AI is changing our lives. Quite literally. Gen AI is being used in many areas that include generally repetitive, routine cognitive and analytical tasks, but also in some areas that were quite unexpected. It was unimaginable a few years back that creative work, for example, can be done by anyone else than humans, but it is not unthinkable any longer. More than that, we see ‘artworks’ such as poetry, stories, music, photography and videos compiled by Gen AI.
Whether they can be considered artworks or not, whether they are truly creative or mainly reshuffles of human-made art (many argue that these are merely a form of ‘algorithmic creativity’ that relies on existing data sets), or whether they are violating intellectual property rights or not – these questions remain and raise serious ethical and originality-related issues, which need to be addressed by the responsible authorities and creative professionals. In any case, regardless of the issues mentioned above, we started to talk about some ‘level’ of creativity of the AI systems, even if their creativity is different from human creativity.
Gen AI versus Human Intelligence
The topic of the creativity of AI systems is also interesting outside the artworld and creative industries. Creativity is not just about originality, it is also about discovery and experimentation. It means following our intuition, trying out new things, taking risks, making mistakes and learning from them. Thus, creativity is highly relevant in many areas of our lives. Machines seem to be incapable of human-like creativity as they lack both imagination and intuition as they are designed to optimize performance, thus their abilities to explore and innovate are quite limited. So, if limited by their nature, can AI systems ever mimic human creativity and intuition to perform accurately in real-life situations?
Even today, when Gen AI applications are largely spread and have proven that they are capable of some level of creativity, we still think that they are better than us only in some areas of analytical intelligence, especially the ones requiring ‘surface-level’ intelligence. They do perform better than humans when it comes to repetitive, routine cognitive and analytical tasks, can solve mathematical or logical problems of certain complexity, and can explain mostly such problems similarly to humans. However, they seem to be incapable of generating new concepts, for example, to ‘invent’ new algorithms or mathematical concepts. The problem of ‘algorithmic creativity’ mentioned earlier seems to persist here too.
At the moment, Gen AI is capable of repeating or combining existing solutions provided by humans, but it seems to lack the capacity to ‘innovate’ or come up with groundbreaking solutions even in the domain of analytical intelligence. Moreover, intuitive intelligence and empathetic intelligence are more characteristic to us, humans. We argue that we are better at thinking creatively, adjusting to new situations and performing risky or complex tasks that require experience and context-knowledge (all these skills adding up to intuitive intelligence). Moreover, we humans are better at recognizing, understanding, and responding appropriately to emotions (all these skills adding up to empathetic intelligence). So, how are machines performing in real-life situations, which are always rich in context, requiring constant adaptability and where decisions and behaviors are often emotion-fueled? And why is this important from the perspective of developing automated driving systems by Bosch?
The importance of contextual knowledge in real-life situations
Think of a situation when a car is driving in a residential area, within normal traffic conditions. You as a driver advance carefully, at low speed, towards your destination. Suddenly, a ball rolls onto your lane and it lands in front of your car. Your instant reaction is to break and/or steer away (if necessary) to avoid crushing the ball. Your behavior is guided also by the expectation that a child (or more children) could appear at any moment running after the ball, and you need to get prepared to avoid any accidents. Now, how would an automated driving system (driverless car) react in such a situation? Would it ‘assume’ or ‘expect’ that children might be playing with the ball, and they might run after it?
While human drivers can assess situations using their contextual knowledge, today’s automated driving systems are still learning how to do it. And in this process, latest Gen AI models might just be of great help. To match the humans’ capabilities, computers need to acquire an ability to make predictions about the perceived situations. We could see earlier, how relevant contextual knowledge is for assessing and understanding real-life scenarios. The strenghts of Gen AI is that it uses vast amounts of data, enabling them to draw conclusions from this data. Having access to and learn from extremely large amounts of data, the system can make predictions about similar situations to their training data-sets. Furthermore, in most training cases today, we keep human researchers/engineers in the loop in to give feedback to the machine on its predictions and adjust them continuously.
Gen AI and the future of automated driving
Applied to the field of AI, understanding context means not only processing large amounts of data and drawing appropriate conclusions based on them, but also ‚understanding’ the human context as well. Latest Gen AI models seem to become better and better at understanding context, making good predictions and drawing appropriate conclusions when trained with real data. By being exposed to real-world knowledge, they seem to become more context-aware and adaptable, thus effectively addressing a variety of use cases, including some scenarios that self-driving systems might face. For example, they can make appropriate predictions and give useful indications for the previous situation when the ball rolls in front of the car. However, for AI systems to understand context in all types of driving scenarios and address highly complex real-life situations is still under development.
In conclusion, there remain two huge challenges for current reserach & development to solve: on the one hand, despite the huge advances in Gen AI capabilities, they still fail when facing completely new, ambiguous situations or scenarios of unprecedented complexity. But for safety and reliable functioning, in the future, they must become able to learn to cope with unexpected and previously unknown situations and to act in highly sofisticated contexts as well. On the other hand, even if researchers succeed in training AI models to react with a high degree of accuracy, efficiency and reliability to all kinds of real-life driving situations, integrating such models with existing automated driving systems and embedding them into the car in order to have a system which senses, thinks and acts live in all kinds of real-life driving scenarios in a matter of seconds, is still a way to go.
Comments