Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
Variational Information Pursuit for Interpretable Predictions
Authors: Aditya Chattopadhyay, Kwan Ho Ryan Chan, Benjamin David Haeffele, Donald Geman, Rene Vidal
ICLR 2023 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, V-IP is 10-100x faster than IP on different Vision and NLP tasks with competitive performance. In this section, through extensive experiments, we evaluate the effectiveness of the proposed method. |
| Researcher Affiliation | Academia | Aditya Chattopadhyay, Kwan Ho Ryan Chan, Benjamin D. Haeffele, Donald Geman Johns Hopkins University, USA, EMAIL Ren e Vidal University of Pennsylvania, USA, EMAIL |
| Pseudocode | No | No structured pseudocode or algorithm blocks with explicit labels were found. |
| Open Source Code | Yes | Code is available at https://github.com/ryanchankh/Variational Information Pursuit. |
| Open Datasets | Yes | CUB-200 (Wah et al., 2011) ... Huffington News (Misra, 2018) ... MNIST (Le Cun et al., 1998), KMNIST (Clanuwat et al., 2018), Fashion-MNIST (Xiao et al., 2017b), CIFAR-{10,100} (Krizhevsky et al., 2009) ... Sym CAT-200 (Peng et al., 2018) ... Mu Zhi (Wei et al., 2018) ... Dxy (Xu et al., 2019). |
| Dataset Splits | No | No explicit validation set split percentages or counts were provided. The paper mentions '60,000 training images and 10,000 testing images' for MNIST datasets and '50,000 training images and 10,000 testing images' for CIFAR, but does not specify a validation split from the training data. |
| Hardware Specification | Yes | All of our experiments are implemented in python using Py Torch (Paszke et al., 2019) version 1.12. Moreover, all training is done on one computing node with 64-core 2.10GHz Intel(R) Xeon(R) Gold 6130 CPU, 8 NVIDIA Ge Force RTX 2080 GPUs (each with 10GB memory) and 377GB of RAM. |
| Software Dependencies | Yes | All of our experiments are implemented in python using Py Torch (Paszke et al., 2019) version 1.12. |
| Experiment Setup | Yes | General Optimization Scheme. The following optimization scheme is used in all experiments for both Initial Random Sampling and Subsequent Biased Sampling, unless stated otherwise. We minimize the Deep V-IP objective using Adam (Kingma & Ba, 2014) as our optimizer, with learning rate lr=1e-4, betas=(0.9, 0.999), weight decay=0 and amdgrad=True (Reddi et al., 2019). We also use Cosine Annealing learning rate scheduler (Loshchilov & Hutter, 2016) with T max=50. We train our networks fθ and gη for 500 epochs using batch size 128. In both sampling stages, we linearly anneal temperature parameter τ, in our straight-through softmax estimator, from 1.0 to 0.2 over the 500 epochs. |