This bundle contains source files for a C++ implementation of Extreme Learning Machines (ELM) for microcontrollers. It is based on MbedOS and a GSL port for ARM Cortex-M4, which is provided as static library.

Using this application, it is possible to train an artificial neural network in the single layer feedforward network (SLFN) topology following the ELM theory, and use it later once trained. It leverages ELM in 2 ways:

• ELM is a super fast method for training artificial neural networks, therefore suitable for the processing power of microcontrollers - no iterations of extensive calculations are needed.
• Since random weights need not to be tuned during the trained step, they can be stored in program memory and used for every new training.

There is though one difference in this application, which is random weight generation according to a binary enconding scheme, so that the random weights connecting the input layer to the hidden layer will have 2^b distinct values uniformly sorted within the interval [-1,1], where b is the bit count. For instance, selecting b=1 will lead to random weights -1 and 1 assigned in the hidden laye. Selecting b=2 will provide a pool containing the values -1, -0.5, 0.5 and 1, from which the random weights will be assigned randomly, and so on for grater values of b.

This application was designed originally to work with ultrasonic signals from weld beads, on a STMF4 Discovery development board, as part of a scientific research. Users are encouraged to employ it in scientific research, engineering and robotic projects and contribute to its development. In case of using this application in reasearch, please cite it:

Lucas C. Silva,Computers and Electrical Engineering, https://doi.org/10.1016/j.compeleceng.2020.10689

This application uses MbedOS, therefore it shall be installed on the working environment. MBED-CLI works fine. Once installed, initialize a new project in the workspace with

mbed new [program_name]

then put the files here inside a new folder (e.g. elm/) in the newly created program_name folder. The file tree should look like

|— program_name
   |— elm/
      |—inc/
      |—lib/
      |—src/
      |—examples/
   |—mbed-os/

Target and toolchain must also be configured. This application is run using DISCO_F407VG and GCC_ARM respectively (install the ARM toolchain in case it is missing.

The file build_config.json shall be invoked in compilation by using, referring to the above example, mbed compile --app-config elm/build_config.json as it contains the include paths and library path flags that are passed to the compiler.

The application is based on 3 classes:

• Slfn holds the configuration of the neural netowrk, such as input node count, hidden neuron count, output neuron count, and so on.
• Elmcontains the random weights, random bias and output weights of the network.
• Organizeris a handler for dealing with sample sets, containing the training set, targets, test sample and builders for multidimensional samples.

The general procedure is to create an Slfn object and use it to manage the sample sets via Organizer object. Once the samples are collected, finally run the training method of the Elm network. If the network was already trained, feed it with test samples in order to obtain the network output.

Please run doxygen in order to obtain the code documentation.

An application example for a multiclass problem is provided in the path examples/multiclass/, which was used in the aforementioned research. This example expects samples and targets in float format (4 bytes per value) whose size is defined in an Slfn object, then trains the network, waits for test samples and calculate the outputs of the trained network.