DeepTrainer is an open source machine learning library that is currently under development.
Available from: https://github.com/bulyaki/DeepTrainer
DeepTrainer currently implements the following algorithms:
- Batch Backpropagation
- Distributed Batch Backpropagation
- Resilient Propagation
- Distributed Resilient Propagation
All algorithms are making use of a matrix implementation that uses automatic partitioning (4×4 or 8×8, single or double precision), and hardware accelerated matrix operations (dot product using SSE2, AVX, AVX512).
The algorithms are implemented using C++11 in Visual Studio on Windows OS, although a portable version is planned too.
Member of the @NVIDIA Inception program
DeepTrainer is also a trading name of Next Tier Limited, registered in England and Wales, reg. 07354214
About the DeepTrainer logo
If you are familiar with mathematical symbols you may have noticed that the logo I chose for the DeepTrainer project is a bit weird. The ∂ symbol is not part of the Greek alphabet, it is a special form of the Latin letter ‘d’, and in mathematics is used to denote partial derivatives in the form of ∂f/∂x, which means that ‘f’ is a function of many variables one of which is ‘x’, and we are only interested in the change of the ‘f’ function along the change of ‘x’ – while holding all other dependent variables constant.
However, the logo suggests partial derivative by time, which is not really used in mathematics or physics – because time is almost always considered as a fixed variable that is passing at a constant rate without changes, so we don’t usually meet functions that are dependent on the rate of change in time.
Well, actually there are a couple of fields in physics where partial derivatives by time make sense. One is easy to guess I think: where else would we be interested in the rate of change in time than General Relativity? In GR time is handled as if it was just another additional geometric coordinate to the usual 3 dimensional space – creating the so called 4 dimensional space-time. This is called the Minkowski space – a coordinate system consisting of the usual Euclidean space coordinates plus time, creating a 4 dimensional manifold on which the space and time intervals depend on the inertial reference frame from which they are observed.
Another field where ∂t is used Quantum Mechanics. A relativistic wave equation of a particle is described by the Dirac equation in which partial derivatives by time – as part of the so-called 4-gradient is playing an important role. The ∂/∂t differential operator is even abbreviated as ∂t in it.
I am a modern physics enthusiast who tries to keep up with the latest revelations about our reality both in cosmology and particle physics, so besides the logo being an abbreviation for DeepTrainer, it also expresses my admiration towards science and scientific research in general. So the logo first of all is just a simple initialism, but with a scientific mindset it can also mean a sort of an acronym forming a mathematical symbol.