Firstly, we feed our machine learning model with as much information as we can from the electromagnetic radiation emitted by stars and the universe. We then get our disrupted data and we apply a fourier transformation so that to isolate the frequency of the data from that of the universe noise. Then our trained neural network performs statistical analysis on the disrupted object and finally it clears the object from white noise. Now, what happens if there is a gap of data at a certain area?
Here is how it goes:
We interpret the 3D objects as a combination of shapes and we feed them in the machine as a matrix input. These images are then used as samples to create a probability distribution, which will allow us to determine the best choice for filling any missing parts in the object. To fill these gaps, we need to gain as much information as we can from the context close to the missing area and also to develop an artificial perception of the image as a whole. We use an adversarial training method -also known as GAN (Generative Adversarial Networks)- where two functions will fight against each other to force the machine to create an artificial substitute that will be both plausible and natural to human understanding. The machine learning model has been trained before so that it can understand the features of some 3d printable objects, by thousands of 3d model objects and shapes, and it has created a distribution of reference. By scanning through a particular image, the machine can decide which features are missing and thus try to regenerate them, based on the new and the distribution of reference. Our deep convolutional neural network is finally able to reconstruct the objects by approximating the original object.