Multigrid Neural Memory
Authors: Tri Huynh, Michael Maire, Matthew Walter
ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our multigrid neural memory architecture on a diverse set of domains. We begin with a reinforcement learning-based navigation task, in which memory provides a representation of the environment (i.e., a map). To demonstrate the generalizability of our memory architecture on domains decoupled from spatial geometry, we also consider various algorithmic and NLP tasks previously used to evaluate the performance of NTMs and DNCs. |
| Researcher Affiliation | Academia | 1University of Chicago, Chicago, IL, USA 2Toyota Technological Institute at Chicago, Chicago, IL, USA. Correspondence to: Tri Huynh <trihuynh@uchicago.edu>. |
| Pseudocode | No | The paper provides mathematical formulations for its components and architectural diagrams, but it does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide any explicit statement or link regarding the availability of open-source code for the described methodology. |
| Open Datasets | Yes | In the second variant, the input is a sequence of twenty 28 28 MNIST images (Lecun et al., 1998). (...) we evaluate its performance on bAbI (Weston et al., 2015a), which consist of 20 question answering tasks corresponding to diverse aspects of natural language understanding. |
| Dataset Splits | No | Training runs for 8 × 10^6 steps with batch size 32. Test set size is 5000 maps. (...) For each task, the dataset contains 10000 questions for training and 1000 for testing. The paper mentions training and testing set sizes but does not specify validation set splits or sizes in the main text. |
| Hardware Specification | No | The paper does not provide specific details regarding the hardware used for running the experiments (e.g., GPU models, CPU types, or cloud computing specifications). |
| Software Dependencies | No | The paper mentions using RMSProp for training and A3C for reinforcement learning, but it does not specify any software versions for libraries, programming languages, or other dependencies required to replicate the experiments. |
| Experiment Setup | Yes | We train each architecture using RMSProp. We search over learning rates in log scale from 10^-2 to 10^-4, and use 10^-3 for multigrid and Conv LSTM, and 10^-4 for DNC. (...) Training runs for 8 × 10^6 steps with batch size 32. |