AI – Transforming Data into Business Decisions
Collecting data is no longer a significant challenge. However, the questions to address are: How should the data be utilized? And how can it be implemented into AI systems? Opportunities arise when data is leveraged to streamline business operations or discover new applications. These opportunities are best realized through a well-defined purpose, an open-minded approach, and meticulous data refinement.
An increasing number of companies are currently exploring artificial intelligence and machine learning on a broader scale. In a study conducted by the Info-Tech Research Group, insights were gathered from CIOs and other IT leaders globally regarding their AI plans for 2024. One out of three respondents mentioned they will introduce generative AI features in the coming year. The majority are looking to leverage AI for automating repetitive tasks and streamlining IT operations, while two out of five will use AI to define their business strategy.
Three levels of maturity
Kevin Söderberg, Senior Data Scientist and AI Tech Lead at HiQ, acknowledges these ambitions. According to him, companies typically embark on exploring AI systems through three entry points.
“Some companies still ride the hype wave without a clear idea of what they want to achieve. However, more commonly, we observe a certain level of maturity where the company has collected data and aims to explore how to use it. The third type consists of companies with a strategy for leveraging their data to create added value,” he explains, continuing:
“We encounter all three levels of maturity. What these companies have in common is a lack of in-house AI expertise, leading to a challenge in fully leveraging the data they have collected. Furthermore, there are few specialists in the market, making it challenging to bring this expertise into the companies.”
The pace of development requires continuous skill enhancement
Kevin Söderberg works together with Albin Sidås, Data Scientist at HiQ, on visual perception for autonomous vehicles in one of Toyota’s future projects. Working on AI systems is complex and demands expertise across various domains. In the current Toyota project, this includes tasks such as semantic segmentation, which involves identifying structures in an image, object identification to pinpoint objects in the image, and object tracking to monitor the identified objects.
“Given that AI is a vast field and development is progressing so rapidly, it is essential for us to work in parallel while staying updated in our respective specialized domains,” Albin Sidås explains.
Right data at the right time
“Many are unaware of the limitations in various types of data and how to use the right data at the right moment,” Albin Sidås points out.
The data applicable to AI models range from unstructured data such as images and audio to semi-structured data such as event reports and HTML objects, as well as structured data typically found in spreadsheets, relational databases, and tabular data generated from sensors.
The type of data you use depends on the specific application. In visual perception, unstructured data in the form of images is predominantly used, while AI systems designed to detect credit card fraud, for instance, often rely on structured data.
“So far, most AI systems have utilized only one type of data, but we are now witnessing the emergence of what we call multimodal models, where different types of input data are combined, resulting in even more flexible models,” Kevin Söderberg explains.
Prioritizing data quality
Even if a company has amassed large volumes of data for a specific purpose, there is no guarantee that the data is sufficient. As an example, Albin Sidås outlines challenges that can arise with recorded real-world data.
“It can be difficult to achieve good quality and sufficient variation in recorded data. To address this challenge, it is essential to work with different lighting conditions, various types of cameras and lenses, and record in different environments, etc. Otherwise, there is a risk of accumulating large quantities of duplicated data that serves no purpose. If that’s not sufficient, to enhance your models even further you can complement it with synthetic data or use open datasets available, usually from universities and colleges,” he suggests.
“It can be difficult to achieve good quality and sufficient variation in recorded data. To address this challenge, it is essential to work with different lighting conditions, various types of cameras and lenses, and record in different environments, etc. Otherwise, there is a risk of accumulating large quantities of duplicated data that serves no purpose.”
Lower the threshold and start small
Both Kevin Söderberg and Albin Sidås observe that Swedish companies show interest and a reasonable level of competence in embracing AI and machine learning. However, there is also uncertainty about how AI systems can be utilized and whether it’s worth the investment.
“A piece of advice is to start with low-hanging fruit and demonstrate results through a Proof of Concept, such as using Large Language Models for document summarization to enhance efficiency. That way, the organization usually realizes that investing in AI systems is worthwhile,” suggests Kevin Söderberg.
More common with AI in product development
AI is primarily found in three areas today: building intelligent products, creating intelligent services, and enhancing a company’s internal processes, often with a focus on products and services.
“The success of an AI initiative is significantly influenced by the company’s technical maturity and its ability to embrace change,” says Kevin Söderberg. He continues:
“It’s more common to invest in AI technology within product and service development. Internal processes often rank lower, even though a more efficient internal process can ultimately increase the pace of innovation.”
Demands for explainability in AI-based decisions
AI systems offer numerous benefits, yet they also come with limitations and challenges. They are reshaping the way many of us work, with intelligent assistants already handling repetitive tasks. However, this transformation also introduces new tasks and responsibilities. For instance, models need updating and retraining with new data, and as of now, the system doesn’t perform these tasks autonomously.
“We have to understand that artificial intelligence is not an exact science. It’s a statistical model that requires a specific mindset to work with,” says Kevin Söderberg
In AI, the focus is on probability rather than causality, a departure from what many are accustomed to. AI is often perceived as a black box with outputs that seem inexplicable. However, there are methodologies to investigate AI models, including XAI (Explainable Artificial Intelligence), which is a framework for analyzing AI models and, to some extent, understanding how decisions are made within the model.
“It can be challenging to provide a mathematical proof for a result, but we can empirically demonstrate how the model operates and how it makes decisions,” says Albin Sidås.
Furthermore, there are requirements for specific outcomes, such as credit ratings, to be explainable. And in the EU, there is the European Centre for Algorithmic Transparency, whose purpose is to ensure transparency regarding algorithms and models.
“We have to understand that artificial intelligence is not an exact science. It’s a statistical model that requires a specific mindset to work with.”
Key success factors for your AI initiative
– Access to data and hardware for training models and analyzing extensive datasets is a fundamental requirement.
– Developing a strategy for how the company will navigate data and AI, offering direction, and ensuring responsible AI practices.
– Implementing a customized project process that recognizes that AI systems require a more circular approach than traditional software development.
– Fostering organizational openness to embrace the changes that may arise from the outcomes of models and analyses.