In the long march of humankind, there have been numerous changes. Technological advancements do play a crucial role in the march so far. Artificial intelligence and cyborg are few of the greatest achievements of humankind in the field of technology. IBM Watson is now being trained to make it much smarter and accurate.
It is definitely a time consuming and an extensive process to retrain the services, so as to get accurate results. Primarily because, most of the time, developers manually extract data from systems to determine how to change their training sets. This can get even more complicated when using something like a visual classifier since images need to be cross-referenced with their classifications and potentially even put into the context of the situation.
Nevertheless, IBM Watson is using a new approach to streamline this process and continuously makes the system better without taking away too much time from development.
The entire system is reliant upon a three-step process of collecting data, reviewing it, and retraining services. In the first step, data is logged from each of the services in the form of text, audio, and images. All of that data is automatically organized and cross-referenced in a way that makes it easier to manually view later if the user decides to do so. The second step is the review step. By utilizing a quiz-based system, the process of reviewing data is made much quicker and more enjoyable. For example, reviewing data from the visual recognition service might involve showing the user a photo, what it was classified as, and how high the confidence was for that classification. The user just has to say whether that photo was classified correctly or incorrectly. If it wasn’t, then the user specifies the correct classification, and that data is used by the system to aid in the retraining process. After the review session, the user is presented a number of statistics on the results of the review process so they can see how accurately their services are performing down to a per-classification level. The third step is where the services are retrained. All of that new data that is created through the review process can now be used to enhance the ground-truth of the system through retraining. Because all of the data brought in by services is real-world data, it is generally much better for use in training than data that is made up.
A good way to think of this is as though Watson is learning through real-world interaction instead of fabricated interaction. The whole point of adopting a new approach to retraining the system is to cut down the efforts and time it consumes.