Why bigger is not always better: on finite and infinite neural networks

The lack of representation or equivalently kernel learning leads to less flexibility and hence worse performance.

Video

References

Why bigger is not always better: on finite and infinite neural networks