ADAGOS: Simplifying AI while Maximising Performance
Mohamed Masmoudi, CEO
Straight from its fictitious applications in fantasy books and films, artificial intelligence (AI) has become an integral part of business operations in the current socio-economic setting, utilised by enterprises in all industries across the globe. The inner workings of current AI programs involve a colossal number of computations, with millions of digital neurons creating numerous data-packed layers to convert inputs into valuable outputs.
Following the development of AI so far, there have been two generations of neural network programming—shallow, then deep, learning—that have paved the way for an AI program to adapt and better perform its preordained operations. However, while second generation neural network manages to simplify and streamline existing workflows, the immensity of deep neural networks poses a significant problem for data scientists. The continually amassing big data has allowed the first two generations to focus on oversized structures where redundancies are necessary, resulting in poorly designed and inefficient neural networks. "Current algorithms can work quite efficiently when the amount of learning data is huge compared to the complexity of the problem, but they are still far from being optimal," asserts Mohamed Masmoudi, a well-established mathematician and industry veteran. Witnessing the need to overcome this obstacle, he founded ADAGOS, a spinoff of the Institute of Mathematics (IMT), Toulouse, France. As ADAGOS’ CEO, Masmoudi has dedicated his efforts toward establishing a more parsimonious approach to building neural networks. The CEO adds, "We have developed NeurEco, a third generation neural network solution that automatically generates intricate yet minimal network structures, while drastically reducing the demand for resources."
NeurEco aims to make AI more accessible to users, offering them a tool for automatically constructing their neural networks.
We have developed NeurEco, a third generation neural network solution that automatically generates intricate yet minimal network structures, while drastically reducing the demand for resources
"There is no prerequisite for the users to have any neural network knowledge. They simply have to provide their data and NeurEco will handle the rest," expresses Masmoudi. The company's parsimonious NeurEco solution outperforms current deep learning methods, especially when the response of a model is continuous or deals with long term, dynamic predictions. It overcomes most of the drawbacks of existing neural network algorithms by increasing precision and robustness, while drastically reducing their computational cost, energy consumption, and associated carbon footprint. NeurEco's ability to deliver efficient results with only a limited amount of learning data further promotes the efficacy of its parsimonious approach. "The number of network parameters to be identified by the learning process is reduced, but the structure of the network is enriched," says Masmoudi. The connections are more intricate than in common layered networks; NeurEco automatically determines the optimal topology of the network using the company's unique optimisation process. Unlike other attempts to prune and simplify redundant neural networks, ADAGOS’ approach achieves an inherent parsimony, without the need to first develop an oversized network. Such features are a revolutionary advancement for AI applications in the areas of IoT, embedded systems, medical and defence systems, and industrial data management.
To further elucidate on NeurEco's capabilities, Masmoudi shares an anecdote of the time NeurEco helped ADAGOS win the grand prize of the CONTINENTAL Start-Up Challenge in 2019. "We used our solution to reduce the size of the client's artificial neural networks, allowing them to eliminate the complexity of programs and embed it in their cars of the future," he adds. ADAGOS helped CONTINENTAL secure an optimised model that gave them a new perspective and understanding of how differently AI can be utilised in their automotive computers.
Another promising area of application for the company's intuitive NeurEco is across energy regimes. In its ongoing partnership with Meteo*Swift, an AI solutions provider for wind power-based organisations, ADAGOS' parsimony has helped create a new level of convenience for AI operability throughout the organisation. NeurEco aided the client's programmers by modifying their coding practices and generating wind prediction reports that were more accurate, thereby enhancing power generation processes for Meteo*Swift's customers.
Owing to the vast and continually growing nature of AI and big data, ADAGOS has also published several scientific papers based on NeurEco's parsimonious modelling that circumvents the necessity for big data-riddled networking. One of the company's papers—titled Parsimonious Neural Networks—published in the frame of the European Cyber Week, elaborates on NeurEco's ground-breaking capabilities and also brings to light its security features. "NeurEco not only gives robust predictions, but can also effectively fend off DeepFool attacks," adds Masmoudi.
Moving forward, the company intends to capitalise on its collaboration with leading IoT companies in order to deliver solutions that reduce the quantity of rare earth elements used in the production of various connected devices. According to the Global System for Mobile Communications Association (GSMA), the IoT market will increase from 6.3 to 25.1 billion devices in 2025. "Our goal is to reduce the energy consumption of these devices, thus improving their autonomy, while lightening their burden on the planet," expresses Masmoudi. Additionally, ADAGOS plans to make AI an affordable commodity to small-medium businesses (SMBs). With the official launch of NeurEco in the first quarter of 2020, the company is set to revolutionise the way unsupervised learning algorithms are architected, increasing predictive analytic capabilities for various technologies across several industries.
This content is copyright protected
However, if you would like to share the information in this article, you may use the link below: