tensor parameter

A tensor parameter is a learnable, multi-dimensional array that stores numerical values that a model optimizes during training to control how inputs are transformed into outputs.

In neural networks, parameters include weight matrices, convolution kernels, embedding tables, and bias tensors, each with a specific shape defining how layers connect and compute.

These tensors are initialized from chosen statistical distributions, allocated on a target device, such as a CPU or GPU, and tracked by automatic differentiation systems so that gradients can be computed and used to update their values.