Towards Interpretable Reinforcement Learning Using Attention Augmented Agents

By Alex Mott, Daniel Zoran, Mike Chrzanowski, Daan Wierstra, Danilo Jimenez Rezende. 2019

213 citations recorded by the source corpus.

Inspired by recent work in attention models for image captioning and question answering, we present a soft attention model for the reinforcement learning domain. This model uses a soft, top-down attention mechanism to create a bottleneck in the agent, forcing it to focus on task-relevant information by sequentially querying its view of the environment. The output of the attention mechanism allows direct observation of the information used by the agent to select its actions, enabling easier interpretation of this model than of traditional models. We analyze different strategies that the agents learn and show that a handful of strategies arise repeatedly across different games. We also show that the model learns to query separately about space and content (`where' vs. `what'). We demonstrate that an agent using this mechanism can achieve performance competitive with state-of-the-art models on ATARI tasks while still being interpretable.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Reinforcement Learning, Safety, Robustness & Interpretability

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 Computer Vision and Pattern Recognition

In Reinforcement Learning

In Safety, Robustness & Interpretability

Citations in the corpus

Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges (2021, background): “…(i.e., posthoc approximations), including tree-based explanations (Coppens et al., 2019; Liu et al., 2018), reward decomposition (Juozapaitis et al., 2019), autonomous explanation (Fukuchi et al., 2017; Hayes and Shah, 2017), and attention-based methods (Zambaldi et al., 2018; Mott et al., 2019).”
On the Binding Problem in Artificial Neural Networks (2020, background): “For example, Mott et al. (2019) incorporates a form of dot-product attention (Vaswani et al., 2017) in an agent to attend to the internal feature maps of a bottom-up convolutional neural network that processes the input image.”
Exploratory Not Explanatory: Counterfactual Analysis of Saliency Maps for Deep Reinforcement Learning (2019, methodology): “SeaQuest Perturbation Y Y Mott et al. (2019) “The most dominant pattern we observe is that the model learns to attend to task-relevant things in the scene.”
Generalization Through Hand-Eye Coordination: An Action Space for Learning Spatially-Invariant Visuomotor Control (2021, background): “Recent works have attempted to equip interactive agents with attention [17, 24, 34, 41].”
On Identifiability in Transformers (2019, background): “work is not to put a hold on attention as explanation since insights into deep learning systems can be gained by looking at the learned attention distributions of some architectures (Xu et al., 2015; Mott et al., 2019). However, interpretations should be drawn carefully and should be 2 Preprint. Wor”
Neuroevolution of self-interpretable agents (2020, unknown): “Most recently, [67] used a soft, top-down attention mechanism to”
Towards Robust Image Classification Using Sequential Attention Models (2019, methodology): “In this work we propose to use a soft, sequential, spatial, top-down attention mechanism (which we abbreviate as S3TA) [38], drawing inspiration from the primate visual system.”
Causal Analysis of Agent Behavior for AI Safety (2021, background): “Other recent approaches for analyzing AI systems have mostly focused on white-box approaches for improving understanding (see for instance Mott et al., 2019; Verma et al., 2018; Montavon et al., 2018; Puiutta & Veith, 2020) and developing safety guarantees (Uesato et al., 2018).”
Visual Attention Methods in Deep Learning: An In-Depth Survey (2022, methodology): “[56] proposed soft, sequential, spatial top-down attention method (S3TA) to focus more on attended regions of an image [157] (as shown in Figure 8 (b)).”
Goal-Aware Cross-Entropy for Multi-Target Reinforcement Learning (2021, methodology): “Other methods include the use of encoding information of sequential frames up to the last step [29], as well as the application of self-attention [6].”
Multi-Task Reinforcement Learning with Context-based Representations (2021, background): “…learning holds the promise of accelerating learning across multiple tasks by sharing useful information (Caruana, 1997; Zhang et al., 2014; Kokkinos, 2017; Radford et al., 2019; Rajeswaran et al., 2016; Ruder, 2017; Liu et al., 2019a; Mott et al., 2019; Vithayathil Varghese & Mahmoud, 2020).”
Unsupervised Object Keypoint Learning using Local Spatial Predictability (2020, methodology): “In particular, the same (overcomplete) set of keypoints can now be used to faciliate a number of tasks, for example by querying (and attending to) only keypoints relevant for the particular task using a top-down attention module as explored in Mott et al. (2019).”