What is the Artificial Intelligence I am studying?

By Richard’s definition, to build the model that has the ability to achieve goals by adapting behaviors. The model is a structured formula, with weights being differentiable. The dataset contains the examples of the behavior + goals, and the model takes the dataset and is measured by the proxy of task completion via a loss function. The loss function provides the signal to guide the weights to a point that optimize the value of loss function via gradient descent.  While the ability to achieve the goals is measured by given any possible data, in-domain or out-domain, the model is expected to achieve it by the measurement of the loss value or the discrete completion of tasks. The adapting of behaviors involves the change in the weight to try to keep the loss low for new data, while maintaining the ability on already seen data points. The training dynamics of change of weight triggered by coming data points and the regularization term. The whole training process leads the model weights to a training loss landscape that is expected to be able to generalize to other datapoints that are from the in-domain distribution. One way to measure the model is to sample from the in-domain distribution, and test on its performance, with the hope that the good performance indicates the model has learned the principle of the task, with built representation, and representation processing. One could expect the built process of the model is the same as what humans are doing, or provides new patterns that are innovative for humans to learn. The check on the exact way requires feature identification on the intermediate output, or a controlled dataset to analysis model’s behavior via the model behavior.  For the out-domain task, the model is expected to adjust its distribution of training data, where the reliable reward signal is expected to be given, whether through a reliable answer to measure the discrepancy of behavior, or a direct reliable reward signal to guide the weight change.  The type of artificial intelligence can be divided by the goals and the behaviors. NLP for language tasks, CV for vision tasks, Agentic for language with image input and language as output, Robotics for vision, language as input physical action as output. Behaviors divide the study by text for NLP, images for CV, text actions for agentic, physical actions for Embodied intelligence.Different forms of models are proposed with inductive biases to fit the completion of goals in different types. Models come with different architecture, where as for a deep model, each block consists of different units. Models come with different loss function, where the terms of the loss function are added to control models’ training dynamics movements, also the landscape and expected landing zone for one training. The training dynamics, involves the trace of the weight movement on the training loss landscape, where it is controlled by the loss function and the gradient direction as well as the data batch for calculating the loss. The gradient direction involves the configuration of the optimizer, as there is the risk of falling into one local minimum, or keep orchestrating without converging to a saturation point. The data batch forms the batch loss landscape, and can provide partial information about the whole training loss landscape, while trading off the effect of falling into a local minimum in the training loss landscape, or to a point that is genuinely good on the real loss landscape. Also, training here is optimizing the weights on the training dataset loss landscape, with the expectation that the landscape is similar to the real goal landscape. Hence, further behavior adaptation involves the change of landscape if it is seeing more new data, and the adaptation needs the movement of the weights. As there are random variables for the input, the output, the accurate modeling of the real input distribution and a conditioned output distribution, provides the probability view of goal completion and model behaviors. Then, the distribution provides more angles to control the behaviors and measure the goal completion. It is also the inductive bias to inject via distribution type, the prior on the hyperparameters, and moreover, to frame the problem in a data trainable way. It allows differentiable parameters, and to be guided to change via data points measured by the loss function.