For a machine learning system to gain user trust, either its reasoning should to be transparent [Rudin, 2019, Freitas, 2014, Lakkaraju et al., 2016, Letham et al., 2015], or it should be capable of justifying

We sincerely thank all reviewers for their feedback. We will clarify the issues raised below and incorporate them into our final version. R2: Reactive baseline, define N + M (L144). "Sequence of episodes" refers to We will extend this with more examples in the main paper. Modeling users' understanding of an AI's mind [A] could provide an explanation component to teach users about the We would like to disentangle two orthogonal aspects of communication, i.e. modeling of other agents and language Our model with emergent language is an interesting extension.

This document contains supplementary material for "Dialog without Dialog Data: Learning Visual The main paper excludes some details which we provide here. Section 12 reports the ablations we use to evaluate the effects of different aspects of the proposed Q-bot. This section describes our architecture in more detail. There is a minor notation difference between this section and the main paper. Note that for the planner there is an additional residual connection at line 16 which augments the hidden state.

Models are trained on a single Titan Xp GPU on an internal cluster. Training time is typically 6-8 hours on 4 CPUs and 32GB of RAM. We train with batch size B = 128 . Like ShapeWorld, RNN encoders and decoders are single layer GRUs with hidden size 1024 and embedding size 500. For additional example games from both datasets, see Figure S1.

Here we introduce a new paradigm for resolving this ambiguity. We think of program synthesis as a kind of communication between the user and the synthesizer.

These tokens in effect become a lexicon used by agents.