The paper presents a novel architecture for a fully-optical neural network (ONN) that leverages the unique advantages of optics to achieve significant improvements in computational speed and power efficiency compared to conventional electronic neural networks. The ONN architecture is designed to perform linear transformations and nonlinear activations at the speed of light, with minimal power consumption. The authors experimentally demonstrate the essential components of their architecture using a programmable nanophotonic processor, achieving a level of accuracy comparable to conventional digital computers for vowel recognition tasks. The ONN architecture is shown to be at least two orders of magnitude faster for forward propagation and provides linear scaling of neuron number versus power consumption. The paper discusses the potential of ONNs for processing large datasets and their advantages in terms of energy efficiency, particularly for deep neural networks. The authors also explore the possibility of using optical bistability instead of saturable absorption to further enhance the power efficiency of the ONN. The paper concludes by highlighting the potential of ONNs for training neural network parameters directly on the photonics chip and the need for further integration of optical interconnects to fully realize the benefits of all-optical computing.The paper presents a novel architecture for a fully-optical neural network (ONN) that leverages the unique advantages of optics to achieve significant improvements in computational speed and power efficiency compared to conventional electronic neural networks. The ONN architecture is designed to perform linear transformations and nonlinear activations at the speed of light, with minimal power consumption. The authors experimentally demonstrate the essential components of their architecture using a programmable nanophotonic processor, achieving a level of accuracy comparable to conventional digital computers for vowel recognition tasks. The ONN architecture is shown to be at least two orders of magnitude faster for forward propagation and provides linear scaling of neuron number versus power consumption. The paper discusses the potential of ONNs for processing large datasets and their advantages in terms of energy efficiency, particularly for deep neural networks. The authors also explore the possibility of using optical bistability instead of saturable absorption to further enhance the power efficiency of the ONN. The paper concludes by highlighting the potential of ONNs for training neural network parameters directly on the photonics chip and the need for further integration of optical interconnects to fully realize the benefits of all-optical computing.