Lucd - Forbes NVIDIAVoice Channel
Updated: Oct 8, 2018
Data comes before AI: Understand the difference between training and inference.
Training an AI model can be likened to a child learning a language. Most children learn language from countless hours of listening to their parents and everyone else around them talk. They consume massive amounts of “data” over time and gradually learn the language. Depending on the time, the inputs, and the child’s effort and aptitude, certain language proficiencies are achieved.
Like a child, the AI model needs an objective. Then, it is developed by exposing the model to training data that reinforces the objective. With the right type of data and volume of data, (potentially “Big Data”), a sufficient amount of time and processing power, a certain level of proficiency (accuracy) can be achieved. Only then, like a child who learned how to speak and understand a particular language can they then utilize this knowledge to act and play. Then the AI model can be used as part of a business capability or process.
Once trained, an AI model and the capabilities that it powers, can then be used in production, i.e. leveraged as part of a new or better product or service (i.e. think Alexa or Siri), a better customer interaction (think chatbots), or a new or better way of doing business (think internet of things and predictive maintenance). This is called “inference." There are numerous pre-trained models and systems that can be leveraged, acquired, or purchased. And if you can find a model that specifically meets your needs, the training phase is already completed. So, if you purchase a self-driving car, that car is constantly using “inference” to make driving decisions and take actions. But, someone else trained that AI system to do that.
The challenge for a business is that their differentiation may come from combining data from external sources with data which is owned by the business itself. In addition, the objective associated with the business value that the business can create may be different from that of a pre-trained system. There is active research around “transfer learning”, i.e. applying the training performed in one scenario to a different scenario. However, ....