# Proposals

List of the proposed projects

**CASF: Comparative assessment of scoring functions for docking problem**

Pharmacological research is concentrated on constructing a molecule, a ligand, which docks a given protein. This research is expensive since it goes in-vitro. We have to propose a deep neural net to forecast the probability of docking, see details.

**3D-image reconstruction for the lens-free microscopy**

One has to reconstruct an image of a small biological object (of hundreds micrometers, like a eucaryotic cell). Its video comes from a device of a brand-new technology: the free lens microscopy. This project combines modelling of interference images and its Fourier transform and GANs, the generative and discriminative deep learning models see details.

**Atom-resolution synchrotron image reconstruction**

For a given synchrotron image one has to reconstruct an object of nano-meter size. There given a reciprocal lattice image. We have to train a neural net to reconstruct the real and the complex part of a sample object, see details.

**Long-live health monitoring with wearable devices**

For daily accelerometer and gyroscope time series we have to reconstruct typical hand movements and represent these movements as clusters in the phase space.

**Time series segmentation in low-dimensional space**

One has to mark-up zero-phase segments of various hand movements using electronic watches. Neural nets on spherical harmonics seems to be a good tool.

**Invariants and compositions of physical activities**

We have to create a model to describe simple hand movements and compositions of movements. The composition of physical models must fit the composition of machine learning models.

**Generating WANN: weighted agnostic neural networks**

We have to generate agnostic nets and ensembles of agnostic nets to solve one of the well-known reinforcement learning problems: controlling toy cars on the road, steps, ans pendulums.

**GAN creates WANN: generative adversarial network generates agnostic networks**

We have to develop a variational autoencoder, VAE, to generate simple agnostic networks, which are evaluated by some discriminative model.