access icon free Controlled accuracy approximation of sigmoid function for efficient FPGA-based implementation of artificial neurons

© The Institution of Engineering and Technology
Received 18/09/2013, Published

A controlled accuracy approximation scheme of the sigmoid function for artificial neuron implementation based on Taylor's theorem and the Lagrange form of the error is proposed. The main advantages of the proposed solution are two: it provides a systematic way to guarantee the required accuracy and it reuses the circuitry of the linear part of the neuron to compute the sigmoid function. The sigmoid derivative is also available for artificial neural networks with online learning capabilities.

Inspec keywords: approximation theory; field programmable gate arrays; learning (artificial intelligence); neural nets

Other keywords: sigmoid derivative; sigmoid function; Taylor theorem; controlled accuracy approximation scheme; Lagrange error form; linear neuron part circuitry; FPGA-based artificial neuron implementation; online learning capabilities; artificial neural networks

Subjects: Interpolation and function approximation (numerical analysis); Logic and switching circuits; Neural computing techniques; Interpolation and function approximation (numerical analysis); Logic circuits

References

    1. 1)
      • 2. Armato, A., Fanucci, L., Scilingo, E.P., De Rossi, D.: ‘Low-error digital hardware implementation of artificial neuron activation functions and their derivative’, Microprocess. Microsyst., 2011, 35, pp. 557–567, (doi: 10.1016/j.micpro.2011.05.007).
    2. 2)
      • 1. Misra, J., Saha, I.: ‘Artificial neural networks in hardware: a survey of two decades of progress’, Neurocomputing, 2010, 74, pp. 239–255, (doi: 10.1016/j.neucom.2010.03.021).
    3. 3)
      • 3. Basterretxea, K., Tarela, J.M., del Campo, I., Bosque, G.: ‘An experimental study on nonlinear function computation for neural/fuzzy hardware design’, IEEE Trans. Neural Netw., 2007, 18, pp. 266–283, (doi: 10.1109/TNN.2006.884680).
    4. 4)
      • 2. Armato, A., Fanucci, L., Scilingo, E.P., De Rossi, D.: ‘Low-error digital hardware implementation of artificial neuron activation functions and their derivative’, Microprocess. Microsyst., 2011, 35, pp. 557–567, doi:10.1016/j.micpro.2011.05.007 (doi: 10.1016/j.micpro.2011.05.007).
    5. 5)
      • 5. Xilinx: ‘LogiCORE IP DSP48 Macro v2.1’, http://www.xilinx.com/ accessed September 2013.
    6. 6)
      • 1. Misra, J., Saha, I.: ‘Artificial neural networks in hardware: a survey of two decades of progress’, Neurocomputing, 2010, 74, pp. 239–255, doi:10.1016/j.neucom.2010.03.021 (doi: 10.1016/j.neucom.2010.03.021).
    7. 7)
      • 4. Zamanlooy, B., Mirhassani, M.: ‘Efficient VLSI implementation of neural networks with hyperbolic tangent activation function’, IEEE Trans. Very Large Scale Integr. Syst., 2012, doi:10.1109/TVLSI.2012.2232321.
    8. 8)
      • 3. Basterretxea, K., Tarela, J.M., del Campo, I., Bosque, G.: ‘An experimental study on nonlinear function computation for neural/fuzzy hardware design’, IEEE Trans. Neural Netw., 2007, 18, pp. 266–283, doi:10.1109/TNN.2006.884680 (doi: 10.1109/TNN.2006.884680).
http://iet.metastore.ingenta.com/content/journals/10.1049/el.2013.3098
Loading

Related content

content/journals/10.1049/el.2013.3098
pub_keyword,iet_inspecKeyword,pub_concept
6
6
Loading