AI 中文总结
研究大规模神经网络信号放大问题，利用金属-绝缘体转变材料制成的两终端器件，通过负微分电阻实现约11.5倍信号放大，可控制放大且能用模型预测增益，为开发优化轴突样放大功能建立框架。
AI 中文摘要
大规模神经网络中的信号耗散会导致信息丢失并最终导致计算失败，因此需要在神经元-突触连接处进行局部信号放大。在生物神经系统中，轴突负责局部信号放大。将轴突功能转化为硬件并非易事，因为模拟人脑意味着构建由约100亿个相互连接的神经元组成的网络，每个神经元都需要一个专用的紧凑且可扩展的放大器。在此，我们展示了由金属-绝缘体转变材料制成的简单两终端器件中的信号放大。通过在相变边缘操作器件并利用负微分电阻，我们实现了高达约11.5倍的稳健信号放大。我们还展示了由真实神经晶体管产生的尖峰序列的放大，为人工神经元和轴突的直接集成开辟了新的令人兴奋的机会。放大可以通过易于调节的实验参数来控制，包括直流偏置、交流激励、串联电阻和温度。我们进一步提出了一个使用易于观察的传输特性来预测增益的模型。我们的结果为在非线性电子材料中开发和优化类似轴突的放大功能建立了一个框架。
英文摘要
Signal dissipation in large-scale neural networks can lead to information loss and ultimately to computational failures, necessitating local signal amplification at neuron - synapse connections. In biological nervous systems, axons are responsible for local signal amplification. Translating axon functionality into hardware, i.e., the ability to amplify and transmit signals without loss, is non-trivial because emulating the human brain implies building networks composed of ~10 billion interconnected neurons, each requiring a dedicated compact and scalable amplifier. Here, we demonstrate signal amplification in simple two-terminal devices made of a metal-insulator transition material. By operating the devices on the verge of the phase transition and taking advantage of negative differential resistance, we achieve robust signal amplification up to a factor of ~11.5. We also demonstrate the amplification of spiking sequences generated by a real neuristor, opening new exciting opportunities for the direct integration of artificial neurons and axons. The amplification can be controlled by easily adjustable experimental parameters, including DC bias, AC excitation, series resistance, and temperature. We further propose a model that predicts the gain using readily observable transport characteristics. Our results establish a framework for developing and optimizing axon-like amplification functionalities in nonlinear electronic materials.
Comments15 pages
