Integrated Artificial Intelligence

In this interview, Marcin Kowalski, a consultant from GFT Poland, answers questions pertaining to application of solutions based on artificial intelligence in retail and investment banking, emphasizing the challenges associated with introducing AI as a functional tool.

Considering the tremendous pace of digitalization of banking services, what are the business consequences brought about by AI?

Marcin: With the emergence of technologies that allow for large scale implementation of AI, companies obtained tools allowing for acceleration and automation of processes, as well as reduction of costs.

Today we have biometrics, self-driving cars, mechanisms in industrial systems that detect failures weeks before any person could do that, antivirus software capable of tracking viruses never seen before, insurance companies that verify declared damages in a matter of seconds…

New opportunities for business are fantastic. It became possible to build competitive advantage in sectors that were, until now, very stable and difficult to shake.

You’ve mentioned a vast spectrum of applications. One could ask: what is the precise meaning behind the notion of artificial intelligence? Especially in the context of the financial sector. And how does Machine Learning relate to that?

The mainstream terminology is a bit incoherent. The term “artificial intelligence” refers to machines being able to perform operations normally attributed to humans, such as speech recognition. Likewise in business contexts: we talk about artificial intelligence in the case of tasks at which, until now, only humans could excel. One use case of speech recognition in the financial sector could consist in automated customer service, where conversations are held with a computer. This is something that cannot be programmed using traditional methods, since each person speaks a bit differently, with various accents and often in noisy environments. Too many complications. You need a different approach.

The term that binds all these methods is machine learning. This boils down to learning on examples. Instead of manual definition of principles by which they would distinguish between sounds, machine learning algorithms obtain sets of recorded utterances together with their meaning. If an algorithm was selected properly and we supply it with a sufficient number of examples, after some time it will learn to understand utterances that it hasn’t heard before.

Machine learning is a method of building artificial intelligences, but it may also pertain to tasks in which people traditionally perform badly. Here’s a good example of such a task: If we call Mr. Smith, will he take our loan? It’s difficult to say, as it depends on a plethora of factors. Computers can be very good at solving such problems, thus limiting the costs of marketing campaigns by huge margins.

If, however, the decision to call Smith is made on basis of rules prepared manually by a team of analysts, then this would neither be machine learning, nor artificial intelligence.

So what we are dealing with here is, in short, automation?

Our clients are mostly interested in this aspect, as it enables them to work quicker and less expensive. It is also easier to consider improving things that already function. But apart from automation, there are other opportunities. For example: OpenDoor, which prepares an offer of real estate sales in a matter of minutes. This is revolutionary and naturally poses a challenge for huge institutions. But automation itself is not easy.

What is the greatest challenge when implementing AI solutions?

Machine learning is about learning on examples, so the biggest challenge lies in data. In particular, the recently popular set of methods called “deep learning” requires enormous amounts of data. The subject of Big Data is as important for good reasons. We are observing a renaissance in AI because computers became much quicker and are able to process much more. Most importantly, data must always be well organized, regardless of the required volume.

80% of the success of AI projects depends on data. Analysis and selection of algorithms are not simple tasks, although in this particular area, organizational or integration-related problems are not as frequent – this process is easier to control. The problem of integration is as old as the IT industry, but in the context of artificial intelligence, it gains particular significance.

An interesting analogy can be made with regulatory projects. They have a similar goal: to gather information from multiple locations, so that regulators may analyze it in terms of incorrectness. One may use AI for such analysis, but this is already the next step. The main challenge is to gather data from very different systems, and these systems are often disconnected within a single organization. Banks are often characterized by long histories of mergers and acquisitions, hence integration may be incomplete.

So in the entire project of introducing AI in a financial institution, developing the “thinking” is merely the tip of the iceberg, whereas it is the process of integration that is actually essential.

Indeed. Our experience clearly shows that many projects face the challenge of insufficient availability of data. Even in the case of simpler solutions, such that can be developed by an analyst without experience in machine learning, the material for analysis is necessary. Our goal is to make our clients realize that hiring a specialist with PhD in statistics will not solve the problem, as he has to have something to work on in the first place.

What are the challenges specifically related to integration of systems in the context of AI?

Everyone involved in integration processes in IT departments knows what kind of challenges they entail. What is worth emphasizing about projects in which a data science team operates, are the team’s special requirements. Is it necessary to deliver data online, or is a daily load sufficient? Is incompleteness an issue? Perhaps a sampled subset would be enough? In the case of deep learning projects, volumes may be challenging and it may imply clustering. And when scalability is on the table, moving a solution to the cloud is worth considering. Moreover, a lot depends on the structure of the data science team. Is it supported by data engineers and DevOps specialists? If not, perhaps we should consider hiring people with these competences? Otherwise, separating the team from technical issues may turn out to be necessary. It is worth noting that the industry is leaning towards such an approach. An example of this may be the Cloudera Data Science Workbench that is now under preparation, which may be treated as a referential architecture of AI operations on basis of the Hadoop stack. Google DataProc is heading towards this direction as well.

Despite the obvious challenges, the market has already decided on investing in AI.

Naturally. Imagine a desk that analyzes transactions in terms of money laundering. Let’s say it employs 1000 people. Improving the efficiency of reporting suspicious transactions by 10% will enable us to reduce a huge number of jobs. We estimate that AML (Anti-Money Laundering) will become one of the main areas of investment in AI solutions. And this is just one of many such areas.

This leads us to the final question. Artificial intelligence is often seen as a threat – for reasons that may be more or less rational – concerning the degree to which it will replace humans. How do you perceive this problem in the context of financial services?

A person performing a boring, monotonous job is not motivated. Artificial intelligence enables our clients and their staff to focus on building strategies, innovative solutions, relations with clients and increasing their market share. Business will always depend on people, but tedious tasks can be relegated to computers. Mass automation has been taking place ever since the industrial revolution and I don’t know anyone who would wish to reverse this progress.

Ultimately, if you fail to introduce an AI strategy, your competition will do it anyway. If they can succeed with 50 people working on a given task – a task that in your organization requires 500 staff – you will simply be more expensive. What are your clients going to do?

Thanks for the interview, Marcin!

Post a Comment

* indicates required

Comment Area

  1. An interesting article.

    “Although the long-term goal of the neural network community remains design of autonomous machine intelligence, the main modern application of artificial networks is in the field of pattern recognition” – Joshi et al. (1997). 

    However, whilst connectional patterns and network topology could be used to define a rough guideline able to determine network performance and estimates in convergence speeds, similarly network size used to solve pattern recognition problems plays an important role in understanding and analysing cost functions, efficiency and generalisation.

    One of the common misconceptions or mistakes in Neural Network design is to inadvertently create networks that are inherently too complex with parameters often containing too many neurons, resulting in overfitting. Similarly, networks designed too simple consisting of too few neurons or parameters could result in non-convergence or over saturation of connectional weights leading to poor performance and generalization problems and a balance needs to be established without trading off performance against pattern fitting.

    I agree with you that quality and quantity of data plays an important part in AI training leading to more predictive networks but whilst Data Mining techniques are used to automatically find associations in datasets, applying pattern recognition algorithms generating approximations of classes, data gathered and classification requires additional off-line processing and analysis to understand and recognize what has been clustered.

    Similarly, knowledge discovery and expert systems created through exploratory processes using data analysis, requires an understanding of data partitioning and correlation problems before being able to develop rule based systems modelled on observations and patterns.

    Zhang and Bivens (2007), performed comparative analytical studies between Bayesian Networks (BN) and Artificial Neural Networks (ANN) specifically focusing on model performance, concluding that Bayesian Networks “are less sensitive to small dataset sizes” or datasets with incomplete data, typically better suited for environments that change rapidly. 

    However, studies have also concluded that Artificial Neural Networks, although less accurate and subjective to training methods are able to achieve faster model evaluation times, suited for environments that demand intensive high volume real-time dependent predictions, such as that found in finance.