Retalon has developed a new process that enables NN to start training on the trustful data first, slowly learning from less perfect data. The new approach also impacts the initialization step.
FREMONT, CA: Retalon, an award-winning AI & Predictive Analytics software provider, has announced a new approach to the training of Neural Nets, named Progressive Learning.
Traditionally data used to train neural net models is less than perfect, leading to errors and biases. This problem is not very new, as bad data has always been an issue, irrespective of what technology was used. With Neural Nets, however, the idea was that the Neural Nets would at least somewhat replace data scientists in the data preparation process.
Retalon has developed a new process that enables NN to start training on the trustful data first, slowly learning from less perfect data. The new approach also impacts the initialization step. Retalon Progressive Learning technology provides one consistent gradual approach to initialize as well as train NN for business-specific application.
"You don't show a 2-year-old the "Game of Thrones" to educate them on how this world works. You start with "Sesame Street", and then add complexity to the established foundation. We found that this is also an important step in Deep Learning of artificial systems. At the end of the day, all systems (whether human or artificial Neural Net) are based on the same principles. Progressive Learning technology bridges this gap in the process of initialization and training of Neural Nets for business-specific applications", stated Mark Krupnik, Ph.D., CEO at Retalon Inc.
The Retalon Progressive Learning approach provides companies with the advantage of training models that will be more suitable for their business process without overfitting. The approach has already shown some noticeable improvement in quality and stability of results in situations with missing or wrong labels, incomplete data, as well as the presence of outliers. Retalon's new Progressive Learning technology automatically finds out at least 80 percent of anomalies in data and requires much less data scientist intervention, resources, and time.