How does this work technically?
Feature transformation and feature extraction
In this step, some filters and pre-processing steps are applied to images. This allows neural networks to discover correct patterns faster. Trendskout applies different types of feature transformation and extraction, in interaction with the hyper-tuning step. For example, a photo can first be transformed via PCA to a lower dimensional structure, high contrast filters can be applied to it, or certain features can be pre-selected via feature extraction. This step is similar to the data processing applied to data with a lower dimensionality, but uses different techniques. As with classification, annotated data is also often required here, i.e. training data on which the system can learn which patterns, objects or images to recognize.
Selection of the algorithm and hypertuning
After the images, or other high-dimensional data, have been transformed via feature extraction and feature transformation, the selection of the algorithm is started. Trendskout performs this autonomously, just as with other AI and Deep Learning applications. Within the family of neural networks, the underlying algorithm, there are many different subtypes, each with their own set of parameters; number of neurons, number of layers of neurons, how the information between neurons is processed and many others. This means that an infinite number of configurations are possible, of which only a few give an optimal result, for example, able to correctly recognize an object in an image or data. Trendskout performs this hypertuning in interaction with feature transformation and feature extraction. In other words, through advanced Deep Learning techniques, Trendskout can find and implement the optimal data processing, feature extraction and transformation, and set up the best performing algorithm and associated parameter configuration. The combined process of data processing, algorithm selection and hyper-tuning provides an optimal model that can apply the desired image recognition or pattern detection to new data in the next step.
Real-time image and pattern recognition
When the training phase, the previous 2 steps, is completed, the model can be deployed in real-time on new data, for example new images in which patterns or objects must be recognized. The output of this step depends on the chosen Trendskout AI flow and can be: identification of the pattern found, a label, a similar image or other output.
The model is periodically re-evaluated and the entire process from the previous two steps is repeated in the background. If a more efficient model is found, the previous model is replaced.