Can recognition systems self-educate with minimal human intervention (“without a teacher”)? Today, almost all traffic projects use EDGE and server software trained utilizing “supervised learning” principles.
This requires a great deal of effort, time and expense in collecting and organizing access to datasets. In addition, software updates are required periodically due to the appearance of new vehicle types and models, as well as new types of license plates, etc.
New classes of problems and limited resources of EDGE solutions require a new approach in the creation of recognition systems. The market expects that the new approach will allow the system to learn independently, save time on software updates and improve their quality.
This is critical for the EDGE solutions market, since their task is to generate accurate data in large quantities, quickly, efficiently and at low costs.
We believe that such algorithms can be trained in a “no teacher” environment or with minimal human participation. A specially developed algorithm helps the system to identify objects (in our case, vehicles) and group them into clusters (classes) based on proximity metrics.
A person only names clusters (model, type, brand, color) and the system will automatically recognize new data.
Each mistake strengthens the “unsupervised” system because it uses the “memory power” rather than relying on a “teacher” (labeled dataset and human). After a while, the system will perfectly recognize and classify vehicles without a software manufacturer.
The integrator will not need to request a software update. When the system detects a new cluster, it will offer the integrator to check it and give a name, if necessary.
The new principle of creating recognition technology can be attributed to a self-developing system that is capable of embedding and coping with errors using the memory power.
Is the market ready for such changes? Let’s discuss this in the comments.