Our Technology
OF THE EXISTING AND ITS LIMITS
AI model design, once under the control of Deeplearning experts, is reaching human limits in the absence of construction mathematics.For 30 years, the problem of the architecture of digital neural networks has not been resolved. The topologies that emerge are the result of test/error resulting from an implicit natural selection fed by experts in neural networks. As a result, the existing models are little or not optimized with an efficiency below the efforts necessary to obtain them. In addition, their disproportionate size makes them heavy consumers of computing power. Logically, companies, which cannot afford to have access to this power, are reduced to using the pre-wired (and therefore unoptimized) super models of AI providers who have supercomputers in their hands.
A new approach
The natural selection thanks to NNTO Technology
It is around these topology issues that the founders of DATAVALORIS turned to systemic natural selection for the construction of AI models like what nature has done for biological brains. This approach is the heart of the engine developed by DATAVALORIS called NNTO (Neural Network Topology Optimizer) and applies to any framework of digital neural networks.
Mutants AI!
NNTO acts as an overlay that drives the chosen Framework. It provides two services
- The generation service produces AIs from the catalog of HOF (Hall of Fame) parent models and applies to them mutations (hence the term Mutant). It builds them according to our rules for the appearance of mutations and your selection criteria (typology of objective or size for example).
- The selection service chooses, according to the scores that are reported and the constraints applied, the models that will join the parents' catalog, thus replacing the less suitable models.
Selection criteria
Among the selection criteria, the most important is the score reported by the Deeplearning Framework.
This score can be:
- A raw return of a Val_acc calculated by the framework
- A MAE calculated by the framework,
- The F1 calculated by the framework,
- A composite of the type [80%val_acc+20c*factor X], the factor X being able to be the result of the set of ethical data (0 or 1) or even a criterion of explicability of models (formal validation…)
Other criteria are taken into account by the selection service and weight the score:
- The size of the model (number of parameters),
- Its speed of execution (average time of an epoch),
- The difference in error compared to catalog models.
A simplified model design process
This approach implies a change in the way models are designed.
CURRENT DESIGN
DESIGN WITH DATAVALORIS
Thus, this new method of designing AI models takes into account business needs and production imperatives from the design process instead of a manual trial/error mechanism with validation of operational constraints at the time of implementation in production.
"From scratch" or "optimized" IA
The question that would remain to be defined would be at what stage the generation of the model should take place. Should we start from a model with an input and output layer or from a SOTA (state of the art) model?
Our technology excels in optimization, so we recommend starting from a model, if it exists, that has already proven itself.
RAISE, the automated platform for the automatic generation of AI models
From our NNTO engine, we have developed the RAISE platform which is a SAAS solution that allows any user to launch the generation or optimization of deep learning models.This solution is part of the experts' work process and is part of the expert's artillery. It is the missing brick of ML Ops on the AI modeling part (ModelOps)
a patented, secure and controlled process
A connector is installed on your calculation server as close as possible to your framework and pilots the latter to carry out the training and evaluation cycle.
This secure architecture does not require the company's data to be removed and makes it possible to control learning hardware costs.
MORE ROBUST MODELS
The competition of models and the selection generates by their design more robust models
A TOOL TO SUPPORT (RELEASE) DATA EXPERTS
As the process is automated, building the model no longer requires human intervention. The expert can therefore now focus solely on the selection criteria (score, weightings, etc.) and the data used.
A CONSTANTLY EVOLVING RANGE OF USE
Based on our technology, we have declined the uses to allow automatic assembly, evolutive learning on non-shareable distributed data, programming the appearance of mutations (phylogenetic replay learning)…
test now
the raise platform
for free!
RAISE Glossary
To use the platform, you must be familiar with the genetic algorithm concepts that we use:
Mutant :
A mutant is an AI model generated from the catalog to which mutations have been applied.
HOF size :
The HOF size corresponds to a number of “champion” models that are kept in the catalog of “breeders”. The higher this number, the greater the genetic diversity, the more the system is able to get out of an evolutionary impasse, but also the more evolution is slowed down. Conversely, a value of 1 implies that only a better child replaces the champion, so it goes quickly but the evolutionary impasse (local minimum) is probable. (The free version is limited to an HOF of 1). A value of 10 is often used.
Population :
The population is the number of mutants that are generated each cycle. The higher the number, the more mutants are generated and it is necessary to wait for the end of all the evaluations for the catalog update (HOF) to be done. This gives more choices but reduces a mutant's ability to enter the HOF. A value of 1 implies one mutant per cycle so the capacity for genetic diversity suffers because in probability only the champion of the catalog produces a child (bonus to the best). A value of 6 to 10 depending on computing capacity is recommended. It is set to 6 by default in free mode.
Cycle :
The cycle corresponds to the ongoing process of assessing the current population. It is not a parameter but information. For example, with a population of 6, the tenth cycle corresponds to the evaluation of mutants 55 to 60.
Generation :
The generation is an indicative value. The generation of a mutant corresponds to the number of ancestors it has since the initial model. So a generation 2 mutant has a father and a grandfather and the grandfather is the initial model. The max generation is the maximum generation of a mutant reached by the mutant generator for the considered project. The number of generations is not always correlated with its cycle of appearance. a son cycle d’apparition.
Growth control :
The Growth control is a weighting parameter for the selection of mutants and their entry into the HOF. The higher it is, the more preference is given to smaller mutants. A value of 0 disconnects this weighting (this does not necessarily imply a growth in the size of the model because mutations sometimes reduce the size), the maximum value on the other hand completely blocks the growth so the mutants which will be generated will be systematically smaller than his ancestors. Using the maximum value is only recommended for expected size optimization and may slow down the scaling process (useful for edge computing).
Project :
A project concerns an initial model and its entire evolutionary process. You can have multiple RAISE projects under your projects.