Reliability of CKA as a Similarity Measure in Deep Learning