Metalearned Neural Memory

Authors: Tsendsuren Munkhdalai, Alessandro Sordoni, TONG WANG, Adam Trischler

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate the effectiveness of our metalearned neural memory (MNM) on a diverse set of learning problems, which includes several algorithmic tasks, synthetic question answering on the b Ab I dataset, and maze exploration via reinforcement learning. Our model achieves strong performance on all of these benchmarks.
Researcher Affiliation Industry Tsendsuren Munkhdalai, Alessandro Sordoni, Tong Wang, Adam Trischler Microsoft Research Montréal, Québec, Canada tsendsuren.munkhdalai@microsoft.com
Pseudocode No The paper contains Figure 1 illustrating the model, but it does not include any blocks or sections explicitly labeled "Pseudocode" or "Algorithm".
Open Source Code No The paper does not provide any statement about making its source code available, nor does it include links to a code repository.
Open Datasets Yes We demonstrate the effectiveness of our metalearned neural memory (MNM) on a diverse set of learning problems, which includes several algorithmic tasks, synthetic question answering on the b Ab I dataset, and maze exploration via reinforcement learning. ... b Ab I is a synthetic question-answering benchmark that has been widely adopted to evaluate long-term memory and reasoning [45]. ... We train MNM agents on a maze exploration task from the literature on meta-reinforcement learning [4, 44]. Specifically, we adopted the grid world setup from [23].
Dataset Splits Yes For the bAbI tasks, ... We used the standard 10K training examples per task and 1K test examples.
Hardware Specification No The paper does not specify any particular CPU or GPU models, memory amounts, or other specific hardware configurations used for running the experiments.
Software Dependencies No The paper mentions the use of "Adam optimizer" and the "Advantage Actor-Critic (A2C) algorithm" but does not provide specific version numbers for any software libraries, frameworks (e.g., TensorFlow, PyTorch), or programming languages.
Experiment Setup Yes For the dictionary task, we train the MNM and LSTM+SALU models for 60K iterations using the Adam optimizer with a learning rate of 10−4. We use a batch size of 20. ... The MNM models have a 3-layer neural memory with 100 units in each layer. ... For the bAbI tasks, we train the MNM model for 100K iterations using Adam optimizer with a learning rate of 10−4 and a batch size of 20. ... All MNM models are trained with 4 parallel read/write heads. ... We used A2C with a discount factor of 0.99 and a learning rate of 7 · 10−4. The controller s hidden size is 200 and the memory s is 100 with 3 layers. The number of heads is 1.