“To live is to transform into consciousness as wide an experience as possible” said André Malraux. Contemporary artificial Intelligence, often revered for its powerful capabilities, is far from these considerations.
This is not to deny the spectacular results obtained since the turn of the 2010s in the field of
Machine Learning (ML), due in part to the democratization of the computing capacities necessary for running these algorithms, and in part to the dimensionality barrier that convolutional neural networks (CNNs) have been able to address. Liberation from these limitations has opened the door to applications across diverse communities and domains. However, the rapidity at which the technology has been adapted and adopted has perhaps created as many illusions as it has promised solutions. Does every problem have an AI solution? Does every problem need an AI solution?
At Geomatys, perhaps due to a deep understanding of both the advantages and limits of AI gained during the doctoral thesis of one of our founders in the mid-2000s, we have, from the start, considered the power of AI as an extension of our activities rather than a completely new technological revolution.
Thus, we implemented CNNs very early on for e.g., object classification from satellite images, but without making it the alpha and omega of our future activities. Though not intentional, this fundamental understanding of the tool prevented us from falling into the collective mythology that was evolving around it. Instead, we continued to consolidate our fundamentals regarding the mastery of spatial information management for big data infrastructures, this tool being one among many others.
Whether the promises of AI are overstated or not, today’s companies are expected to boast their embrace of the technology when applied. It is out of this necessity – fad as it may be – that we are taking this opportunity to highlight our long history of applying AI tools, and to affirm that yes!, we take full advantage of this powerful advance in technology. Here we present some of the ways we integrate AI into our geospatial data-focused activities.
Today, imagery intelligence (IMINT) greatly mobilizes convolutional neural networks (CNNs) to efficiently automate the tasks of object recognition within images. With strong training data, this includes the ability to capture spatial correlations. Many companies have therefore positioned themselves squarely on this IMINT segment of activity. In contrast, we have decided to position ourselves on what we believe is the bulk of the still under-exploited potential of Machine Learning, and have started distinct projects on three of these domains over the last year.
- Axis 1, AI for predictive purposes: Indeed, CNNs used for automatic classification purposes produce their results on an observation at a certain time, so when it is analyzed, it is already in the past. They do not natively allow the combination of detected information to predict and correlate future behavior. However, faced with the flow of information likely to be detected, a human will often still need to prioritize particular elements and put them into perspective before making a decision. Yet machine learning tools, specifically complex architectures such as reinforcement learning or generative neural networks, have this capacity. One current project, aiming to automate the production of complex nautical cartography from heterogeneous and contradictory information that previously has relied on human evaluation, has allowed us to confirm the potential of complex neural network architectures. Illustrating the broad utility of this method, we are also using it in a separate project to predict the presence of animal populations based on environmental conditions.
- Axis 2, Edge Computing: In some contexts, the transit of an image from its acquisition site to its processing site is not feasible. In order to minimize the cost of the message, it is therefore necessary to classify the image on the spot in order to transmit only the deduced information. We are currently working with partners to test this type of device in real-world conditions.
- Axis 3, Detecting Weak Signals: Detecting weak signals implies exploring large amounts of data, and learning from this type of signal requires even larger amounts of data. However, in some cases, centralizing data in a single infrastructure is not realistic, either for reasons of data confidentiality or, more pragmatically, for reasons of the size of the storage space involved. Running part of the learning model as close as possible to the data and federating the results is a solution that we are currently implementing in the context of a project combining geolocation and health data.
Together with Geomatys’ 15 years of expertise in the field of interoperability, processing and massive geospatial data infrastructures, and the consolidation of this expertise in the Examind software suite, we are now working on transforming our experience in the field of machine learning into easily re-mobilizable functionalities for our customers.