Original link: https://arxiv.org/abs/1803.03067
MAC network is a new fully micro-neural network architecture, the goal is to enhance the expressive power of inference. Inspired by the principles of computer architecture, MAC does not use the traditional neural network black box architecture, and instead adopts a design that promotes transparency and multi-use.
As shown in the figure, the components of the MAC network are an input neuron, a core recurrent network, and an output neuron. The input neuron transforms the breakthrough into a feature, and the cyclic network decomposes the problem into a series of operations, extracting information from the picture to form a cyclic memory. Through these operations, the network performs logical reasoning according to the sequence.
The core of the MAC network lies in the middle recurrent neural network. Unlike the traditional network structure, this part of the recurrent network consists of MAC neurons. p MAC neurons are strung together, and each neuron is responsible for an inference step, transferring and updating control and memory step by step, and finally output.
The author said that the design of the MAC network was inspired by the computer architecture. I guess it refers to the pipeline CPU in the computer composition-divided into five pipeline stages, data and control signals are passed step by step, and the data path includes two processes of reading and writing The signal controls the reading and writing of data.
As shown in the figure, the MAC neuron of each layer reads the control and memory of the upper level, and passes it to the control module and read-write module of the neuron respectively.
The control module uses the superior control information Ci-1 and the information q extracted from the question to generate the inferior control information Ci.
The read-write module uses the upper memory Mi-1 and the picture KB to read in the new information Ri, and uses the Ri and Mi-1 to write the lower memory Mi. The reading process and the writing process are controlled by Ci-1 and Ci, respectively.
In the experiment, the author compares the accuracy of MAC and other models on different tasks of the CLEVR data set. The model achieves state-of-the-art accuracy, and also has high computational efficiency and data efficiency. For better results, the amount of data required is only 1/6 of other models.