M. Sc. Kirsi Louhelainen

M. Sc. Kirsi Louhelainen works developing tailored-made software for real-life business problems. In other words, her work mainly concerns customized Machine Learning (ML) tools to automate ordinary business transitions such as forecasting demand, predicting bankruptcy, dynamic pricing, etc. She studied a bachelor's degree in computer science and obtained a master's degree in the Helsinki University of Technology (TKK) in Finland. Currently she works in Amazon Web Services (AWS), helping businesses with technology, cloud, data, and Artificial Intelligence (AI).

 

ML/AI in real life: how to get value instead of getting academic exercise

 26/marzo/2021  Seminario PISIS-UANL 2021      Asistencia : 30

Introduction

AI and ML have an enormous potential for helping businesses to reach a higher performance and to improve their operations. However, there is a mismatch between the technological realities and their practical applications. M.Sc. Louhelainen outlines a map of the key problems and hindrances in implementing AI and ML solutions to companies, and shares some of the experiences in project development and gives us some recommendations to reduce the risks of failure.

Summary:

As the title of the talk suggests, Kirsi aims to make ML and AI operational enough to solve many hindrances ordinary business face on a daily basis. The problems may vary but it is required more than producing software,deal with data analysis and even with team managing. This talk is a summary of her own learning process and lessons absorbed during her professional career. It ispointed out that very few ML projects can make it into production, as In 2020, only 53 percent of them make it into the final stage. There are many issues stopping the potential of ML mainly regarding operations, maintenance and even the timing in business processes or coordination and communication between different kinds of professionals and professions.

She narrates five stories embedded with key lessons for the audience. The first one deals with a project for automating an accounts paying system. An important lesson here is assessing the nature of the problem well before developing an algorithm, it would be better to look at the tools available first and select the most fitting one. There are cases that do require a specific algorithm, like the case of a company running a global warehouse which they helped out, but most operational improvements can be tackled in other, more efficient ways. The second case also turned out complicated and she highlighted the many problems between data scientists and software developers and how the true dimensions of the problem go beyond coding. One piece of advice she provides is to avoid letting a data engineer rewrite the code for production. It is much better to have the same person responsible for doing the whole process, from testing data to production. The third story concerns the importance of good data analyzing (in this case contained in XMLs files) by revising and fixing the data; they managed to have a 12 per cent improvement in three weeks. So, she concluded that the quality of the data is critical. In the end, the algorithm was secondary yet and it was important to figure out the best. The most typical problems she has found are connected with typos, human data errors, discrepancies between systems, erroneous specifications and data structure changes on the fly. In complicated cases a data scientist may be extremely relevant. The fourth story revolves around an accommodation website that needed dynamic pricing and the problem was maintenance. She mentions basic maintenance issues such as alerts, debugging, server upgrades, library updates and security patches. Yet a critical point concerns where to store the models, keep safe its different versions in case of any hazard and ensure the model keeps updated even if data changes resorting to CI-CD practices. The fifth story concerns human resources evaluations as a way of updating their recruitment software. The final story pertains to the case of a construction company that only needed a better way of updating their data and she proposed only to rely on Google Forms. The later cases mirror a larger problem she wants to address concerning the lack of maturity in data practices, methods, and basic analytics that could be easily incorporated into machine learning processes but remain poorly developed in businesses. In a way, she concludes that projects should align themselves with specific needs and curiosity is a key skill to devise the nature of the problem and its specific solution.

She considers it necessary to combine well the different skills from data scientists and software developers to build up a team able to detect problems with how data is gathered and how it will be used, updated and maintained. Teamwork and trust also comes in handy as a way to solve problems and a good antidote against all eventualities.

Conclusions

The main conclusion hints at finding fitting solutions to problems beyond developing an algorithm. Most problems can be addressed with other tools and strategies. In fact, for standard problems and practices in business there are already well developed softwares that can be employed successfully after a good evaluation and test. The key of the matter consists on understanding well the nature of the business and its problem and devising a solution accordingly. In this fashion, it is more likely to come up with an idea better aligned with the business needs. A major hindrance comes with the quality of the data. After all, after fixing the data the problem may disappear. In the end, she discovered more problems related to administration and implementation of solutions rather than with the specifics of programming (code versioning, model repositories, storage features, tools for continuous model building and testing and monitoring models).

Enlaces relevantes

Reseñas anteriores