Unsupervised Learning of Overlapping Image Components Using Divisive Input Modulation
(a)–(e) Example basis vectors learnt by each
algorithm when trained on the CBCL Face Database, with . (f) The change during the course of training of the
mean Euclidean distance between the input and the reconstructed input. Results
show mean of 10 trials performed using each algorithm. The best and worst
performance over these 10 trials is shown by the error bars (which are very
small in each case). Note that the total training time used for each algorithm
(and hence the meaning of the value “20" on the x-axis of this graph) was 200
epochs for , 2000 epochs for algorithms and , and 20 epochs for all the other algorithms.
Training times are therefore not all directly comparable: at any particular
time has seen 10 times more data and been updated
10 times more than , while and have seen 100 times more data but been updated 24 times less than .
Article of the Year Award: Outstanding research contributions of 2020, as selected by our Chief Editors. Read the winning articles.