Causal Self-Talk - 42Papers

Home

Discover

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Machine Learning

Artificial Intelligence

Computation and Language

Causal Self-Talk

Explainability Via Causal Self-Talk

We define a set of desiderata that capture both the aspirations of the x-ray artificial intelligence (xai) community and the practical constraints of deep learning.

We describe an effective way to satisfy all the desiderata : train the ai system to build a causal model of itself.

We implement this method in a simulated 3d environment, and show how it enables agents to generate faithful andsemantically-meaningful explanations of their own behavior.

These learned models provide new ways of building semantic control interfaces to artificial intelligence systems.

Authors

Nicholas A. Roy, Junkyung Kim, Neil Rabinowitz

Related Topics

Introduction

We argue that providing explanations of a system s behavior is, in essence, a task of building a of that system.

We thus train the base system to supply a self-model alongside its main representations.

The different form allows the self-model to deliver supplementary utility that the base system does not, e.g.

Providing an interface to external users, and enabling them to understand, predict, and control the base system in convenient ways.

Result

We train agents to model their belief state using messages in a synthetic language form.

We focus on the metrics of faithfulness previously introduced.

To measure this, we run evaluation episodes where, at the start of each trial, we inject a message indicating that the instructed tag is in some random room.

We find that the greatest gain of this strategy is its ability to imbue the self-model with causal faithfulness.

Read the Paper

Read Paper

Bookmarks

Bookmarks

Keep Up With Latest Trending Papers. Computer Science, AI and Machine Learning and more.Subscribe

CC BY

Source arXiv

Machine Learning

Artificial Intelligence

Computation and Language

Causal Self-Talk

Explainability Via Causal Self-Talk

We define a set of desiderata that capture both the aspirations of the x-ray artificial intelligence (xai) community and the practical constraints of deep learning.

We describe an effective way to satisfy all the desiderata : train the ai system to build a causal model of itself.

We implement this method in a simulated 3d environment, and show how it enables agents to generate faithful andsemantically-meaningful explanations of their own behavior.

These learned models provide new ways of building semantic control interfaces to artificial intelligence systems.

Authors

Nicholas A. Roy, Junkyung Kim, Neil Rabinowitz

Related Topics

Introduction

We argue that providing explanations of a system s behavior is, in essence, a task of building a of that system.

We thus train the base system to supply a self-model alongside its main representations.

The different form allows the self-model to deliver supplementary utility that the base system does not, e.g.

Providing an interface to external users, and enabling them to understand, predict, and control the base system in convenient ways.

Result

We train agents to model their belief state using messages in a synthetic language form.

We focus on the metrics of faithfulness previously introduced.

To measure this, we run evaluation episodes where, at the start of each trial, we inject a message indicating that the instructed tag is in some random room.

We find that the greatest gain of this strategy is its ability to imbue the self-model with causal faithfulness.

Read the Paper

Read Paper

Home

Discover

Bookmarks

◐ Recommended Follows

◐ Latest Activity