地点：心理学院301会议室

摘要：In many computational neuroscience and machine learning models (e.g. predictive coding and backpropagation), symmetrical feedback weights was the one key ingredient for the implementation. However, there are no empirical evidence supporting the existence of such symmetric reciprocal connections, resulting one of the greatest obstacles for linking the machine world with the real world. In this talk, I will propose a simple neural network model to address this issue. This model is implemented using current-based neurons (leaky integrate-and-fire neurons) and widely observed selective excitatory feedforward connections and non-selective excitatory feedback connections. We demonstrated that, taking advantage of the temporal information of the neuron spikes, the simple neural model can imitate the feedforward weights in the feedback pathway with reversed signs.

简历