Grounded Language Learning Fast and Slow
Authors: Felix Hill, Olivier Tieleman, Tamara von Glehn, Nathaniel Wong, Hamza Merzic, Stephen Clark
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We find that, under certain training conditions and with a particular memory writing mechanism, the agent s one-shot word-object binding generalizes to novel exemplars within the same Shape Net category, and is effective in settings with unfamiliar numbers of objects. We further show how dual-coding memory can be exploited as a signal for intrinsic motivation, stimulating the agent to seek names for objects that may be useful later. Together, the results demonstrate that deep neural networks can exploit meta-learning, episodic memory and an explicitly multi-modal environment to account for fast-mapping, a fundamental pillar of human cognitive development and a potentially transformative capacity for artificial agents. We compared the different memory architectures with and without semi-supervised reconstruction loss on a version of the fast-mapping task involving three objects (N = 3) sampled from a global set of 30 (|G| = 30). As shown in Table 1, only the DCEM and Transformer architectures reliably solve the task after 1 109 timesteps of training. |
| Researcher Affiliation | Industry | Felix Hill, Olivier Tieleman, Tamara von Glehn, Nathaniel Wong, Hamza Merzic, Stephen Clark Deep Mind London, UK {felixhill, tieleman, tamaravg, nathanielwong, hamzamerzic, clarkstephen}@google.com |
| Pseudocode | No | The paper describes architectures and methods in prose and with diagrams, but does not include structured pseudocode or algorithm blocks. |
| Open Source Code | No | We leave this possibility for future investigations, which we will facilitate by releasing publicly the environments and tasks described in this paper. |
| Open Datasets | Yes | The objects include everyday household items such as kitchenware (cup, glass), toys (teddy bear, football), homeware (cushion, vase), and so on. We conducted an analogous experiment by exploiting the category structure in Shape Net (Chang et al., 2015). Shape Net contains 3D models of objects with a wide range of complexity and quality. To guarantee that the models are recognizable and of a high quality, we manually filtered the Shape Net Sem dataset... Chang et al., 2015 |
| Dataset Splits | No | The paper discusses training and evaluation, but does not provide specific numerical percentages or sample counts for train/validation/test splits, nor does it cite a predefined split method that specifies these details. For example, it says 'In all training and evaluation episodes, the initial positions of the objects and agent are randomized.' but no explicit split percentages. |
| Hardware Specification | No | The paper does not specify any particular hardware components such as GPU models, CPU types, or memory specifications used for running the experiments. |
| Software Dependencies | No | The paper mentions various software components and algorithms (e.g., Unity game engine, DeepMind Lab, Res Net, LSTM, Transformer, IMPALA algorithm, Adam), but it does not specify their version numbers or other software dependencies with versioning information. |
| Experiment Setup | Yes | Table 3: Agent hyperparameters (independent of specific architecture). The return cost (not discussed in the main text) is used to weight the baseline estimate term in the V-trace loss. Table 4: Hyperparameters for NGU. |