Hamiltonian Neural Networks (HNNs) are introduced to endow neural networks with better inductive biases by learning exact conservation laws in an unsupervised manner, inspired by Hamiltonian mechanics. HNNs parameterize the Hamiltonian as a neural network and learn it directly from data, allowing them to conserve quantities analogous to total energy. The paper evaluates HNNs on problems where energy conservation is crucial, such as the two-body problem and pixel observations of a pendulum. HNNs outperform regular neural networks in terms of training speed and generalization, and they are perfectly reversible in time. The authors also explore the use of HNNs in modeling larger systems, including the two-body problem and a pixel-based pendulum, demonstrating their ability to learn and conserve energy over long timescales. Additionally, HNNs exhibit useful properties such as the ability to add and remove energy and perfect reversibility. The paper discusses related work and highlights the potential of HNNs in combining the strengths of deep learning and physics-based models.Hamiltonian Neural Networks (HNNs) are introduced to endow neural networks with better inductive biases by learning exact conservation laws in an unsupervised manner, inspired by Hamiltonian mechanics. HNNs parameterize the Hamiltonian as a neural network and learn it directly from data, allowing them to conserve quantities analogous to total energy. The paper evaluates HNNs on problems where energy conservation is crucial, such as the two-body problem and pixel observations of a pendulum. HNNs outperform regular neural networks in terms of training speed and generalization, and they are perfectly reversible in time. The authors also explore the use of HNNs in modeling larger systems, including the two-body problem and a pixel-based pendulum, demonstrating their ability to learn and conserve energy over long timescales. Additionally, HNNs exhibit useful properties such as the ability to add and remove energy and perfect reversibility. The paper discusses related work and highlights the potential of HNNs in combining the strengths of deep learning and physics-based models.