between the two subsystems. To push the envelope for speed and energy efficiency, a radically different computing paradigm that allows in situ computation within the memory, or in-memory computing (IMC), is revolutionary to address the issues associated with abundant data movement. [2,3] In particular, IMC based on analog memristors holds promise to provide a low latency and energy-efficient approach to implement data-centric applications such as image processing by means of neural network training. [3,4] Thus far, the development of convolutional neural network (CNN), an important model for image recognition, has been experimented using memristor devices, which are plagued with fundamental scientific issues related to interdevice variability, nonlinearity, and sneak path current. [5][6][7] Such a twoterminal memristor is typically integrated with metal oxide as a switching medium, which relies on the formation and rupture of filaments in the amorphous medium for resistive switching (RS). [5,6,[8][9][10][11][12] However, the stochasticity of the ions' movement is difficult to control and thus results in poor spatial and temporal variations in the RS performance. [13][14][15] The variation has a significant influence on the computing accuracy loss, as most reported high computing accuracies were only realized with small cycleto-cycle variation. [7,13,16] Compared with single-layered oxide, switching medium made of double-layered oxide offers betterIn-memory computing based on memristor arrays holds promise to address the speed and energy issues of the classical von Neumann computing system. However, the stochasticity of ions' transport in conventional oxidebased memristors imposes severe intrinsic variability, which compromises learning accuracy and hinders the implementation of neural network hardware accelerators. Here, these challenges are addressed using a low-voltage memristor array based on an ultrathin PdSeO x /PdSe 2 heterostructure switching medium realized by a controllable ultraviolet (UV)-ozone treatment. A distinctively different ions' transport mechanism is revealed in the heterostructure that can confine the formation of conductive filaments, leading to a remarkable uniform switching with low set and reset voltage variability values of 4.8% and −3.6%, respectively. Moreover, convolutional image processing is further implemented using various crossbar kernels that achieve a high recognition accuracy of ≈93.4% due to the highly linear and symmetric analog weight update as well as multiple conductance states, manifesting its potential beyond von Neumann computing.