TopHat is not able to map to novel non-canoncal introns (without GT-AG, GC-AG or AT-AC). The reason why you are detecting non-canonical introns must be due to the presence of non-canonical introns the transcriptome annotation that you are using. So there are 2 easy solutions to solve this:

1. Pre-filter the transcriptome annotation file in order to exclude all the transcripts that have non-canonical introns.
2. Do not use the reference-guied approach and lets TopHat do a ab intio detection of splice junctions.

Now, let me tell you that we have done a comprehensive analysis of non-canonical splice sites in the human transcriptome. You probably want to exclude non-canonical introns because you think that they are source of alignment artifacts. In fact the most of the non-canonical introns that we initially detect are artifacts, but through a set of a very stringent filters we build a high confident catalog of non-canonical introns present in the human transcriptome, we found interesting features associated to them and we even validate some of these by RT-PCR. If you're going to do a genome-wide analysis yo can ignore them, but if you want to properly annotate a transcriptome you should consider them. Our work is under revision, but if anybody wants to know more, can email us.

Cheers,