Bayes Centre

Neurolabs Case Study

A case study from our Bayes Member, Neurolabs.

Quick overview of your company

Neurolabs builds object recognition algorithms based solely on synthetically generated data. This makes us faster, cheaper and better than our competition: using pixel perfect annotations for algorithm training allows us to reach beyond human level accuracy. Foodservice is our beachmarket, aiming to extend into quality control  in manufacturing, and healthcare. We deploy our solution through software integrators, whilst partnering with hardware providers.

What problems are you trying to solve, any examples of progress to date?

It is commonly understood that the barrier to entry for leveraging machine learning is very real, largely due to the amount of data needed to kickstart the process, e.g. hand-labelled images for object recognition. With advances in computer graphics, it is easy to generate an image that is indistinguishable from reality, and because all of the depicted information is intrinsically understood by the computer, we can automate and massively accelerate the human-driven laborious task of annotating images. Producing this synthetic data is simple and scalable, however, minor differences between synthetic and real images reduce the quality of the deep learning model. Reducing this so-called domain gap is at the forefront of Neurolabs's research efforts. At the same time, we are producing and deploying functioning object recognition models now, as assisted checkout solution in cafeterias and starting soon in manufacturing.

What are your thoughts about the company's vision and direction, and your role in helping achieve them?

Our vision is to enable every industry to take advantage of object recognition by providing them with a platform that takes care of both generating the synthetic data and producing the machine learning model. Our current direction is more localized into validating our hypothesis that synthetic data for object recognition can revolutionize all industries, meaning that we are focusing our efforts on building bespoke solutions for different industry verticals. My role is to oversee the production of the synthetic data and digital asset procurement, which lands me right in the middle of creating or acquiring 3d models and providing the ML team with the data that they need.

What are the biggest challenges on the horizon?

Our challenges are placed in one of two categories: fundamental and auxiliary. The former focuses on answering the questions that are currently being tackled through research in academia and industry for which we don't currently have the answers: machine learning models that can recognize 10,000 different classes and bridging the aforementioned domain gap, to name a few. The latter focuses on the support that Neurolabs provides to answering the fundamental questions and their consequent industrialization. This boils down to our resource pool and knowing how to allocate it ideally.

What are your goals for the next 6-12 months? When you tell people about your job and company, what is the thing that surprises them, or gets them excited about the work you do?

Our goals for the next year comprise of productizing and commercializing our object recognition solution in three separate industries. While our long-term goal for providing our platform is to stop at creating the machine learning models, many industries currently do not have the technical know-how to embed such a model into their product offering. We believe that as the benefits of object recognition become more apparent in all industries, so will the shift in machine learning competence across every domain.

Many people are aware of advances in artificial intelligence and machine learning and their applicability to our life, however, a lot do not know just how much data is required by today's data-hungry models. The seemingly closed circuit of using a computer to generate images that will in turn be used to train a computer to recognize images is intriguing to a lot of people as well.

Is there a way people can get involved with your work?

For now our focus is to work with software integrators that work for a number of clients in a given industry. If you are such a company and know that your clients would like to take advantage of object-recognition, please get in touch!

 

-Markus Schläfli, Head of Simulation