摆脱GPU依赖，Nature发布「物理神经网络」综述：实现大规模、高效AI训练与推理

Core Insights - The article discusses the transformative potential of Physical Neural Networks (PNNs) in AI computation, emphasizing their ability to utilize physical systems for more efficient training and inference compared to traditional digital GPUs [1][4][21] - A recent review published in Nature highlights the development of PNNs from a training perspective, suggesting that with sufficient research investment, PNNs could revolutionize AI computing [3][21] Group 1: PNNs Overview - PNNs leverage physical systems such as light, electricity, and vibrations for computation, potentially overcoming the limitations of traditional digital chips [1][4] - PNNs are categorized into two types: Isomorphic PNNs, which strictly adhere to predefined mathematical transformations, and broken-isomorphism PNNs, which allow for approximate physical transformations [7][8] Group 2: Training Methods - Various training techniques for PNNs include: - In silico training, which uses digital twin models for weight gradient calculations [8] - Physical-Aware Training (PAT), which allows for gradient extraction through approximate predictive models [9] - Feedback Alignment (FA) and Direct Feedback Alignment (DFA), which enable training without transferring weights between forward and backward passes [10] - Physical Local Learning (PhyLL), which simplifies training by using cosine similarity between positive and negative samples [11] - Zero-order and no-gradient training methods, which estimate gradients through sampling strategies [12] - Gradient descent training through physical dynamics, which directly utilizes physical principles for weight updates [13] Group 3: Commercial Viability - Despite the large physical size of AI models, PNNs may offer significant efficiency advantages over digital systems when designed appropriately [16][19] - The scalability of PNNs is crucial, as their physical characteristics could lead to better energy efficiency compared to traditional digital devices [16][19] Group 4: Future Challenges - PNNs face challenges such as noise accumulation during computation, which can affect precision [20] - There is a need for optimization of PNN architectures to better align with the capabilities of physical hardware [20] - Balancing neural and physical characteristics in PNN design is essential for effective implementation [20]