What occurs in the infinite-width NTK regime?

Under the classical Neural Tangent Kernel infinite-width limit and suitable parameterisation, which description is most accurate?

The NTK becomes the identity matrix, so every training example is learned independently

The NTK converges to a deterministic limiting kernel and remains approximately constant during gradient-based training

The network's hidden representations continue to change substantially, while only its final layer behaves like a kernel method

The network is equivalent to a Gaussian process throughout training because its parameter posterior remains Gaussian