TKIL: T KERNEL OPTIMIZATION FOR CLASS B …
TKIL, that addresses the imbalances in memory-based incremental learning. 2) The core of TKIL is a novel tangent kernel (GTK) for finite-width neural networks. We mathematically prove that mini-mizing GTK loss associated with TKIL, expressed as the cosine distance of two Jacobian matrices, avoids local optimums and achieves balanced ...