According to this post, "neural networks are a special form [of a computational graph]". I think that one can infer from that that all neural networks are computational graphs.

My question then would be: are there computational graphs that aren't neural networks?


a computational graph is made of node where is done "operation" on incoming variables. (see first paragraph from the link done in OP)

A neural network use perceptron or "neuron" model for each node. a generic example of neuron model is : Each incoming value is multiply by a "synaptic" weight, then are sum and the result will be pass to an activation function. Which will give an output result that could be project to next nodes. Neural Network can "learn" by adapting synaptic weights thanks to methods such as backpropagation or direct feedback alignment.

Any computational graph where nodes are not a kind of neuron/perceptron model are not neural network.


Any computational graph that does not yield neural network functionality is a computational graph but not a neural network.

Such as this image from the link you have provided:

enter image description here

If you consider this directed graph as it is, then you have a computational graph. Being a neural network requires a goal that the network has to reach gradually.

  • $\begingroup$ I`m not convinced that a goal (or loss function, if that is what you meant) is required for a neural network to be defined. Could you provide a reference? $\endgroup$ – Lay González May 4 '18 at 13:19

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.