DeepSensor research ideas

Are you interested in using DeepSensor for a research project? Thankfully there are many interesting open questions with ConvNPs and their application to environmental science. Below are a non-exhaustive selection of research ideas that you could explore. It would be helpful to ensure you are familiar with the literature and resources in the Resources page before starting.

Why not join our Slack channel and start a conversation around these ideas or your own? :-) You can join by signing up for the Turing Environment & Sustainability stakeholder community. The form includes a question on signing up for the Slack team, where you can find DeepSensor’s channel.

Transfer learning from regions of dense observations to regions of sparse observations

Since the ConvNP is a data-hungry model, it does not perform well if only trained on a small number of observations, which presents a challenge for modelling variables that are poorly observed. But what if a particular variable is well observed in one region and poorly observed in another? Can we train a model on a region of dense observations and then transfer the model to a region of sparse observations? Does the performance improve?

Sensor placement for forecasting

Previous active learning research with ConvNPs has only considered sensor placement for interpolation. Do the sensor placements change when the model is trained for forecasting?

See, e.g., Section 4.2.1 of Environmental sensor placement with convolutional Gaussian neural processes.

U-Net architectural changes

The ConvNP currently uses a vanilla U-Net architecture. Do any architectural changes improve performance, such as batch normalisation or dropout?

This would require digging into the neuralprocesses.construct_convgnp method and replacing the U-Net module with a custom one.

Extension to continuous time observations

The ConvNP currently assumes that the observations are on a regular time grid. How can we extend this to continuous time observations, where the observations are not necessarily on a regular time grid? Can we do this without a major rework of the code and model? For example, can we pass a ‘time of observation’ auxiliary input to the model? What are the limitations of this approach?

Training with ablations for interpretability

Since the ConvNP operates on sets of observations, it is possible to ablate observations and see how the model’s predictions change. Thus, the ConvNP admits unique interpretability opportunities.

However, the model would need to be trained with examples of ablated observations so that it is not out of distribution when it sees ablated observations at test time. For example, when generating Tasks with a TaskLoader, randomly set some of the context_sampling entries to 0 to remove all observations for those context sets. Then, at test time, ablate context sets and measure the change in the model’s predictions or performance.

Monte Carlo sensor placement using AR sampling

The GreedyAlgorithm for sensor placement currently uses the model’s mean prediction to infill missing observations at query sites. However, one could also draw multiple AR samples from the model to perform Monte Carlo sampling over the acquisition function.

How does this change the sensor placements and what benefits does it yield? Do the acquisition functions become more robust (e.g. correlate better with true performance gains)?

The Environmental sensor placement with convolutional Gaussian neural processes paper will be important background reading for this.