The challenges of a deep-learning startup

When I founded my deep-learning company five years ago, the deep-learning landscape was still in flux. Innovative startups, like ours, were at the forefront of a technological revolution. Our goal was clear: leverage deep learning to tackle problems previously deemed impossible in the domain of high-tech systems. In our case, we found our niche in developing smart vision systems for the next-generation agri-robots that will revolutionize the agriculture sector. As this domain is still in its infancy, we extended this niche to applying deep learning to the more mature domain of manufacturing.

But as any founder of an innovative high-tech startup can attest, each forward step came with its inevitable challenges and stumbling blocks. Here, I reflect on the most significant hurdles I faced building a deep-learning company.

Data, data, and more data: effective deep learning requires vast datasets. In the agricultural sector, there’s a lot of variation between plants, greenhouse environments and lighting conditions. Gathering and annotating image data from greenhouses is a massive challenge and typically a job nobody likes to do as it’s ‘boring.’ Also, data collection can only be done during specific seasonal periods. Consequently, it demands good planning of your project.

Shifting technologies: the speed at which deep-learning technologies evolve can be both a blessing and a curse. While new tools and frameworks can make work easier, it also means constant re-education and the risk of months of effort suddenly becoming obsolete. We tackled this by developing a deep-learning operations platform that abstracts away the details of specific tools and frameworks. This way, when a new deep-learning framework becomes available, there’s a low cost for switching to this new framework.

The talent hunt: finding experienced talent specialized in deep learning is like searching for a needle in a haystack. Therefore, our strategy was to attract young AI talent and create the time and space for them to become talented application engineers. One part of this is creating a good mix of junior and senior-level engineers. Also, company culture plays an important role in keeping talent.

Real-world implementation: an algorithm that performs flawlessly in a controlled lab environment can falter in the field. Our agri-robots need to be resilient to variable conditions, from changing light conditions to unexpected obstacles. In the end, this just means a lot of testing is required. Our motto therefore has become: AI happens in the field, not behind your keyboard.

Financing and scalability: securing investments for a technology that promises a lot but is still in its infancy is an art in itself. To be independent of finding investments, we did both consultancy projects to generate income and product development to prepare for our next business stepping stone. Eventually, we found an investor that both provided financial security as well as opened doors to manufacturing customers.

The journey of a deep-learning company, like ours, is characterized by continuous adaptation, learning from failures and celebrating small victories. As we’ve advanced, we learned more and more about how to apply deep-learning technology within our specific domain.

Five years into our venture, I can affirm that deep-learning solutions are significantly impacting the agricultural sector. Numerous agri-robots and systems, powered by deep-learning algorithms, are now aiding farmers in enhancing their yield and optimizing resources. As we gaze into the future, we recognize that the technological landscape is rapidly evolving. With the advent of transformers, multi-modal models, foundational models and vector databases, there’s a wealth of opportunities awaiting exploration.