The scientific method: AI’s value added is the unknown

As the most advanced form of Artificial Intelligence, Deep Learning is proof that new opportunities can be created by people who trust in the power of experimentation. By putting hypotheses into practice, data science turns the scientific method into fertile ground for producing the technologies of tomorrow.

Gaining understanding from experiments

The journey from a black box to the development of Artificial Intelligence models that are explainable and reproducible is more of a long and winding road, rather than an easy single step. With AI, you cannot understand everything, or explain everything. So, using the scientific method is very important. And it's essential when it comes to understanding the results – a complex task for the data scientist, but a simple one for a machine.

The scientific method is a process based on experimentation. In AI, it starts from a basic assertion, which could be a malfunction, an unexplained result or a model that performs better than expected… What follows is a series of hypotheses that seek to explain the result. Which of the variables was decisive in producing that result? By testing these different hypotheses, and adjusting one variable at a time, experimentation can identify some of the factors and explain the final outcome. But before any particular hypothesis can be accepted, many others have to be dismissed, turning the whole exercise into something of a scientific marathon. If a hypothesis seems to make sense and can be verified, it will be repeated until its validity can be absolutely confirmed. Clearly, the method is far from being the kind of predictable process usually associated with software development, for example, where the end result is known – along with the way to achieve it. There is no straight line when it comes to understanding AI. The possibilities are numerous and provide the opportunities of the future… provided the method is applied rigorously. 

Finding the decisive variable

Without the use of the scientific method, AI becomes something approximate, nebulous and even dangerous. Imagine the predictions that might be made about a particular market, based on a model that had not been rigorously validated by the scientific method. The model would appear to be efficient and successful, but in truth, the expert behind it would have only partly understood what was happening, and would be unaware of the impact being made by the most influential variables. Chance would then play its part, making the model unstable and unpredictable. On a large scale, the damage caused could be substantial.

For a practical example, take the development of car batteries. Imagine one designed for use in Europe – how effective would it be in a hotter, colder or more humid region? Should we expect it to overheat? Would drivers be haunted by fears of being unable to start the engine? For a car battery, the external temperature is actually the crucial variable for its performance. If we apply this to AI, the scientific method provides a way of identifying the crucial variable in each model. It accumulates knowledge about the AI’s black box, and ensures the reproducibility of the results and guarantees their reliability.

Going further, to create value

Opting for the scientific method also provides a guarantee that an AI system will operate transparently, and therefore more ethically. By ethically, we mean there is certainty about both its performance and its reliability. This takes time to develop, but it provides a more dependable outcome. For AI, the unknown is a source of value added, by offering models that are more powerful than a basic software package. It encourages constant innovation.

At Homiwoo, we are very much part of this long-term scientific approach, this search for viable, continuous innovation. We guarantee our clients that they will receive new models, an approach that will make a difference when it comes to understanding the finer details of the property market. We guarantee the provision of reliable, long-lasting models. Our commitment to excellence by constantly questioning what we are doing recently brought us success in the Deeptech category of the 2020 i-Nov Innovation Competition organised by BPI France.

We are proud of this recognition of our company culture by our peers. Our aim is to be the pioneers of the new application of AI. To be more innovative, but also more rigorous. 

William Violet, CEO, and Adrien Bernhardt, CTO of Homiwoo