Research On Visual Logic Learning Oriented To Image Description IQ Problem

Logical reasoning is the core of human intelligence and a key and challenging research topic in the field of artificial intelligence.Artificial intelligence has made significant progress in image recognition and classification tasks,but the disadvantage is that the current recognition system lacks reasoning ability.One of the goals of artificial intelligence is to develop machines with similar human logical reasoning capabilities,so it is necessary for us to deeply understand the learning and reasoning in machines.How to make computer learning possess logical reasoning ability similar to human is a very important research content.The current artificial intelligence reasoning technology is still immature,and it is more difficult for machines to reason directly.It is often necessary to provide prior knowledge or a given prior reasoning mode for reasoning.In response to this shortcoming,we propose logic learning to study machine inference problems.The purpose is to enable machines to directly learn logic patterns from data without prior design or only a small amount of prior knowledge.Intelligence quotient(IQ)testing is one of the most commonly used methods to define and test human computing and logic understanding capabilities.We want to use IQ test questions to conduct logic learning research,so can machines directly learn logic from IQ questions?Based on this problem,this article has carried out the following research:(1)A visual logic learning task of image description IQ problem is proposed.The purpose is to allow the computer to directly learn the logical mode between IQ images without knowing the relationship between the images and the meaning of the content contained in the images.Defined and formalized the IQ problem visual logic learning task,and designed a complex IQ problem data set Fashion?IQ for this task.Examples of each problem include three contextual problem pictures and four candidate answer pictures,of which three are context question pictures are generated sequentially according to a certain transformation,and the computer needs to learn the logical transformation mode between the context pictures to reason and select the correct candidate answer.(2)The feasibility of the visual logic learning task of image description IQ problem is verified.Experiments are conducted on three traditional neural network models to explore the logic patterns existing between IQ problem images for reasoning without designing any prior reasoning models for the neural network.Through the analysis of the inference accuracy of different models on the IQ problem of different transformations of the data set Fashion?IQ,the experimental results show that the neural network can learn the logical pattern between the IQ problem images through training and the correct reasoning is driven by pure data.s answer.This verifies the feasibility of the visual logic learning task of image description IQ problem.(3)A new method-time-series relational network is proposed to solve the problem that the existing neural network model performs poorly on the visual logic task of image description IQ problem.Based on the temporal characteristics of IQ problems,this paper proposes a logical reasoning model of temporal relational network.The model draws on the idea of LSTM processing time series data,and uses relational network module to extract the spatial characteristics of IQ problem images and LSTM extracted by CNN.The time sequence feature of the IQ problem image is extracted and combined to obtain the time sequence-spatial relationship feature for further reasoning.Through experimental comparison and verification,the logical reasoning method proposed in this paper has achieved the best performance in solving the visual logical reasoning task of the image description IQ problem.The IQ problem Turing test was carried out.By comparing the temporal relational network model with the inference accuracy of human subjects when doing the complex IQ problems designed in this article,it shows that the machine can achieve close to humans and even surpass humans in some IQ problems.This article uses neural networks to learn and reason similar to human IQ test questions as an example to explore the use of machines to learn and reason logical patterns between images without a priori pattern,and verifies that logical learning uses a purely data-driven method to learn logical patterns.The feasibility of reasoning,a new logical reasoning task is designed for visual logic learning,and new ideas and directions are provided for the research direction of logical reasoning by machines.This has certain significance and application value for the research of machine reasoning tasks.