A Lagrangian Duality Approach to Active Learning

Authors: Juan Elenter, Navid Naderializadeh, Alejandro Ribeiro

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the benefits of our approach in a variety of classification and regression tasks
Researcher Affiliation Academia Juan Elenter University of Pennsylvania elenter@seas.upenn.edu Navid Naderi Alizadeh University of Pennsylvania nnaderi@seas.upenn.edu Alejandro Ribeiro University of Pennsylvania aribeiro@seas.upenn.edu
Pseudocode Yes Algorithm 1 Primal-dual constrained learning (PDCL). Algorithm 2 Active learning via Lagrangian duality (ALLY).
Open Source Code No The paper's reproducibility checklist states that code is included, but the main text does not provide a direct URL or explicit statement (e.g., 'We release our code at...') for the authors' own implementation code. It mentions adopting baselines from a PyTorch implementation [63], but not for ALLY itself.
Open Datasets Yes In the classification setting, we use standard datasets that commonly appear in the active learning literature, namely STL-10 [54], CIFAR-10 [55], SVHN [56] and MNIST [57]. Lacking an established benchmark regression dataset for active learning, we evaluate ALLY on the Parkinsons Telemonitoring dataset (PTD) [58].
Dataset Splits No The paper mentions an 'initial labeled set' and evaluating on a 'test set', but does not specify explicit training/validation/test dataset splits (e.g., percentages or sample counts for each split).
Hardware Specification No The paper does not provide specific details about the hardware used for experiments, such as GPU models, CPU types, or cloud computing resources.
Software Dependencies No We adopt the PyTorch [62] implementation of the baselines from [63]. The paper mentions PyTorch and libact but does not specify their version numbers.
Experiment Setup Yes Input: Labeled dataset L, primal learning rate ηp, dual learning rate ηd, number of PDCL iterations T, number of primal steps per iteration Tp, constraint vector ϵ. ... In all experiments, the initial labeled set L0 consists of 200 randomly drawn samples, and the budget is set to either b = 200 or b = 1000. We use a ResNet-18 architecture [59] with an embedding size of 128. ... We use a multi-layer perceptron (MLP) with two hidden layers, each with 256 neurons and rectified linear unit (ReLU) activation, leading to an embedding size of 256. The dual regression head fω is a MLP with 5 hidden layers, ReLU activations and batch normalization. ... we cluster the embeddings of the unlabeled samples...using the k-MEANS clustering algorithm [60], where k = b is a hyperparameter. ... we set ℓ ( , ) = ℓ( , ). We use the cross-entropy loss for both ℓ( , ) and ℓ ( , ) and set ϵ(x) = 0.2, x. ... We use mean-squared error for both ℓ( , y) and ℓ ( , y) and set ϵ(x) = 0.1, x.