Why image recognition does not have the industrial success it deserves?
Deep Learning and its application to computer vision has attracted a lot of attention in recent years, due to the wide range of tasks to which it applies and the level of performance achieved. However, despite all its promises, it is difficult to find examples of large-scale industrialization of these systems…
And yet, after having helped many clients automate their visual tasks on an industrial scale, we have seen how deep learning promises can be delivered and massively unlock business value. And just because it’s you, we’ll share our secrets.
An outdated academic approach
We have found that there is one main reason behind the lack of significant industrial achievements powered by modern computer vision. This reason is that the academic, theoretical approach taken by most industry actors to building and industrializing computer vision projects is ill-adapted to business. As it turns out, it is very close to software development methodologies – such as the waterfall model – used in the first days of software development, and which have been replaced in the mid-90s.
“Since the emergence of deep learning use in business, the academic approach of developing AI has prevailed. By focusing on finding the best model on fixed hypotheses, the resulting systems became abstracted from the real-world problems and failed to perform in production.”
Aloïs Brunel, CPO at Deepomatic
Indeed, here are the 5 main reasons why the academic approach is fundamentally flawed in a business context:
- The conditions of production are difficult to anticipate and evolve over time because of operational circumstances.
- The data set used to train the algorithms does not have to be fixed.
- Developing in-house neural network architectures is time-consuming and expensive, with little guarantee of improving performance.
- It takes a lot of time to go from the design phase to the production phase, which makes it difficult to detect and fix any problem.
- For high-impact AI projects, it is often required to deploy the AI systems at the edge, which implies spending more time on setting up an IoT infrastructure.
Our experience has repeatedly shown that going from a linear development model to an iterative and agile one has the capacity to unlock high-impact applications.
The LEAN AI software revolution: a method tailored for real-world business contexts
So, we have created a new framework inspired by recent trends in Agile, Lean and Devops, called Lean AI. The goal of this methodology is to minimize the risk of failure in the field of industrial AI, as well as to maximize the performance of AI systems.
In the context of AI systems, the final product is generally already well understood: we know how it is going to be used in production and the business value it is meant to generate. The uncertainty lies on how to produce the right AI that implements that product, and the uncertain conditions of production in which the system is going to operate.
Lean AI will save you some precious time by both optimizing the product building process and navigating the uncertainty of production conditions. And that’s the key. To do so, we have theorized the 3 main ingredients of Lean AI, that are “The Minimum Viable AI”, “The Lean AI Loop”, and “Short cycle implementation”.
Each ingredient is fully detailed in our white paper – please feel free to go and check it out if you’d like to learn more about the Lean AI methodology and tools, as well as its practical implications, challenges, and means to overcome them.
Lean AI is the methodology you were looking for in order to deploy your AIs at industrial scale. We are convinced it has the potential to unleash the true potential of AI for businesses. What about you?
