![]() The question remains open and I guess only the future can tell where AI will lead us to. For example, when AI is left to make decisions affecting people on a personal level. In the case of Hubble this wasn’t a problem- but we can easily imagine a situation where the impact may be huge. AI, on the other hand, is like a magic box: you feed it a set of examples and you get a set of results, without any insight into how the input triggered the particular output. With traditional code, you can always look into it and understand how ‘the system’ arrived at the feedback it gave. At the same time we’re aware of the risks. we’re very keen to investigate and implement the possibilities it offers. ![]() It can’t be denied AI is a huge opportunity, and at Flock. We even started to wonder whether it will make us, software engineers, obsolete? Will our work be fully replaced by AI, will we step in solely as a Human-in-the-Loop? AI technology turned out to be a better option than the traditional approach in this case. That was a reason to celebrate but it also left us somewhat puzzled. After we’d adjusted this mistake, the system returned correct results in 100% of the cases. It turned out to have been a simple human mistake - out of the 1000 invoices we fed into the model, we labelled one licence plate wrong by accident. The output had an amazing 99% accuracy, a result we were thrilled with, but obviously we were still eager to understand where the incorrect one percent came from. Subsequently, we fed all of them into Google Document AI (Entity Extraction Model) and trained the system to extract what we needed. We proceeded with labelling the licence plates on 1000 invoices. if we didn’t try the latest hype: an AI solution. That was the moment we realised we needed a novel approach. Needless to say, we weren’t quite happy with the result as yet.Īt a closer look we realised that our code identified certain postal codes and IBAN numbers as number plates, too, returning the false positives which accounted for the 20% the programme missed. 80% isn’t enough, and to arrive at 100% manual intervention was required. A simple adjustment in the code did the job, we ran the tests and arrived at 80% accuracy. In addition to tracking changes to software code, managing a product’s configuration is also necessary. It’s true that some number plates came with - and some without - dashes, but that wasn’t really the challenge. Software life cycle management has also improved in recent years. After all, it was basically about simple pattern recognition. When Hubble asked us to write a programme to detect license plates on the various invoice formats, we expected the task would be quite straightforward.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |