Input data types

RAMjET is focused on two main forms of input data: lightcurve data and raw flux pixel data. Lightcurves provide a simple, largely human interpretable form of input data. These lightcurves have been produced through some process from original flux pixel data. Often, the transformation process from flux pixel data to lightcurves discards valuable information, because the hope is to produce a human understandable lightcurve. Because of this, allowing the network to work with the more raw pixel flux data will allow the network to utilize more subtle and complex patterns discarded by the lightcurve production process. Throughout the rest of the documentation, we will often refer to a lightcurve or a series of flux frames as an “example”. This is because the neural network will use lightcurve or a series of flux frames as its input.

Output data types

We are interested in several types of data characterizations (labels). These labels are the information we want the network to be able to automatically predict. For example, we may wish to know whether a lightcurve contains a transit or microlensing event. Each of these would be a binary type label for each lightcurve. For the moment, RAMjET is focused on two types of output predictions, each of which can be applied to many problems. The first is a single binary label per example. That is, the entire input example (e.g., the entire lightcurve) has a single yes/no label, specifying if that example contains an instance of the phenomena we’re looking for. The second type of label we’re currently focused on is a binary label per time step. That is, each time step of the input example has a corresponding binary label. For instance, a lightcurve may have each time step labeled as to whether a transit event is occurring during that time step. While it’s more effort to produce this kind of label for use in training the network, it means the network will also produce this more meaningful label in its predictions, and, furthermore, it can make the training process easier as well.