algorithms can break down, scale importance and relevance and summarize scientific studies on AI and physics, there is no reason why it can’t do the same for financial regulation or other relevant regulatory events such as anti-money laundering notices.
The MIT Research
In a paper entitled “Rotational Unit of Memory: A Novel Representation Unit for RNNs with Scalable Applications” by researchers from MIT and the Qatar Computing Research Institute build on existing AI algorithms used for language analysis and text summation.
Existing AI algorithms struggle with language analysis and text summation of lengthy and complicated texts for the same reasons they struggle in mimicking human conversation and interaction accurately, namely they lack the ability to learn and adapt their understanding over extended periods of time. The MIT researchers deploy a new method called “Rotational Unit of Memory” or RUM to improve language analysis and text summation to a standard they term “state of the art”.
Taking a step back: What are AI and Neural Networks
AI or machine intelligence is the ability of an agent, using a table of historical observations, actions and awards, to take action in response to an observation that would lead to a desired reward. The complexity of an agent, its ability to observe, analyse rewards and respond accordingly correspond to its intelligence.
Neural networks are a form of AI and try to copy human learning. Computers running neural network algorithms “learn” or recognize patterns and then deploy this recognition in tasks such as identifying objects in surveillance video or even fraudulent financial transactions.
What this means for financial services
RUM algorithms being deployed in regulatory impact assessment and analysis would mean that the underlying infrastructure of regulation would not have to change to make it machine-readable. RUM algorithms would also enable regulated firms to develop more coherent 3 Lines of Defense framework within AI environments.
For example, in 2017, the FCA conducted a TechSprint demonstrating a proof of concept for machine-readable regulations for regulatory reporting. Last year it announced, and is now conducting, a pilot. But with improved AI using RUM algorithms, very few changes would be needed to enable FCA consultations and regulations to be machine readable.
But while machine-readable regulations are a worthy goal, there are incremental and valuable uses of these algorithms before the industry and regulators arrive at an agreed approach for machine-readable regulations. Trade and social media surveillance, internal investigations, but more importantly regulatory impact assessment are all feasible use cases of these algorithms.
3 Lines of Defense
It won’t just be front office programmers that will be responsible for developing and utilizing new AI resources. Compliance, Legal and Audit functions will also need to be intimately involved in the deployment of AI by regulated firms. These functions should have the capability to not only monitor but to confirm that the AI algorithms are operating properly within regulatory and operational risk appetites. Second and third line functions will also be responsible for developing overarching policy frameworks.
Additionally and ultimately, regulators will need to build more expertise in this area, potentially approving or releasing specific algorithms designed for specific regulatory purposes. Regulatory oversight should be at a minimum comparable to the oversight they perform regarding AI used in other areas such as automated trading.
These new algorithms could be revolutionary. For larger firms, AI technology could free up staff to focus more on clients or allowing firms to invest more resources into the economy by reducing regulatory and operational overheard for smaller transactions. For smaller firms, it would allow more innovation and activity, increasing their ability to expand geographies and client bases by reducing compliance risk.
But in order to make this happen, regulators will need to step up their game.
Here is abridged sample text from a scientific article used in the MIT study to demonstrate the efficiency of their algorithm:
Story (abridged) The raccoon that topples your trashcan and pillages your garden may leave more than just a mess. More likely than not, it also contaminates your yard with parasites — most notably, raccoon roundworms baylisascaris procyonis(…) That istrue in varying degrees throughout North America, where urban raccoons may infect people more than previously assumed. Led by Weinstein, the UCSB researchers wondered if most human infections went undetected… Their study, appearing in the CDC Journal Emerging Infectious Diseases, found that 7 percent of surveyed individuals tested positive for raccoon roundworm antibodies. That was news to Weinstein, who said the researchers wouldn’t have been surprised if they’d found no evidence of human infection… Over 90 percent of raccoons in Santa Barbara play host to this parasite, which grows to about the size of a №2 pencil and can produce over 100,000 eggs per day (…) Sometimes they reach the brain, with potentially devastating consequences. This infection, termed “baylisascariasis,” kills mice, has endangered the allegheny woodrat and has caused disease like blindness or severe brain damage in dozens of people, including a toddler in Santa Barbara back in 2002.
Here is a summary produced by commonly-used AI LTSM algorithms today:
“baylisascariasis,” kills mice, has endangered the allegheny woodrat and has caused disease like blindness or severe consequences. This infection, termed “baylisascariasis,” kills mice, has endangered the allegheny woodrat and has caused disease like blindness or severe consequences. This infection, termed “baylisascariasis,” kills mice, has endangered the alleghenywoodrat.
Here is a summary produced using the new RUM algorithms developed by MIT researchers:
Urban raccoons may infect people more than previously assumed. 7 percent of surveyed individuals tested positive for raccoon roundworm antibodies. Over 90 percent of raccoons in Santa Barbara play host to this parasite.
Dangovski, R., Jing, L., Nakov, P., Tatalovi, M., Soljaci, M. (2019). ‘ Rotational Unit of Memory: A Novel Representation Unit for RNNs with Scalable Applications’, Transactions of the Association for Computational Linguistics, Vol. 7, pp. 121–138.
Banwo, A. (2018). ‘Artificial intelligence and financial services: Regulatory tracking and change management’, Journal of Securities Operations & Custody, Vol. 10 №4, Autumn 2018.
Zubek, R., Dasgupta, A. and Doyle, D. (2017). ‘Predicting the impact of legislative texts: An application of supervised machine learning to statutory instruments in the United Kingdom, 2005–2015’, Applied Quantitative Text Analysis Conference, 2017, p. 4.