TracIn
A gradient-based method for estimating the influence of training samples that approximates a sample's influence on model outputs by computing the inner product between the gradient of the training sample and that of the test sample.
Timeline (last 90 days)
Data attribution aims to quantify the contribution of each sample in the training data to a given prediction of the model, and common methods include influence functions, gradient similarity (such as TracIn) and leave-one-out.
The classic gradient attribution method TracIn estimates influence by computing the inner product between the gradient of a training sample and that of a test sample, but it requires a separate backpropagation for each sample, and its computational complexity grows linearly with data size, making it hard to apply to million-scale datasets.
All Facts (2)
Data attribution aims to quantify the contribution of each sample in the training data to a given prediction of the model, and common methods include influence functions, gradient similarity (such as TracIn) and leave-one-out.
50%UnverifiedThe classic gradient attribution method TracIn estimates influence by computing the inner product between the gradient of a training sample and that of a test sample, but it requires a separate backpropagation for each sample, and its computational complexity grows linearly with data size, making it hard to apply to million-scale datasets.
50%