Research On Pure Data Driven Logic Learning Of Visual Digital Sequence

Logical reasoning has always been a major research difficulty in the field of artificial intelligence.It needs to define some reasoning pattern in advance.But the study of logical reasoning is easy to get into trouble without the prior knowledge.As far as human beings are concerned,although they have certain logical reasoning ability,they can use mathematical background knowledge for reasoning and learning,but if they directly mine the logical relationship which is hidden in the complex data without prior knowledge,they can not find the internal logical relationships quickly,or even can not carry out logical pattern reasoning.In view of the shortcoming in the current research field,the thesis proposes purely data-driven logical learning,which aims to enable the computer to autonomously learn logical pattern in data without prior knowledge.And sequence logic learning as a representative task of logical learning,how to accurately analyze and predict sequence has become a hot topic for researchers.The thesis uses the visual digital sequence as an example to explore the feasibility of mining the logic relationship hidden in the data by using the data-driven method.The main research contents are as follows:(1)Two large datasets are constructed for the logic learning task of image number sequence and the prediction task based on image sequence of arbitrary position,which provide data support for the subsequent experiments.Each large dataset is divided into four different types based on its solutions,namely: ‘Linear'sequence,‘Multiplication'sequence,‘Fio' sequence and ‘Nested'sequence.And the dataset contains various image sequences,and all of these images are black and white images with picture size of 15×85,align right,and size-normalized.The maximum number of digits in each image is set to 10,and can be positive or negative number.(2)The traditional sequence solving technique is based on the known numerical meaning and rules,and through a lot of background knowledge to build the pattern,then predicts subsequent numbers.But for unknown numbers and rules,it will be invalid.Aiming at the problems of pattern construction of such unknown numbers and rules,a logic learning task of image number sequence is proposed.And tests on four typical deep neural networks(MLP,LSTM,CNN-MLP,Res Net),so that the computer can autonomously learn the inherent logical pattern without prior knowledge of the meaning of the content in the image or of the relationship between images,so as to predict the subsequent content of image sequence.The experimental result shows that the method proposed in the paper can solve the problem of pattern construction of unknown numbers and rules to a certain extent.(3)Most of the existing researches on the prediction of number sequence are aimed at the prediction of subsequent numbers.However,the prediction of arbitrary position sequence has great significance and application value for human beings,artificial intelligence,production and life.Such as intelligence test,game prediction,cryptography,etc.In this paper,the problem of sequence prediction at any position is studied experimentally,instead of being limited to predicting the content of subsequent image.And a new idea of image position limitation is provided by designing a framed empty image to limit the position of the image sequence,instead of using the traditional one-hot encoding.In particular,this paper proposes a simple convolutional neural network model(S-CNN)composed of six convolution layers,which uses the convolution kernel size of 3×5 and combines the sequential logic reasoning task based on deep learning with the perceptual recognition task of optical character recognition(OCR).And tests on six typical deep neural networks(MLP,LSTM,CNN-MLP,Res Net18,Res Net34,S-CNN).The experimental result shows that the proposed model has good result on other types of dataset except ‘Multiplication'sequence,which proves that the proposed model is feasible to predict arbitrary position image sequence to some extent.In the thesis,visual digital sequence is taken as an example to explore the feasibility of mining hidden logical relations in data by using purely data-driven method,which provides a new idea and direction for the study of sequence prediction.In view of the shortcomings of the traditional sequence prediction methods,the sequence logic learning task is designed from two different perspectives.This further verifies the feasibility of logic learning research,which has certain research significance and application value in the field of logic learning.