On Training Efficiency and Computational Costs of a Feed Forward Neural Network: A Review

Table 8

Summary of Piecewise Linear Approximations.

Ref.	Method	Convergence	Precision	Computational costs	Notes
[44]	Piecewise linear; “VHDL-C”	N/A	MSE = 0.00049	213 clock cycles	Resources used: Flip Flop Slices: 1277; 4 input LUTs: 3767; BRAMS: 4
[45]	“Bajger-Omondi” method	N/A	Absolute error: up to 10^−6 for 128 pieces with 18-bit precision	N/A
[46]	PWL approximation	N/A	N/A	Propagation delay: 1.834 ns (100 ns more than LUT approach)	Resources used: 4 input LUTs: 108 (79 less than LUT approach); Slices: 58 (44 less than LUT approach); Total gates: 1029 (329 less than LUT)
[47, 48]	A-Law	N/A	% error = 0.63; 84% accuracy (classification problem)	Propagation delay: 3.729 ns	Resources used: Slices: 185; LUTs: 101; Total gates: 1653
[47, 49]	Alippi	N/A	% error = 1.11	Propagation delay: 3.441 ns	Resources used: Slices: 127; LUTs: 218; Total gates: 1812
[47, 50, 51]	PLAN	N/A	% error = 0.63; 85% accuracy (classification problem)	Propagation delay: 4.265 ns	Resources used: Slices: 127; LUTs: 218; Total gates: 1812