Environmental drivers of systematicity and generalization in a situated agent

By Felix Hill, Andrew Kyle Lampinen, R. Schneider, Stephen Clark, Matthew Botvinick, James L. McClelland, Adam Santoro. 2019

111 citations recorded by the source corpus.

The question of whether deep neural networks are good at generalising beyond their immediate training experience is of critical importance for learning-based approaches to AI. Here, we consider tests of out-of-sample generalisation that require an agent to respond to never-seen-before instructions by manipulating and positioning objects in a 3D Unity simulated room. We first describe a comparatively generic agent architecture that exhibits strong performance on these tests. We then identify three aspects of the training regime and environment that make a significant difference to its performance: (a) the number of object/word experiences in the training set; (b) the visual invariances afforded by the agent's perspective, or frame of reference; and (c) the variety of visual input inherent in the perceptual aspect of the agent's perception. Our findings indicate that the degree of generalisation that networks exhibit can depend critically on particulars of the environment in which a given task is instantiated. They further suggest that the propensity for neural networks to generalise in systematic ways may increase if, like human children, those networks have access to many frames of richly varying, multi-modal observations as they learn.

Subfields: Artificial Intelligence, Cultural Studies, Reinforcement Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Cultural Studies

In Reinforcement Learning

Citations in the corpus

Learning Transferable Visual Models From Natural Language Supervision (2021, background): “, 2015) have demonstrated exciting emergent behaviors such as systematically accomplishing zero-shot tasks (Hill et al., 2019).”
The Next Decade in AI: Four Steps Towards Robust Artificial Intelligence (2020, background): “Likewise, as one recent paper (Hill et al., 2019) put it, “the degree of generalisation that networks exhibit can depend critically on particulars of the environment in which a given task is instantiated” § The idiosyncrasy and inability to extrapolate beyond a training distribution is at odds with…”
Imitating Interactive Intelligence (2020, background): “…studying how to build agents that understand the “grounding” of language (Harnad, 1990) within their sensorimotor embodiment is both fundamentally interesting (Hill et al., 2019a; McClelland et al., 2019) and of obvious use for building robotic and other personal assistant artificial intelligences.”
On the Binding Problem in Artificial Neural Networks (2020, background): “These failures at systematically reusing knowledge suggest that neural networks do not learn a compositional knowledge representation (although some mitigation is possible (Hill et al., 2019, 2020)).”
The MAGICAL Benchmark for Robust Imitation (2020, background): “[22] previously found that an egocentric view improved generalisation to unseen instructions or unseen visual objects, despite the fact that it introduces a degree of partial observability to the environment.”
Towards mental time travel: a hierarchical memory for reinforcement learning agents (2021, methodology): “For example, the visual encoder for the ballet tasks is set to have a ﬁlter size of 9 because this is the resolution of each square in the grid, and the entropy cost for the ballet tasks was chosen from our prior work [18] which used a similar grid world action space.”
A Benchmark for Systematic Generalization in Grounded Language Understanding (2020, background): “2017a; Hill et al., 2019), but they do not speciﬁcally test for a model’s ability to perform rule-based generalization, or do so only in limited contexts.”
Grounding Language in Play (2020, background): “Prior work typically studies restricted observation spaces (e.g. games (Goyal et al., 2019), 2D gridworlds (Yu et al., 2018)), simpliﬁed actuators, (e.g. binary pick and place primitives (Hill et al., 2019)), and synthetic language data (Hermann et al., 2017).”
Using cognitive psychology to understand GPT-3 (2022, background): “The upside of our findings is the recommendation that to create more intelligent agents, researchers should not only scale up algorithms that are passively fed with data but instead let agents directly interact and engage with the world (39, 40).”
Mind's Eye: Grounded Language Model Reasoning through Simulation (2022, background): “…work has tried to augment text inputs to LMs with audio signals (Tsai et al., 2019) or visual perception (Bakhtin et al., 2019; Yi et al., 2018), for improved game playing (Lee et al., 2022), faster language learning (Hill et al., 2020), or better decision making in general (Reed et al., 2022).”
Are Neural Nets Modular? Inspecting Functional Modularity Through Differentiable Weight Masks (2020, background): “It has repeatedly been argued that NNs lack compositionality due to their failure at systematic generalization (Lake & Baroni, 2018; Bahdanau et al., 2019; Barrett et al., 2018; Hupkes et al., 2020; Hill et al., 2020).”
Placing language in an integrated understanding system: Next steps toward human-level performance in neural language models (2020, background): “77, an agent was trained to identify, lift, carry, and place objects relative to other objects in a virtual room, as specified by simplified language instructions.”
Extending Machine Language Models toward Human-Level Language Understanding (2019, background): “In (74), an agent was trained to identify, lift, carry and place objects relative to other objects in a virtual room, as specified by simplified language instructions.”
Graph Density-Aware Losses for Novel Compositions in Scene Graph Generation (2020, background): “This argument aligns well to the works from other domains [14], showing that generalization strongly depends on the diversity of samples during training.”
Word meaning in minds and machines (2020, methodology): “After learning not with respect to dozens of object types, generalization to new objects is below 50% correct in a 2D grid world (20 training object types), and somewhere between 60% and 80% in a richer 3D environment (Hill, Lampinen, et al., 2020).”