Page 26 - GS210901

CoverStory

chines can do in seconds, researchers noted. On the Oppor- ally as well as in concert." Regarding heightened probabil-
tunities and Risks of Foundation Models, a study published ity, he added, "Better than trying to predict the future (you
Aug. 18, 2021, by the Center for Research on Foundation know you'll be wrong, just not by how much) is to under-
Models, drew sharp distinctions between traditional AI stand the distribution of potential outcomes and simulate
and today's advanced AI capabilities. the likelihood, impact and subsequent actions for poten-
tial future scenarios."
"Most AI systems today are powered by machine learn-
ing, where predictive models are trained on historical data Explaining feedback loop development, Snitkof said, "[T]
and used to make future predictions," CRFM researchers he best organizations build a learning machine, investing
wrote. "The rise of machine learning within AI started in data quality, thoughtfully designed data architecture,
in the 1990s, representing a marked shift from the way and the ability to construct a feedback loop that can in-
AI systems were built previously: rather than specifying crease the accuracy of analytics over time."
how to solve a task, a learning algorithm would induce it
based on data—i.e., the how emerges from the dynamics Berkhahn stated that next-gen AI is generally more con-
of learning." textual than prototypical AI solutions, which helps service
providers authenticate legitimate customers and thwart
AI foundation models use neural networks and self-su- criminal activity. "We want to know the true owner of a
pervised learning to assess massive swaths of data, CRFM company, what type of company and the risk threshold
researchers noted while also cautioning against misuse. of the company's transaction monitoring," Berkhahn said
"Foundation models are scientifically interesting due to about contextual analysis. "That contextual information
their impressive performance and capabilities, but what helps identify and block criminal activity."
makes them critical to study is the fact that they are quick-
ly being integrated into real-world deployments of AI Renshaw cited automation and machine learning as ad-
systems with far-reaching consequences on people," they ditional components of next-gen AI. Regarding automated
wrote. processes, he said, "Predictive analytics solutions should
also look for efficiencies beyond model performance. Au-
Defining characteristics tomating various stages of the data science life cycle, from

David Snitkof, vice president of analytics at Ocrolus, men- data migration to model creation to continuous model
tioned the biggest change he has seen in predictive analyt- governance, will reduce operational overheads and ensure
ics is the sheer mass of diverse and disparate data sets, in- that the solution continues to stay robust."
cluding data from unstructured documents used in credit
risk modeling. "[O]ur world is awash in data, but much of On the role of machine learning, Renshaw said, "Robust
this data is unstructured, noisy or in hard-to-reach plac- and effective predictive analytics solutions should be
es," Snitkof said. "With the use of AI to extract valuable based on the ML best practices to achieve performance,
signals from documents as well as digital data feeds, it is speed and adaptability." He added that solutions should
possible to leverage a greater variety of data and use it to be tailored to customer requirements.
offer more personalized financial products." Early-stage growth

Nicole Newlin, vice president of solutions at Ocrolus, Reflecting on AI's evolution, Renshaw identified three
agreed, stating solution providers need to focus on their stages of growth: early-stage, predictive and prescriptive.
inputs. "The human plays a big part in providing the data Early-stage financial fraud solutions were driven by rules
a machine learns as it becomes self-sufficient," she said. that looked for certain characteristics in transactions to
"[Solution providers] need to remove bias or discrimina- identify, block and report fraud, he stated. Today, a rules-
tion and manage incomplete data to build the best possible based approach is still relevant but has limitations and
model." may even be counterproductive from an operational effi-
ciency and management perspective, he added.
Snitkof further noted that legacy AI solutions typically
focus on credit bureau data, which may not present a bor- "Legacy fraud prevention solutions detected and alerted
rower's full financial picture. He advised lenders to con- on malicious behaviors associated with device-centric
stantly evolve data collection, heightened probability and data,." Renshaw said. "Initially, this approach created
feedback loop processes. ‘deny lists’ and ‘allow lists’ to identify good vs. bad users
and were notoriously out of date."
Dynamic analytics

"The best lenders collect a lot of data, both in terms of As early models evolved, they became capable of detecting
quantity as well as variety," Snitkof said about data collec- malicious activity by identifying malware, devices, and
tion. "They evaluate signals from many different sources networks associated with fraudulent behavior, Renshaw
and explore how they correlate with outcomes, individu- said. From there, they attempted to identify individual
users by analyzing their behavior and actions, he noted,

26