A Q&A with DeepSea’s Director of AI Research, Antonis Nikitakis
Q: What is your most exciting piece of AI research at the moment ?
Explainability is critical for building trust in AI systems. Therefore, our most exciting AI research currently involves researching methods that improve models’ explainability – to enable ever-greater trust to be built amongst stakeholders. AI can be seen as a “black box” whose input and operations may not be transparent and understandable to outsiders, particularly when they are based on complex algorithms or deep learning models. AI can be vulnerable to bias or discrimination, especially if it is trained on biased datasets or if it is designed without sufficient consideration of ethical concerns. This can lead to concerns about the accuracy or reliability of the system, which can compromise trust and confidence in its use.
One of the most promising ways to enable trust is to promote models and methods that allow humans to understand the logic behind the system’s decisions. At DeepSea, we conduct research on models that integrate concepts and principles from Physics and Naval Engineering. This enables us to explain how the models operate and that physical principles are always respected. Additionally, we enhance explainability by introducing new methods and metrics for evaluating the performance of AI models in extrapolation regimes (i.e., outside the given datasets). Having a reliable approach to evaluating performance regardless of the available data quantity and quality is a significant step forward in promoting explainability.
Q: How is AI changing the competitive landscape of the shipping industry? What advantages can AI offer to early adopters?
AI is transforming the shipping industry in various ways, from optimising operations and increasing efficiency to improving decision-making capabilities with vessel-specific weather routing. The advantages that AI offers to early adopters are significant. We are already seeing the major impact that AI can have in our work with Wallenius Wilhelmsen. Utilising DeepSea’s Performance Routing technology, the vehicle carrier operator became the first global shipping company to adopt a fully AI-based approach to voyage optimisation. As an early adopter, Wallenius has already started to reap the quantifiable benefits of fuel savings and a reduced environmental impact.
AI can also provide early adopters with better decision-making capabilities reducing the risk of costly mistakes and improving their maintenance and operational efficiency. For example, with DeepSea’s Cassandra today we can very precisely estimate the fouling state of a vessel and its trend to help the operations team to optimally schedule a cleaning.
Q: What types of data are being collected and analysed by AI in the shipping industry?
Today there is a very large volume of data that is being collected. We have various sensor data from onboard sensors that is being collected regarding all key machinery and subsystems of the vessel such as hull, main engine, electronic chart display and information system (ECDIS), and generator engines, among others. This information is important for optimising vessel routes, predicting maintenance needs, and ensuring that each vessel is operating at maximum efficiency. Then, there is weather data, which is critical for ensuring the safety of vessels and crew members. There is live weather data from sensors, and forecast weather and historical data collected from third-party providers. Furthermore, there is manually-reported operational data (e.g. noon reports) and also live and historical data from the automatic identification system (AIS) which is responsible for transmitting a ship’s position so that other ships are aware of its position. By analysing all these types of data and turning them into practical insights, we can help shipping organisations to optimise their fleet performance, improve their operational efficiency, save on fuel costs, and decrease their carbon footprint.
Q: What are some of the challenges associated with implementing AI?
Most of the challenges in implementing AI today involve data-related challenges, challenges related to the model evaluation and its robustness, and stakeholder trust which I mentioned earlier. One major challenge is that of distributional shift. Put simply, distributional shift has to do with the mismatch that could exist in an AI model between the training environment and the actual deployment. This “shifted” data can be challenging for machine learning models. In shipping, for example, we have a prominent ‘latent phenomenon’ which is hull fouling. The fouling state of a vessel is a dynamic phenomenon that prevents a generic model from fitting the data. Data drift due to sensor miscalibration can also lead to a distributional shift. Special considerations and methods should be taken into account to ensure a consistent fit across the dataset of a vessel. DeepSea has developed sophisticated data-driven methods for unsupervised anomaly detection which can detect when such issues may appear.
Q: Are there any ethical considerations or potential risks associated with using AI in the shipping industry? How can they be addressed?
For every decision support system based on AI there is always concern whether the recommended actions are biased due to the underlying models. In the era of deep learning, the biases can be driven directly from the data without it being obvious which parts of the data led to such biases – or how to correct them. At DeepSea, we try to alleviate such issues using a multi-faceted approach to improve a model’s interpretability, and to measure and guarantee models unbiasedness across the board. To learn how we boost model robustness, have a look at our paper on ‘How to judge the real-world effectiveness of AI in shipping’.