Key application areas include computer vision, natural language processing, forecasting, classification, recommendation engines and clustering. They are all driven by data, of course.
Classification, forecasting and recommendation engines aren’t new, but they are useful for improving processes. Computer vision and natural language processing are tougher, but new deep learning techniques have made these solutions more accessible. Models have been trained using large datasets (ImageNet, COCO) and released for free use.
Take an example: an application that identifies whether a picture contains a hotdog or not. To create this tool, you’ll need a dataset of hotdog pictures and not-hotdog pictures – and lots of them. However, you can drastically reduce the volume of training images needed by applying transfer learning: chopping off the ‘head’ of an open source neural network and retraining that head using hotdog images.
The same technique can be applied to natural language processing. Pretrained word embeddings capture similar word meanings they contain. The word ‘bus’ will be more closely related to ‘train’ than ‘chihuahua’. Word embeddings will capture synonyms by default, reducing your implementation workload.
You’re definitely going to need skilled people to tackle AI challenges, but while it might take some searching, it’s possible to find the right profiles. These include: