Attention-based Iterative Decomposition for Tensor Product Representation

Authors: Taewon Park, Inchul Choi, Minho Lee

ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental In experiments, we apply AID to several recent TPR-based or TPR equivalent approaches to show its effectiveness and flexibility.In all experimental results, AID shows effective generalization performance improvement for all TPR models.
Researcher Affiliation Collaboration Taewon Park1, Inchul Choi2, , Minho Lee1,2,* 1Kyungpook National University, South Korea 2ALI Co., Ltd., South Korea
Pseudocode Yes Algorithm 1 Attention-based Iterative Decomposition module.
Open Source Code Yes 1The code of AID is publicly available at https://github.com/taewonpark/AID
Open Datasets Yes SAR task: We use a set of arbitrary 1,000 words to construct each word set, as outlined in Table 15. bAbI task: The b Ab I task (Weston et al., 2015). Sort-of-CLEVR task: The Sort-of-CLEVR task (Santoro et al., 2017). WikiText-103 task: Wiki Text-103 task (Merity et al., 2016).
Dataset Splits Yes Wiki Text-103 task (Merity et al., 2016)... The training set consists of 28,475 articles, while the validation and test sets contain 60 articles each.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running experiments.
Software Dependencies No The paper mentions optimizers (Adam, RMSprop) and model architectures (LSTM), but does not provide specific software dependency details with version numbers (e.g., Python, PyTorch, TensorFlow, CUDA versions) needed for replication.
Experiment Setup Yes SAR task: we utilize the Adam optimizer with a batch size of 64 and a learning rate of 1e 3, β1 of 0.9, and β2 of 0.98 for training iterations of 30K. ... sys-bAbI task: embedding size to 179. ... Adam optimizer with a batch size of 64 and a learning rate of 1e 3, β1 of 0.9, and β2 of 0.99 for 100 training epochs. ... Wiki Text-103 task: Adam optimizer with a batch size of 96, an initial learning rate of 2.5e 4, and a learning rate warmup step of 2,000 for 120 epochs. ... Tables 6, 7, and 8 show our module s hyper-parameter settings for each task.