Behavior Is Everything: Towards Representing Concepts with Sensorimotor Contingencies
Authors: Nicholas Hay, Michael Stark, Alexander Schlegel, Carter Wendelken, Dennis Park, Eric Purdy, Tom Silver, D. Scott Phoenix, Dileep George
AAAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we perform an in-depth evaluation of our approach, highlighting its ability to ground sensorimotor concepts in environmental interactions and enable effective reuse of learned representations. We give examples of learned classification and bring-about SMCs, and we quantify the impact of using SMCs in compositional hierarchies in connection with different curricula, considering classification and bring-about concepts. |
| Researcher Affiliation | Industry | Nicholas Hay, Michael Stark, Alexander Schlegel, Carter Wendelken, Dennis Park, Eric Purdy, Tom Silver, D. Scott Phoenix, Dileep George Vicarious AI, San Francisco, CA, USA nick@vicarious.com |
| Pseudocode | No | The paper describes its methods and algorithms in prose but does not include any formal pseudocode blocks or algorithm listings. |
| Open Source Code | Yes | Pixel World and supplementary material containing all curricula, dataset specs, and supplemental figures can be found at https://github.com/vicariousinc/pixelworld. |
| Open Datasets | Yes | We introduce Pixel World (PW), a novel learning and benchmark suite of presently 96 related sets of environments. ... Pixel World and supplementary material containing all curricula, dataset specs, and supplemental figures can be found at https://github.com/vicariousinc/pixelworld. |
| Dataset Splits | No | The paper states that test performance is evaluated on the best performing iteration (as determined on the training set), but it does not specify a separate validation dataset or split for hyperparameter tuning or early stopping. It only mentions a dedicated training set and held-out test set. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running the experiments, such as GPU models, CPU types, or memory. |
| Software Dependencies | No | The paper mentions software components like "Gated Recurrent Unit (GRU)", "natural policy optimization (NPO)", and "RLLab", and cites papers for them. However, it does not specify version numbers for RLLab or any other software libraries, which is necessary for reproducible software dependencies. |
| Experiment Setup | Yes | We run NPO for 200 iterations with a batch size of 2,000 and a maximum trajectory length of 100. We initialize our networks following (Glorot and Bengio 2010), adjusting the output layer s bias to reduce the initial probability of performing a signal, increasing initial expected episode length and exploration. |