Broad Learning System With Sparse Neuron Selection And Its Application

Deep structure learning is limited by its complicated structure,difficult theoretical analysis,and time-consuming training process.Broad Learning System?BLS?,which is proposed based on the single-hidden layer feedforward network?SLFN?,offers an alternative way to solve the problems.It has the advantages of simple implementation,reduced structure,and fast training speed.So,it has vast application prospects.However,there are still some problems in BLS,such as the randomness of the hidden nodes,the redundancy of the network architecture,and the adaptability to different dimensional data.To address the problems in BLS,this thesis will finish the following works:?1?Structure optimization based on BLSSNS.Since BLS adopts randomized functional-link neural nodes,to guarantee the function approximation capabilities,BLS may have more randomized hidden nodes to maintain information of input data.Considering the possible re-dundant hidden nodes in BLS,in this thesis,the sparse representation method based on the?₁norm regularization was proposed to select the powerful hidden nodes.Furthermore,an efficient and effective method based on the separable surrogate function?SSF?for solving the sparse rep-resentation problem in Broad Learning System with Sparse Neuron Selection?BLSSNS?was proposed.A vast number of numerical experiments on 19 benchmark datasets show that the pro-posed BLSSNS achieves the state-of-the-art performance with fewer hidden nodes than BLS,ELM and SVM.?2?Improvement of sparse representation method.Considering the possible over-penalization of the?₁penalty,the sparse broad learning system based on?₀and?₁norm regular-ization has been proposed from a novel Bayesian perspective.It solves the problems of unstable solution of the system of equations under the constraint of?₀norm regularization and over-penalization of?₁norm regularization.The solution of the proposed algorithm achieves more sparse results than?₁norm regularization.In addition,the SSF algorithm is extended to solve the?₀and?₁norm regularization problem.Lastly,the experimental results show the sparseness of the proposed method.?3?Inspection of the adaptability of BLSSNS and enriching the application of system.In order to test the adaptability of BLSSNS on different dimensional data,the author further uses a face dataset to test the face gender recognition performance of BLSSNS.In this thesis,the algorithm based on the BLSSNS and fusion feature with singular value decomposition?SVD?has been proposed for the application of gender classification.In this algorithm,after dividing the image into blocks,the singular value vector from both the whole image and the partial blocks are obtained.After that,the global and local singular value vectors are combined as the fusion features.Then the fusion features are used to train the BLSSNS for gender recognition.The results show that the fusion feature with more blocks,the BLSSNS has better performance.What's more,we have conducted some experiments on gender dataset to confirm the efficiency of the fast incremental learning algorithm of BLS.