Despite its alleged conservatism, the financial industry is deeply engaged in digitalization processes. Financial institutions actively explore various technological initiatives to provide more lean, efficient and up-to-date services. Among the most intriguing of them are the applications of machine learning models, the algorithms that allow computer programs to automatically improve through experience (Medium, 2019). Currently, these algorithms are steadily gaining prominence in multiple different sectors of the financial industry. The use cases include chatbots assisting the customers with small inquiries, valuation of financial instruments, option hedging, marketing and many other tasks which were traditionally performed by human employees. Although it sounds exciting that artificial intelligence takes over huge volumes of challenging human work, it would be irresponsible not to wonder how credible and accurate these systems are. Therefore, in this blog entry, we explore the flaws and opportunities of machine learning algorithms using machine vision solutions as an example.
What is machine vision? Machine vision is a system of hardware (such as cameras), image sensors and image processing software that provides image-based automatic inspection and analysis. In essence, machine vision allows the robots to "see" their surroundings (Emerj, 2019a). For example, the machine vision can use the satellite images to evaluate high-level changes in economic circumstances, recognize the language in the written documents as well as read human emotions by detecting the changing facial expressions. In this article, we focus on the machine vision technologies that use the neural networks as an image processing method. Unlike other work performed by neural networks, the image processing tasks can be easily benchmarked to human vision, which allows showcasing the advantages and disadvantages of these robots in a more concrete fashion.
How do we use machine vision in the context of the financial industry? For instance, satellite imagery recognition could unlock new opportunities for insurance companies (Venturebeat, 2019). By accessing property features data, the insurance agent could provide a quotation for services in a faster and cheaper way. Moreover, the solution could be of use for the reinsurers that could monitor client portfolios and test the accuracy of the provided data. In other words, such a solution would ensure a more thorough risk management at a lower cost.
Furthermore, machine vision can bring value by generating investment insights. For instance, the Orbital Insight platform uses a multitude of geospatial data sources to cater to the information needs of hedge funds and large corporates (Bloomberg, 2019). The satellite information may help analysts to form more accurate investment predictions.
Additionally, machine vision techniques may prove useful for automated document pre-population (Emerj, 2019b), which could ease the bureaucratic burden within large organizations and enable the companies' clients to allocate their resources to more value-generating activities.
What could go wrong? It is easy to get overly enthusiastic about the bright and innovative future of finance provided by machine vision technologies. Indeed, the potential cost-cutting and the upside from the investing accuracy sound promising. However, the neural networks, a family of machine learning models that often power the image processing software, have their flaws.
A failure to correctly classify adversarial examples is one example of such a flaw. The neural networks utilized in machine vision are particularly susceptible to this issue (Towards Data Science, 2019). Introducing adversarial examples means adding carefully constructed noise to the pictures "shown" to the robot. The perturbations create an "optical" illusion for the neural network that misleads the robot to "see" a wrong outcome with high confidence. In the picture below, taken from a 2014 study by (Goodfellow, I. J. et al, 2014), this effect is demonstrated by adding noise to the image of a panda, leading the robot to recognize it as a gibbon.
How can banks and insurers defend themselves? Unfortunately, the presently available remedies against the adversarial attacks are not particularly efficient. The first method is the adversarial training, which entails mimicking the attack against the tested network and then training the model not to be misled by them. This method improves the generalization of the model but fails to provide acceptable levels of robustness. (Towards Data Science, 2019) Moreover, this strategy ends up being an exercise of catching up with an elusive attacker, which has no guarantees of effectiveness. The second method is the defensive distillation. The data scientists train the secondary model with a smoothed surface in the directions on attacker would typically try to exploit. As a result, it becomes more difficult for the perpetrators to discover adversarial input tweaks that could mislead the network. This method worked with the earlier variants of adversarial attacks but has been beaten by the more recent and more advanced ones, like the Carlini-Wagner attack (Towards Data Science, 2019).
It is unlikely that algorithms spotting visual patterns in financial context would be more resilient to adversarial attacks than the system used in the study above. However, in the context of the financial industry, the price of a deceived system may become very high. The adversarial attack to the system evaluating the investment or contract opportunities could lead to misjudgements worth millions and billions of customers' funds. As a result, a vulnerability in such a system would be capable of inflicting lasting financial and reputational damage to the institutions.
Innovation is necessary for the long-term survival of financial institution, so there is hardly any doubt that value-adding machine learning solutions such as machine vision systems will become a new norm of tomorrow. However, it is essential to include monitoring and validating of these offerings as part of a value proposition. This step is a crucial element of building trust, ensuring the quality and security of financial products as well as managing risks attributed to cutting-edge technology. Moreover, it is only a question of time until the regulators come up with the risk management requirements for machine learning-driven systems. For instance, the British FCA has already begun analysing the special features of the neural networks and has specifically highlighted the importance of interpretability of solutions using machine learning methods.