This paper introduces a classical-quantum hybrid algorithm for machine learning on near-term quantum processors, called quantum circuit learning (QCL). QCL uses a low-depth quantum circuit to learn tasks by iteratively tuning its parameters. The framework allows a quantum circuit to approximate nonlinear functions, as shown by numerical simulations. QCL combines a quantum circuit with a classical computer to perform supervised or unsupervised learning tasks. In supervised learning, the algorithm learns to output results close to a teacher's data by optimizing parameters. In unsupervised learning, it minimizes a cost function without teacher data. QCL is compared with existing algorithms like quantum variational eigensolver (QVE) and quantum approximate optimization algorithm (QAOA). Unlike these, QCL directly tunes the circuit parameters rather than the weights of a classical computer. Theoretical analysis shows that QCL can approximate any analytical function if the circuit has enough qubits. Numerical simulations demonstrate that a 6-qubit circuit can fit the dynamics of 3 spins in a 10-spin system with a fully connected Ising Hamiltonian. QCL's ability to approximate functions is based on the tensor product structure of quantum systems, which allows for exponential growth in the number of basis functions. This enables QCL to represent complex functions that are intractable for classical computers. The unitarity of the quantum circuit also helps avoid overfitting by constraining the parameters. Numerical simulations show that QCL can perform tasks like function approximation, classification, and fitting quantum many-body dynamics. The paper also discusses the optimization procedure for QCL, showing how gradients of expectation values can be calculated using quantum circuits. The framework is shown to be effective in both regression and classification tasks, with results demonstrating its ability to learn complex functions and classify data. The results indicate that QCL has the potential to outperform classical methods in certain tasks, particularly those involving high-dimensional data. The paper concludes that QCL provides a promising approach for machine learning on near-term quantum devices.This paper introduces a classical-quantum hybrid algorithm for machine learning on near-term quantum processors, called quantum circuit learning (QCL). QCL uses a low-depth quantum circuit to learn tasks by iteratively tuning its parameters. The framework allows a quantum circuit to approximate nonlinear functions, as shown by numerical simulations. QCL combines a quantum circuit with a classical computer to perform supervised or unsupervised learning tasks. In supervised learning, the algorithm learns to output results close to a teacher's data by optimizing parameters. In unsupervised learning, it minimizes a cost function without teacher data. QCL is compared with existing algorithms like quantum variational eigensolver (QVE) and quantum approximate optimization algorithm (QAOA). Unlike these, QCL directly tunes the circuit parameters rather than the weights of a classical computer. Theoretical analysis shows that QCL can approximate any analytical function if the circuit has enough qubits. Numerical simulations demonstrate that a 6-qubit circuit can fit the dynamics of 3 spins in a 10-spin system with a fully connected Ising Hamiltonian. QCL's ability to approximate functions is based on the tensor product structure of quantum systems, which allows for exponential growth in the number of basis functions. This enables QCL to represent complex functions that are intractable for classical computers. The unitarity of the quantum circuit also helps avoid overfitting by constraining the parameters. Numerical simulations show that QCL can perform tasks like function approximation, classification, and fitting quantum many-body dynamics. The paper also discusses the optimization procedure for QCL, showing how gradients of expectation values can be calculated using quantum circuits. The framework is shown to be effective in both regression and classification tasks, with results demonstrating its ability to learn complex functions and classify data. The results indicate that QCL has the potential to outperform classical methods in certain tasks, particularly those involving high-dimensional data. The paper concludes that QCL provides a promising approach for machine learning on near-term quantum devices.