HTTPS SSH
Compilers Artifacts Status Coverage
Full build GCC tests with BLAS, tests without BLAS, examples Build Status
Test build GCC, Clang tests with BLAS Build Status Coverage Status

NOTCH

Feed-forward Neural Networks in C++11 for the rest of us.

This is (supposed to be)

  • a header-only C++11 library

  • with no extra dependencies (CBLAS dependency is optional)

  • cross-platform (Linux and Windows)

  • reasonably fast (for a CPU-only implementations, see benchmarks)

  • but not at the cost of algorithms' readability and portability

This library is named after Notch, a transmembrane protein which acts as a receptor and has a role, amonth other things, in neuronal function and development.

Motivation

This library was born out of frustration.

  • Most of the neural network frameworks are notoriously difficult to install and deploy. Some of them work only on a particular operating system flavor, or have very specific hardware requirements. Some of them require hundreds of megabytes of dependencies to be installed.

  • Many neural network frameworks are designed only to train neural networks. Few care about using them and integrating neural networks into the end-user software.

  • Some neural network libraries are just too generic and too verbose, though it can be subjective.

Notch is supposed to lower the barrier to entry and be a tool which works anywhere where a modern C++ compiler is available. Just copy a header file. No need to install anything. (but if you can throw in also a BLAS library, that helps)

Notch is also designed to be embedded into other software. Standalone executable size starts from approx. 100 KiB, without extra dependencies to bundle.

How to use

See Getting started, examples, and API reference.

Note: Notch is still in the early stage of development. Some important features are still missing. Interfaces change. Hey, it doesn't even have a version number yet!

Conditional Compilation

Set these compilation flags before including "notch.hpp" and other Notch headers:

  • #define NOTCH_USE_CBLAS

    if you're linking your program with a CBLAS library and want to use an efficient implemtation of the linear algebra.

  • #define NOTCH_USE_OPENMP

    to parallelize Notch computations using OpenMP. If you use OpenMP and BLAS together, make sure that the BLAS library is compatible with OpenMP (OpenBLAS should be compiled with OpenMP support).

  • #define NOTCH_ONLY_DECLARATIONS

    If you include Notch headers in more than one compilation unit (source file), then, to suppress multiple definitions, #define NOTCH_ONLY_DECLARATIONS before Notch includes in all but one of the source files. To see an example, look how Notch is used in test/ source files.

Naming Conventions

Different neural network libraries name the same things differently. This library sticks to the notation of NNLM3.

For neurons with activation potential $v_j = W_{ji} y_i + b_j$, activation function $\phi(v_j)$, and the network loss $E$, the naming convention is summarized in the following table:

| CONCEPT                      | NOTCH                 | TORCH7 NN           | CAFFE          |

| network layer                | ABackpropLayer        | Module              | Layer          |
| loss E calculator            | ALossLayer            | Criterion           | LossLayer      |
| entire network               | Net                   | Sequential          | Net            |
| network initializer          | Init                  | ???                 | Filler         |
| network optimizer            | SGD                   | StochasticGradient  | SGDSolver      |

| forward propagation step     | .output()             | :forward()          | .Forward()     |
| back propagation step        | .backprop()           | :backward()         | .Backward()    |
| update weights               | .update()             | :updateParameters() | .Update()      |
| read weights W_{ji}          | .getWeights()         | .weight             | .layer_param() |
| read bias    b_j             | .getBias()            | .bias               | .layer_param() |

| \partial E / \partial W_{ji} | weight sensitivity    | .gradWeight         | ???            |
| \partial E / \partial b_j    | bias sensitivity      | .gradBias           | ???            |
| \partial E / \partial y_i    | error signals, errors | .gradInput          | ???            |
| \partial E / \partial v_j    | local gradient        | .gradOutput         | ???            |

Bibliography

  • NNLM3: Neural Networks and Learning Machines / Simon Haykin. -- 3rd ed. (libgen)

  • PRML: Pattern Recognition and Machine Learning / Christoper M. Bishop. -- (libgen, web)

  • NNTT: Neural Networks: Tricks of the Trade / GrĂ©goire Montavon et al (eds.) -- 2nd ed. (libgen)