Training is the way to add new skills to AI. During training, data containing information about the new skill and the desired answer is provided to the AI.
In the training process, data is first fed to the AI, which then produces an answer. The correctness of this answer is calculated and used to update the model parameters during the backpropagation process. This cycle is repeated until the desired level of correctness is achieved or the model can no longer learn more with the available data. In the latter case, the model’s architecture, used features, and provided data are investigated. Possible modifications to the model architecture, such as adding task-specific layers, are made. Used features are redefined, or new features are added. The quality or amount of data is increased. To succeed in the training process, the domain must be fully understood.
At the moment, many so-called foundation models are available. A foundation model possesses the basic knowledge needed to perform a task. For example, in an LLM, the foundation model understands language and its meaning but doesn’t have information about a specific domain. Then, only specific domain information is needed to train the model. The use of a foundation model simplifies the training process and significantly reduces the amount of data required.
The training of an AI model begins with understanding the domain. The training objective is defined based on the skill to be trained. Examples of training objectives include classification (determining if something is of good or bad quality) or predicting the next word. Task-specific parameters and features are then defined and implemented in the model. Parameters are values where knowledge is stored inside the model, and features are measurable properties of the input data that describe the task at hand. These features are used to achieve the training objective. Training data is collected and prepared into training, validation, and evaluation datasets. The training method and learning parameters are defined, along with the criteria for validation success. Training loops are run iteratively using the training data, with periodic validation using the validation data and updates to the learning parameters. Upon completing the training, the new skill is evaluated with the evaluation dataset. If the evaluation meets the success criteria, the model is deployed to production; otherwise, training continues by evaluating the training performance and the training process used.
