Quantum computing is becoming a reality, at least on a small scale. However, designing a good quantum algorithm is still a challenging task. This has been a huge major bottleneck in quantum computation for years. In this work, we will show that it is possible to take a detour from the conventional programming approach by incorporating machine learning techniques, speci cally neural networks, to train a quantum system such that the desired algorithm is \learned," thus obviating the program design obstacle. Our work here merges quantum computing and neural networks to form what we call \Quantum Neural Networks" (QNNs). Another serious issue one needs to overcome when doing anything quantum is the problem of \noise and decoherence". A well-known technique to overcome this issue is using error correcting code. However, error correction schemes require an enormous amount of additional ancilla qubits, which is not feasible for the current state-of-the-art quantum computing devices or any near-term devices for that matter. We show in this work that QNNs are robust to noise and decoherence, provide error suppression quantum algorithms. Furthermore, not only are our QNN models robust to noise and decoherehce, we show that they also possess an inherent speed-up, in term of being able to learned a task much faster, over various classical neural networks, at least on the set of problems we benchmarked them on. Afterward, we show that although our QNN model is designed to run on a fundamental level of a quantum system, we can also decompose it into a sequence of gates and implement it on current quantum hardware devices. We did this for a non-trivial problem known as the \entanglement witness" calculation. We then propose a couple of di erent hybrid quantum neural network architectures, networks with both quantum and classical information processing. We hope that this might increase the capability over previous QNN models in terms of the complexity of the problems it might be able to solve.