Pure Data is a visual programming language where the program executes as flows of data between executing operators in a patch, a visual canvas representing the program (a dataflow programming langugage).

While there has been interest in having patches that can run artificial neural networks or even interoperatibility with neural network libraries this idea relate more seeing a neural network working. In that regard, this is closer to a live version of TensorBoard where the operator can add training instances and change the network architecture in real time (no warranties that will lead to a working, well training neural network).

It can be aesthetically pleasing and might help understand how the network works.

Each neuron thus is an object in the patch and its state at a given moment can be seeing in the patch. Similarly, the weights are also objects. The gradients and weights calculations can be triggered through messages. The calculations can be done using libraries such as GridFlow or by building an extension wrapping TensorFlow.

Pure Data was popular many years ago. These days things are very quiet in the space but the HTML5 port might be the way to go. Possibly working with TensorFlow.js.

On a personal note, I have done some work on PD and I enjoyed both the work and the community quite a bit.