Newsletter
Magazine Store

50 Best Companies to Watch 2020

Harness the features of parsimonious neural networks to proliferate in AI application: ADAGOS

thesiliconreview-mohamed-masmoudi-ceo-adagos-20.jpg

The Artificial Neural Network (ANN) is basically a series of algorithms that endeavors to identify underlying relationships in data through a process that is similar to the way human brain functions. ANN can adapt and change according to inputs and generate the best possible result without the need for redesign. The concept of ANN is deeply rooted in artificial intelligence, and it is quickly gaining popularity. An ANN contains layers of interconnected nodes. ANN are primarily used in applications for financial operations, enterprise planning, business analytics, and product maintenance.

To address the challenges in the implementation of machine learning, ADAGOS has developed NeurEco. A new neural network approach based on parsimony. NeurEco reduces the resources (size of learning data, energy consumption, size of neural network and memory requirement, computing time, development time) required to implement machine learning methods by several orders of magnitude when compared with the current state of the art. Their first significant fields of application are embedded systems and the Internet of Things. By achieving the smallest sufficient network structure, they minimize the network’s battery consumption and make it possible to embed AI on small devices and to increase their autonomy. Thanks to its parsimony, NeurEco requires less data for learning. It is even possible to embed the learning process and to update the neural network on the fly, avoiding the need for waiting weeks and months to collect enough data. Their second major field of application is healthcare. Thanks to parsimony, any prediction given by ADAGOS’s neural networks comes with a clear explanation. For example, if a patient is declared positive for a disease, the neural network response is supported by an explanation of which features of the input data led to this conclusion. Due to their redundancy, conventional neural networks cannot give such a clear explanation. In the case of COVID-19, which ADAGOS is currently focusing on, this explanation would help elucidate the mechanisms of the disease.

In conversation with Mohamed MASMOUDI, CEO of ADAGOS

Q. Explain your company’s successful journey in a short story.

Our world has entered an era where technological progress can no longer depend on the exploitation of natural resources, and it has become essential that new advances not only improve upon the existing technology but also reduce their energy consumption. The art of artificial intelligence is mostly inspired by the biological brain, including its redundant nature. While this redundancy may ensure the continued functionality of the living brain despite regular, and sometimes accidental, loss of the neural cells, the same argument does not hold for artificial neural networks, which are made from inert matter.

We have realized that it is possible to do much better. The founding team of ADAGOS was well known in academic circles for their contribution to topological optimization techniques. These techniques have primarily been developed to optimize the design of structures like bridges, fluid transport networks, printed circuit boards and microstrip antennas, and MEMS (Microelectromechanical systems). A major issue in topological optimization is deciding which element to add or to remove from the structure. We introduced the novel idea of using topological optimization criteria for designing neural networks; this provided a method for selecting the neurons and links that should be added to achieve the optimal network performance. This approach achieves native parsimony, present from the network’s conception, and drastically outperforms existing pruning methods for simplifying the structure of oversized neural networks. The resulting parsimonious neural network is a powerful tool for solving large and complex problems while using fewer computational and data resources - by several orders of magnitude - than classical neural networks. Also, the method allows for the construction of networks to be achieved by an automatic algorithm; no prerequisite knowledge in machine learning is required, and a user needs only to provide their data. Thus, we can spare not only material resources but also the hours and hours of tedious development spent on the creation of classical neural networks.

Q. What motivated you to venture into such an advanced yet sophisticated field of technology?

We drew our inspiration from Occam’s Razor, a fundamental scientific principle that dates back to the 14th century; Einstein once elegantly phrased it as: “Everything should be made as simple as possible, but no simpler.” The immensity of neural networks obtained using classical AI tools is in complete contradiction with this principle of parsimony. We were certain that it was possible to achieve superior performance with smaller networks that require fewer resources.

Q. What kind of problems do neural networks and deep learning work well for?

Conventional AI algorithms are mainly oriented to qualitative zero-one responses like classification, pattern recognition, and natural language processing. NeurEco outperforms conventional methods, particularly when the response of the model is continuous (quantitative), as a dynamic prediction. Most of the needs in the fields of IoT, health applications, energy, and the environment are related to such continuous responses.

Q. Implementing Machine Learning doesn’t guarantee success. Experimentations need to be done if one idea is not working. How do you make sure your time, effort, and money are used efficiently and are not spent on unsuccessful projects?

An idea is not inherently good or bad. If you are convinced that it is a good idea, with a little determination, you will finally find a way to make it work. Before starting a project, we discuss its chances of success, and we do not start until we have reached a minimum consensus on it. This reduces the risk of abandoning a project.

Q. What are your plans for the future development of your company?

Thanks to our collaboration with a leading player in the field of IoT, it appears clearly that our solution is particularly suitable for IoT and embedded systems. According to GSMA, the market of IoT will increase from 6.3 to 25.1 billion devices in 2025. Our goal is to enable the simplification of these devices by providing resource-sparing solutions and reducing the number of rare earth elements they use and increasing their autonomy. Another goal of ours is to make our tool popular for health applications. We plan to improve upon existing decision support tools, which have been developed with classical AI approaches. NeurEco’s parsimonious networks will offer clinicians more robust evidence to support them in their clinical decision making.

Meet the leader behind the success of ADAGOS

Mohamed MASMOUDI is the CEO of ADAGOS. He is also a professor at the Institute of Mathematics, Toulouse, France. Mohamed dedicated much of his career to sharing the power of mathematics with industry.

“NeurEco drastically reduces resources required for implementing AI.”

NOMINATE YOUR COMPANY NOW AND GET 10% OFF