Neural network is an active research field which involves many different (unsolved) issues, for example, different types of configuration of the network architectures, training strategies, etc. Amongst these active issues, the choice of loss (or cost) functions plays an important role in how a neural network model is to be optimized (trained) and how the model will perform after the training. Given the choice of measurement criteria, loss functions measure how far an estimated output is from its true value. And the measurement criteria can change depending on the task in hand and the goal to be met. The objective of this project is to understand the role of different loss functions and to evaluate the dependence of the performance on the loss functions using the language prediction problem.

Volumetric light scattering caused by participating media has been and still is a complicated, but visually pleasing lighting effect that increases the sense of realism of digital scenes tremendously. Various solutions to rendering these effects in real-time exist, basing themselves upon the single scattering approximation. Transparency is another difficult subject within the field of computer graphics, due to the constraint of having to somehow find the order of semi-transparent objects, or to approximate them independently of order. However, no prior research has focused on scenes in which both participating media and semi-transparent objects are present. While a technique such as Depth Peeling (Everitt, 2001) can be used to solve the problem, this method has some major downsides and often does not obtain real-time results. In this thesis we attempt to obtain faster results, while trying to pertain as much quality as possible. To this end, we take the Depth Peeling method as our reference.

By partitioning scenes into what we call 'depth zones' as seen from the light, we can obtain necessary color and depth information to then approximately render the required result using ray-marching. Partitioning scenes is done in various ways (uniformly, non-uniformly) and special data structures are used to store all necessary information efficiently. The combination of partitioning and data storage led to two new methods being implemented: Multisample Texture Depth Zones (MSTDZ) and histogram-based MSTDZ, of which the last sorts fragments of a scene into bins to determine where to partition depth zones in a less naive way than standard MSTDZ. Both of these methods have the ability to approximate Depth Peeling's results very well, at higher speeds and, depending on the approximation, using lower amounts of memory.
A final addition was implemented which groups texels of the histogram texture used in histogram-based MSTDZ as to use smaller textures. In low complexity scenes, this gives the possibility to lower the amount of consumed memory, at the cost of time and a small amount of quality, compared to the reference method.