Quantum Superposition Spiking Neural Network
Quantum Superposition Spiking Neural Network
Quantum brain as a novel hypothesis states that some non-trivial mechanisms in quantum computation, such as superposition and entanglement, may have important influence for the formation of brain functions. Inspired by this idea, we propose Quantum Superposition Spiking Neural Network (QS-SNN), which introduce quantum superposition to spiking neural network models to handel challenges which are hard for other state-of-the-art machine learning models.For human brain, grasping the main information no matter how the background changes is necessary to interact efficiently with diverse environments. As an example, it is easy for human to recognize the digits whether it is white character with black background or inversely black character with white background. While if the current machine learning models are trained with one of the cases (e.g. white character with black background), it will be nearly impossible for them to recognize the color inverted version. To handel this challenge, we propose two-compartment spiking neural network with superposition states encoding, which is inspired by quantum information theory and spatial-temporal spiking property from neuron information encoding in the brain. Typical network structures like fully-connected ANN, VGG, ResNet and DenseNet are challenged with the same task. We train these networks on original image dataset and then invert the background color to test their generalization. Result shows that artificial neural network can not deal with this condition while the quantum superposition spiking neural network(QS-SNN) which we proposed in this paper recognizes the color-inverse image successfully. Further the QS-SNN shows its robustness when noises are added on inputs.
量子叠加加标神经网络
量子脑作为一种新的假设，指出量子计算中的一些非平凡机制，例如叠加和纠缠，可能对脑功能的形成具有重要影响。受此想法的启发，我们提出了量子叠加加标神经网络（QS-SNN），该技术将量子叠加引入尖峰神经网络模型，以应对其他先进的机器学习模型难以应对的挑战。.. 对于人脑而言，无论背景如何变化，都必须掌握主要信息，才能有效地与各种环境交互。例如，无论是黑色背景的白色字符还是白色背景的反向黑色字符，人类都容易识别数字。虽然如果当前的机器学习模型是用其中一种情况（例如，白色字符和黑色背景）进行训练的，则它们几乎不可能识别出颜色反转的版本。为了应对这一挑战，我们提出了一种具有重叠状态编码的两格尖峰神经网络，该网络受量子信息理论和大脑中神经元信息编码的时空尖峰特性的启发。完全连接的ANN，VGG，ResNet和DenseNet面临相同任务的挑战。我们在原始图像数据集上训练这些网络，然后反转背景色以测试其概括。结果表明，本文提出的量子叠加加标神经网络（QS-SNN）能够很好地识别彩色逆像，而人工神经网络无法应对这种情况。此外，当在输入上添加噪声时，QS-SNN表现出其鲁棒性。 （阅读更多）
