We investigate the synergy between (simulated) noisy quantum computers and classical deep neural networks to address the challenge of quantum error mitigation. Our approach combines noisy quantum data with classical circuit descriptors to train scalable convolutional neural networks capable of predicting accurate expectation values of quantum circuits with more qubits than those used for training. Additionally, we explore the use of circuit knitting techniques to implement quantum circuits for the neural network training set, enhancing the performance of Variational Quantum Eigensolver algorithms.