Hybrid Quantum Learning
Artificial Neural Networks is learning pattern in data with physics (Nobel Prize in Physics 2024)
Training neural networks, especially in the context of reinforced learning for real-world indeterministic environments, is also a process of learning complex probability distributions and correlations. In the process, a model starting at a high entropy state decreases its entropy as it encodes learned information from training into its neural network. The model's ability to learn the intrinsic probability distribution of its learning goal is critical for both convergence and generalizability. As it has been both theoretically and experimentally shown that quantum processors could generate nontrivial probability distributions that are hard for classical algorithms to simulate, such novel platforms offer a new and powerful way for next-gen AI models based on Deep Artificial Neural Networks to evolve via a quantum leap forward.
Hybrid Quantum Reinforced Learning
Reinforced learning (2024 ACM A.M. Turing Award) is a cornerstone in modern machine learning / artificial intelligence model development and their applications to large-language models (LLMs) and advanced robotics control (Embodied AI).
The World's Most Advanced Quantum Reinforced Learning for Humanoid Robotics
To highlight the potential of hybrid quantum enhanced ANN in solving real-world problems with performance on par with or even exceeding comparable classical counterparts, here we did an experiment that combines parametrized quantum circuits (PQCs) with classical deep neural networks (DNNs) and the TD3 architecture to solve the characteristic Humanoid-v4 environment from Gymnasium library. With novel methods invented by Anyon added in the training steps, we were able to train the hybrid model efficiently and scalably, resulting the most sophisticated reinforced learning environment successfully addressed by hybrid quantum methods so far. The experimented Hybrid Deep Quantum Neural Network (hDQNN) even outperforms all classical counterparts where we replace the PQC with a fully-connected (FC) neural network, a random binary generator (RBG), or a constant 0 feedforward (0L) while keeping the rest of the reinforced learning agent and training hyperparameters the same. The test rewards are shown in the right figure below. The work pave the way for further research and development by the AI/ML community to leverage quantum devices for enhancing the development of AI and advanced robotics.
​
See paper published online soon for details of the experiments and methods.
