Humans detect tactile stimuli through a combination of pressure and vibration signals using different types of cutaneous receptor. The development of artificial tactile perception systems is of interest in the development of robotics and prosthetics, and artificial receptors, nerves and skin have been created. However, constructing systems with human-like capabilities remains challenging. Here, we report an artificial neural tactile skin system that mimics the human tactile recognition process using particle-based polymer composite sensors and a signal-converting system. The sensors respond to pressure and vibration selectively, similarly to slow adaptive and fast adaptive mechanoreceptors in human skin, and can generate sensory neuron-like output signal patterns. We show in an ex vivo test that undistorted transmission of the output signals through an afferent tactile mouse nerve fibre is possible, and in an in vivo test that the signals can stimulate a rat motor nerve to induce the contraction of a hindlimb muscle. We use our tactile sensing system to develop an artificial finger that can learn to classify fine and complex textures by integrating the sensor signals with a deep learning technique. The approach can also be used to predict unknown textures on the basis of the trained model. A tactile sensing system that can learn to identify different types of surface can be created using sensors that mimic the fast and slow responses of mechanoreceptors found in human skin.