In this thesis we seek to make advances towards the goal of effective learned compression. This entails using machine learning models as the core constituent of compression algorithms, rather than hand-crafted components.
To that end, we first describe a new method for lossless compression. This method
allows a class of existing machine learning models – latent variable models – to be
turned into lossless compressors. Thus many future advancements in the field of
latent variable modelling can be leveraged in the field of lossless compression. We
demonstrate a proof-of-concept of this method on image compression. Further, we
show that it can scale to very large models, and image compression problems which
closely resemble the real-world use cases that we seek to tackle.
The use of the above compression method relies on executing a latent variable
model. Since these models can be large in size and slow to run, we consider how
to mitigate these computational costs. We show that by implementing much of the
models using binary precision parameters, rather than floating-point precision, we
can still achieve reasonable modelling performance but requiring a fraction of the
storage space and execution time.
Lastly, we consider how learned compression can be applied to 3D scene data – a
data medium increasing in prevalence, and which can require a significant amount of
space. A recently developed class of machine learning models – scene representation
functions – has demonstrated good results on modelling such 3D scene data. We
show that by compressing these representation functions themselves we can achieve
good scene reconstruction with a very small model size.