We proposed Ising learning algorithm, the first technique to train multilayer feedforward neural networks on Ising machines (quantum computers). The training time is reduced by 90% compared to CPU/GPU.

We unified the filtering and control capabilities into a single policy network in RL, achieving SOTA noise robustness and action smoothness in real-world control tasks.

We proposed DACER, an online reinforcement learning algorithm that utilizes a diffusion model as the actor network to enhance the representational capacity of the policy.

we proposed a variant of neural ODE, called SmODE, to smooth out control actions in RL. A mapping function is incorporated to estimate the changing speed of system dynamics.

We proposed a policy network for RL with low-pass filtering ability, named Smonet, to alleviate the action nonsmoothness issue by learning a low-frequency representation within hidden layers.

We proposed LipsNet, a smooth and robust neural network with adaptive Lipschitz constant, to deal with the action fluctuation problem in RL (reinforcement learning).