For 30 years, neural network architecture has remained an unsolved problem, handled by trial and error. DataValoris applies natural selection to AI model design.
DataValoris chose systemic natural selection to build AI models, mirroring what nature did for biological brains. This is the core of our NNTO engine (Neural Network Topology Optimizer), which works with any neural network framework. NNTO belongs to the NAS family (Neural Architecture Search) — the automation of architecture design, seen as the next step beyond AutoML — but stands apart by searching topologies through evolution rather than exploring predefined configurations.
NNTO acts as a layer that drives the chosen framework. This engine powers our four answers — model obfuscation, bias counterfight, data drift adaptation and multi-objective optimization — through two services:
Produces AIs from the catalog of parent models, including from an existing or open-source base, then applies mutations according to our rules and your selection criteria (objective type, size and more). The point is not merely to obscure a model: it is to differentiate it, optimize it, make it more robust and adapt it to dataset changes. The resulting topology is dedicated to your problem and remains unknown to the competition, making it hard to reverse-engineer or copy.
Chooses, based on reported scores and your constraints, which models join the parent catalog, replacing the least suitable ones.
RAISE creates new architectures from the most promising models. Beyond simple obfuscation, mutations shape a topology dedicated to your problem, more robust — and unknown to the competition. The resulting model becomes a proprietary asset that is hard to reverse-engineer or replicate.
Each model is evaluated against your objectives: score, size, execution speed, robustness, biases to avoid, production constraints or adaptation to dataset changes.
The best models join the parent catalog and replace the least suitable architectures. Selection relies first on the score reported by the Deep Learning framework (val_acc, MAE, F1 or a composite), weighted by model size (number of parameters), execution speed and the error difference compared with the other catalog models.
A concrete example of evolution: on the left, a minimal initial model; on the right, the mutant champion obtained after several generations of mutation, competition and selection. The topology has grown connections and structures that nobody drew by hand — and that no competitor knows.
AutoML benchmark
Building a high-performing AI faces a near-infinite space of topologies and hyperparameters. Our AutoML benchmark shows that RAISE's evolutionary approach — rooted in phylogenetics and natural selection — ranks among the best AutoML tools on the market, outperforming reference approaches, including XGBoost. “The results of our evaluation campaign show that the mutants created on RAISE perform significantly better than the original model.” — a French transportation company, object recognition on conveyor.