On: Every Model Learned by Gradient Descent Is Approximately a Kernel Machine

[alexilliamson]

You only need the "Support Vectors" to make predictions, not the whole dataset.