Instance-Dependent Noisy Label Learning via Graphical Modelling

PASS: Peer-Agreement based Sample Selection for training with Noisy Labels

“…This issue has prompted the development of innovative learning algorithms aimed at tackling the problem of noisy labeling. Within the domain of noisy labels, many methods have emerged [19,39], each tailored to tackle the challenges posed by distinct settings of noise, namely instance-independent noise (IIN) [22] and instance-dependent noise (IDN) [68]. Early studies in the field of noisy label operated under the presumption that label noise was IIN, that is, mislabeling occurred irrespective of the information regarding the visual classes present in images [22].…”

“…In IIN, a transition matrix is generally employed, which comprises a pre-determined probability of flipping between pairs of labels [72]. Nevertheless, recent studies have progressively redirected the field's attention toward the more realistic scenario of IDN [10,19,71], where label noise depends on both the true class label and the image information.…”

“…proaches incorporate robust loss functions [52], designed specifically to operate effectively with either clean or noisy labels, as well as probabilistic modeling approaches that model the data generation process [19,71]. Furthermore, training regularization [43,64] imposes a penalty term on the loss function during training, thereby reducing overfitting and ameliorating generalization.…”

“…Furthermore, training regularization [43,64] imposes a penalty term on the loss function during training, thereby reducing overfitting and ameliorating generalization. Various techniques integrate sample selection strategies as a key algorithmic step [12,13,19,39], allowing for the identification and segregation of noisy and clean labels. One popular criterion for this sample selection process is the loss value between the prediction of the trained classifier and its label, whereby it is typically assumed that the noisy data exhibits a large loss [22,30,32,64] or greater magnitude of the gradient during training [63].…”

“…This sample selection relies on a thresholding algorithm [53] that distinguishes samples based on how high the agreement between the peer classification predictions is. Our proposed method, named as peer-agreement based sample selection (PASS), can readily be integrated into existing models in noisy label learning, such as In-stanceGM [19], DivideMix [39], SSR [18], FaMUS [70], AugDesc [51], and Contrast-to-Divide (C2D) [77]. The primary contributions of our method can be delineated as follows:…”