This paper presents a fully-optical neural network (ONN) architecture that leverages the advantages of photonics to achieve significant improvements in computational speed and power efficiency compared to conventional electronic neural networks. The ONN uses integrated photonic circuits to perform matrix multiplications and nonlinear transformations, enabling high-speed, energy-efficient neural network computations. The architecture is demonstrated using a programmable nanophotonic processor with 56 Mach-Zehnder interferometers and 213 phase shifting elements. The ONN is tested on a vowel recognition dataset, achieving an accuracy comparable to conventional digital computers. The ONN's performance is evaluated under varying levels of phase encoding and photodetection noise, showing that it can achieve high accuracy with minimal power consumption. The ONN's architecture is particularly effective for tasks requiring high-speed, low-power processing, such as speech recognition and image processing. The paper also discusses the potential of ONNs for future applications, including the use of non-volatile phase-change materials to reduce power consumption. The ONN's architecture is shown to be significantly more energy-efficient than conventional electronic computers, with power efficiency at least three orders of magnitude better for conventional learning tasks. The paper concludes that ONNs have the potential to revolutionize the field of artificial intelligence by enabling faster, more efficient neural network computations.This paper presents a fully-optical neural network (ONN) architecture that leverages the advantages of photonics to achieve significant improvements in computational speed and power efficiency compared to conventional electronic neural networks. The ONN uses integrated photonic circuits to perform matrix multiplications and nonlinear transformations, enabling high-speed, energy-efficient neural network computations. The architecture is demonstrated using a programmable nanophotonic processor with 56 Mach-Zehnder interferometers and 213 phase shifting elements. The ONN is tested on a vowel recognition dataset, achieving an accuracy comparable to conventional digital computers. The ONN's performance is evaluated under varying levels of phase encoding and photodetection noise, showing that it can achieve high accuracy with minimal power consumption. The ONN's architecture is particularly effective for tasks requiring high-speed, low-power processing, such as speech recognition and image processing. The paper also discusses the potential of ONNs for future applications, including the use of non-volatile phase-change materials to reduce power consumption. The ONN's architecture is shown to be significantly more energy-efficient than conventional electronic computers, with power efficiency at least three orders of magnitude better for conventional learning tasks. The paper concludes that ONNs have the potential to revolutionize the field of artificial intelligence by enabling faster, more efficient neural network computations.