A Mechanism of Visual Concept Learning via Neuronal Random Projection
Koplik, Peter Sebastian
MetadataShow full item record
We explore the effectiveness of the random projection method, a biologically plausible, computationally efﬁcient,and data-independentmethodofdimensionalityreductionindistinctionbetweencategoriesofvisualstimuli. We observe that a neural network tasked with approximating the original stimulus from the reduced domain generally excludes information not useful in distinguishing visual categories. This suggests that random projection may be useful in the efﬁcient recall and recognition of visual concepts even though the projections only contain small fractions of the original information. Our ﬁndings indicate that the reconstruction of visual stimuli from the random projected domain preserves best the features most typical of that particular category of stimuli.