Neuton takes a different approach to building perceptron neural networks. Instead of using a fixed architecture or enumerating candidate structures like most NAS methods, Neuton grows the network neuron by neuron. The minimum structural unit in the optimization process is the neuron's input, which reduces the granularity of the architecture search and produces compact networks without sacrificing accuracy.
Neuton uses a patented global optimization algorithm to identify network parameters, which avoids the local extrema and plateau problems associated with gradient descent. The algorithm also supports parallelization across multiple hosts and GPUs without loss of accuracy, making cross-validation feasible within acceptable training times.
During training, the network grows automatically with built-in overfitting control. Neurons are added until the network reaches its maximum generalization ability. The global optimization algorithm keeps each neuron efficient, which limits overall network size while maintaining accuracy.
Note
The only required inputs are the data, a target variable, and a metric. Training, validation, and model selection all happen automatically.
The key differences between Neuton and traditional neural network frameworks are:
| Criteria | Neuton | NAS |
|---|---|---|
| Architecture definition | Automatic neuron-by-neuron growth | Automated search over candidate structures |
| Moment of size optimization | During training | Before or after training |
| Optimization algorithm | Proprietary (global optimization) | Gradient descent |
| Accuracy loss from size reduction | No | Possible |
| Manual hyperparameter tuning | Not required | Required |
| Built-in overfitting control | Yes | No |
| Training cost | Free (through Edge AI Lab) | Very high (multiple full trainings) |