A Q&A with Konstantinos Kyriakopoulos, CEO and Co-founder of DeepSea Technologies
There’s been a lot of news recently about “sentient” artificial intelligence. Will DeepSea’s AI be sentient any time soon?
Sentience is the ability to experience emotions and sensations. Machine learning algorithms are trained on data – pairs of inputs and outputs – and they learn how to reconstruct the outputs from inputs by themselves. A lot of AI nowadays is trained on inputs and outputs that come from humans – for example, language models – which are given large amounts of text and learn how to reproduce those conversations. You can even hide pieces of the conversation, and the AI will eventually learn to fill in the gaps. If you start talking to it – or asking it questions – the AI will then figure out what is statistically the most likely next phrase that a human would say or answer.
When you start talking about AI that is trained on emotion recognition and emotional language, emotional pictures and so on – that’s where it starts to get somewhat philosophical.
When a machine starts to process internal representations in a similar way to how a human “feels” something, at what point can you say that the algorithm itself is truly experiencing the sensation in question?
That is what the conversation about sentient AI is about. It is primarily a philosophical conversation, but it is also a great opportunity to discuss how these neural networks work from the inside.
DeepSea’s AI is not trained on humans. It is trained on ships, on ship engines and on the environment in which the ship moves, it is trained on the sea. It obtains the properties of these elements rather than the properties of humans. So, sentience is not relevant in our case.
What is the coolest thing about AI?
The coolest thing about the application of AI and deep learning is the sheer flexibility that it has.
You can train these systems to do essentially anything you want.
If you start off with any problem, you can always come up with a way to give the AI the right data and structure to learn how to solve it. This opens up an enormous number of opportunities in all sorts of different fields.
One of the most exciting problems that we have worked on at DeepSea is how to figure out the level of fouling of a vessel, even though this is not measured. Something very real has happened to the vessel – barnacles are becoming stuck to it. It is impacting the performance of the vessel – but it cannot be measured directly. With intelligent AI, this becomes something we can understand quickly and accurately.
So what does DeepSea’s AI learn about my ship?
One great thing about AI is that the same techniques and technologies can be applied to radically different fields. One such technique which we have applied very successfully at DeepSea is transfer learning – which essentially means having one vessel learn from another.
At DeepSea, we have been able to apply transfer learning so we can deploy our optimisation solutions on ships where we have little, or poor-quality, data – for example data that comes from daily reports rather than high-frequency sensors. Every vessel has some data available – AIS, vessel particulars, daily report data – but that is not enough to train AI from scratch. However, what we have found is that if you have a vessel-independent model that learns the key characteristics about motion, dynamics and physics that are common to all vessels, you can use the same technique of transfer learning to take that model and transfer it to a new vessel.
How is DeepSea’s AI different from that of other companies?
AI is a relative newcomer to shipping, and there has been a lot of critical foundational work to do – DeepSea has assembled the leading AI team in the industry to do it.
Our research focuses on a number of vital areas. Perhaps most important of all in driving real impact for our customers is understanding causality – i.e. what factor affects what. What this means is that at DeepSea we can predict how much fuel your vessel will consume in any arbitrary conditions (eg. the weather, the speed, the draft and so on) as each of them varies over the course of a voyage. This factor-specific prediction is vital for figuring out what speed and what route your vessel should take in order to minimise energy requirement and therefore fuel consumption.
This is an area where most people who are trying to do modelling of any kind – especially ship modelling – fail. When trying to model a ship, there are all sorts of different spurious correlations we have to contend with. A less sophisticated machine learning model will not necessarily determine correctly which part of the effect is due to which factor.
As far as I am aware, we are the only organisation that has effectively solved this critical problem. What does this actually mean for our customers? Well, typically, these sophisticated models drive our Performance Routing platform, and directly translate into an average 8% fuel saving.
These models are the difference between a standard weather routing approach – which often leads to even greater fuel consumption than sailing without it – and a true optimisation of a voyage.
What made you start DeepSea?
My background is in deep learning from a research perspective. My research at the University of Cambridge focused on language and speech. The idea for DeepSea came about in a conversation that I had with Roberto Coustas, the other founder of the company. I was looking for applications of Al that differ from the ‘standard’, well-worn paths. Roberto, coming from a shipping perspective and having studied AI himself, was aware of a key problem in the industry that seemed insoluble – truly being able to predict ship behaviour.
Shipping companies struggle to do things like fouling detection or true voyage optimisation because they do not have the ability to accurately predict, in different conditions (and especially in conditions of extreme weather), what the precise fuel consumption would be. As a result, all the decisions they take lack a very important piece of information – what is the predicted impact? If we can predict this accurately, a shipping organisation can coordinate far better to minimise its fuel expenditure – and therefore its emissions and environmental impact. Without this ability, there is huge waste. By applying deep learning, we felt we could actually solve this major problem.
Is AI the answer to everything?
Sadly not! AI is a specific technology that solves specific types of problems – particularly problems that have to do with lack of predictability.
AI has to work side-by-side with humans. There are things that only humans can do – for example, captaining the vessel, making day-to-day safety decisions, inspecting the engine, etc. There are many things the AI does not know – issues with traffic, complicated port constraints, geopolitical factors.
The task, for us and for modern shipping companies, is to take people that have been used to working in one way and support them to work in a different way – a way where they can exploit the powers of AI within the context of their own skills, goals and insights.
Should I trust this ‘black box’?
Ah the old ‘black box’. Yes, it’s true – this is one of the major challenges of Al (actually not just for users – it’s equally frustrating for researchers and those working in the field!).
The problem is that, if you want humans to engage with this technology interactively, the AI needs to be as transparent as possible about why it is making the suggestions that it is making. What we are working toward is moving from the ‘black box’ to the ‘translucent box’ – so you can start seeing what is happening inside.
On the optimisation side, we have a tool that makes available the impact of different factors in the equation. For example, it can explain exactly how much you are saving when you make a certain change – why a certain route is more optimal than another – and so forth. Across the entire AI world, this area of ‘explainability’ is really a work in progress and the subject of a lot of academic research.
But I think you can also trust something you don’t fully understand, as long as there is evidence that it works, and there is sufficient support for the approach. Right now, we are involved in an important initiative with a number of leading universities, called the Shifts Project, which is leading the way globally in finding new ways to tackle some of the greatest challenges to the real-world application of AI. I’d like to think that the backing of universities such as Cambridge and Johns Hopkins, alongside our ever-growing list of clients, should be enough to get hardened AI-sceptics to engage us in conversation.
The world’s most advanced shipping companies are starting to adopt AI at a rapid pace. To understand why – and what you could achieve with AI – download our white paper AI: Harnessing decarbonisation’s secret weapon now: