Harnessing the power of deep learning
“How to harness the power of deep learning?” This question was discussed in the “Deploying deep learning-based models in projects and real-world cases – challenges and solutions” webinar organized by FCAI on March 11, 2021.
In the beginning of the event, Arno Solin, Assistant Professor in Machine Learning at Aalto University, gave a short introduction to FCAI and to FCAI’s Next-generation Deep Learning research program. He also introduced the content of the event. The speakers, representing different viewpoints, came from the large technology-driven industrial company Saab, rapidly growing AI expert company Silo AI, and from the cutting-edge AI research community within FCAI.
The first talk argued that AI will transform the entire defense sector. However, despite this vision, the steps are cautious as high trust in AI-driven systems is required in all aspects, says Petteri Alinikula, the CTO of Saab Finland. Still, Saab invests a lot in R&D to enhance and to create new capabilities for existing sensor products used in surveillance.
To increase understandability and to support decision-making Saab is, together with FCAI, developing deep learning algorithms that can also express uncertainty. Saab has also recently developed a high-performance big data and computing platform, Mimir’s Well, that is able to receive and store massive amounts of all kinds of data like satellite images and videos. According to Petteri Alinikula, Mimir’s Well will enable new kinds of predictive applications, e.g., by combining situational awareness data and time series data.
Since the foundation of Silo AI three and half years ago, the company has finalized more than 100 different AI application projects together with its customers. The growth of the company has been fast and Silo AI is constantly recruiting new experts. In his talk, Silo AI’s Head of Technology Niko Vuokko emphasized the importance of trust and a strategic approach in the deployment of AI systems. Many AI technologies are new and may affect many parts of an organization.
According to Vuokko, building deep learning models is relatively easy compared to the big challenge of cost-effectively putting the AI systems running in production. When problems occur, deeper understanding is required on the underlying reasons, on the impact on the customer, as well as on how much it costs to fix things. It has also become clear to Silo AI that one cannot successfully bring AI into production unless there is access to the latest expertise. The development in the field of AI is so fast that, with the latest technologies and approaches, companies may reach even ten times better results than without.
During the past ten years there have been many research breakthroughs: for example, AI enabled image recognition, image generation, language translation, and antibiotic development. Deep learning models are extremely powerful and already routinely used in many industries. However, it is not always easy to work with them and there are several known challenges, as Research Fellow Markus Heinonen pointed out during his talk.
One proposed solution to increase transparency and understandability could be to use convolutional Gaussian Processes or other so-called calibration techniques to calculate the probabilities of results. The method selection depends on the complexity of the problem and available data. For simple problems, one should use simple models like linear functions. For more complex problems, e.g., convolutional neural networks, multi-layer perceptron or ensemble models are required.
According to Markus Heinonen, the interest in Bayesian Deep Learning has been increasing recently, as this viewpoint allows for the modelling of uncertainties and is shown to be more explainable and robust than many alternative methods. At the end of his talk, Heinonen listed known issues with deep learning techniques and challenged the industry to help in prioritization.
The panel discussion chaired by Professor Arno Solin covered a wide range of questions that the audience had sent beforehand and during the event. The questions ranged from making Finland the leading AI export country to how to preserve languages and alleviate loneliness with the help of AI. Some of the key messages from the panel were the following:
More people who want to build upon AI and put it into applications are needed.
AI is diverging in more and more different approaches, similarly to software or electric technologies. It is not realistic to expect that one specialist knows everything in AI nowadays. For companies, this is of course a big competence challenge.
The trust and ethics aspects of AI are important to consider. However, contemporary AI is a tool just like a knife or a car. The use of it depends on us people. Already today, it has been shown how AI can be used to create major impact, e.g. in healthcare. Also, the explainability and improved understandability of the AI algorithms in future solutions will help.
Nothing can be accomplished without adequate funding and the right combination of competences and human resources. The event ended with a presentation by Heikki Ailisto, Research Professor from VTT and Lead of FCAI’s Industry and Society Program. He highlighted the current funding instruments in Finland and European Union that support collaborative joint development.
Link to webinar website containing presentation materials >>