EEG signals based on motor imagery are electrical signals acquired from different brain regions when subjects perform different motor imagery tasks. Influenced by the factors such as the difference of brain structures and scalp states of different subjects, the acquired signals of different motor imagery tasks are usually confused, which will result in the misclassifications for a large number of motor imagery signals. A classification algorithm for motor imagery tasks based on improved deep forest is proposed. Firstly, the powerful optimization ability of the variable-length particle swarm algorithm is adopted to search for the optimal weights for probability values of each class predicted by the random forest and completely random forest in the deep forest. Secondly, the weights are assigned to the corresponding class probability values, so as to realize the purpose of the result correction. Finally, the data set 2a of BCI competition IV is adopted to evaluate the efficacy of the proposed algorithm. The results show that the proposed algorithm achieves a higher classification accuracy for the four classes of motor imagery signals in comparison with the traditional deep forest. The proposed algorithm can learn from the predicted results of the classifier, which is of great significance for the improvement of the classifier performance.