Learning position-tolerant object representations