Large Language Models (LLMs) have demonstrated impressive abilities in symbol processing through in-context learning (ICL). This success flies in the face of decades of critiques asserting that artificial neural networks cannot master abstract symbol manipulation. We seek to understand the mechanisms that can enable robust symbol processing in transformer networks, illuminating both the unanticipated success, and the significant limitations, of transformers in symbol processing. Borrowing insights from symbolic AI and cognitive science on the power of Production System architectures, we develop a high-level Production System Language, PSL, that allows us to write symbolic programs to do complex, abstract symbol processing, and create compilers that precisely implement PSL programs in transformer networks which are, by construction, 100% mechanistically interpretable. The work is driven by study of a purely abstract (semantics-free) symbolic task that we develop, Templatic Generation (TGT). Although developed through study of TGT, PSL is, we demonstrate, highly general: it is Turing Universal. The new type of transformer architecture that we compile from PSL programs suggests a number of paths for enhancing transformers’ capabilities at symbol processing. We note, however, that the work we report addresses computability, and not learnability, by transformer networks.

Compositionality is a core property of natural language, and compositional behavior (CB) is a crucial goal for modern NLP systems.The research literature, however, includes conflicting perspectives on how CB should be defined, evaluated, and achieved.We propose a conceptual framework to address these questions and survey researchers active in this area.We find consensus on several key points.Researchers broadly accept our proposed definition of CB, agree that it is not solved by current models, and doubt that scale alone will achieve the target behavior.In other areas, we find the field is split on how to move forward, identifying diverse opportunities for future research.

Neural networks continue to struggle with compositional generalization, and this issue is exacerbated by a lack of massive pre-training. One successful approach for developing neural systems which exhibit human-like compositional generalization is $\textit{hybrid}$ neurosymbolic techniques. However, these techniques run into the core issues that plague symbolic approaches to AI: scalability and flexibility. The reason for this failure is that at their core, hybrid neurosymbolic models perform symbolic computation and relegate the scalable and flexible neural computation to parameterizing a symbolic system. We investigate a $\textit{unified}$ neurosymbolic system where transformations in the network can be interpreted simultaneously as both symbolic and neural computation. We extend a unified neurosymbolic architecture called the Differentiable Tree Machine in two central ways. First, we significantly increase the model’s efficiency through the use of sparse vector representations of symbolic structures. Second, we enable its application beyond the restricted set of tree2tree problems to the more general class of seq2seq problems. The improved model retains its prior generalization capabilities and, since there is a fully neural path through the network, avoids the pitfalls of other neurosymbolic techniques that elevate symbolic computation over neural computation.

In the context of structure-to-structure transformation tasks, learning sequences of discrete symbolic operations poses significant challenges due to their non-differentiability. To facilitate the learning of these symbolic sequences, we introduce a differentiable tree interpreter that compiles high-level symbolic tree operations into subsymbolic matrix operations on tensors. We present a novel Differentiable Tree Machine (DTM) architecture that integrates our interpreter with an external memory and an agent that learns to sequentially select tree operations to execute the target transformation in an end-to-end manner. With respect to out-of-distribution compositional generalization on synthetic semantic parsing and language generation tasks, DTM achieves 100% while existing baselines such as Transformer, Tree Transformer, LSTM, and Tree2Tree LSTM achieve less than 30%. DTM remains highly interpretable in addition to its perfect performance.

Yichen Jiang, Asli Celikyilmaz, Paul Smolensky, Paul Soulos, Sudha Rao, Hamid Palangi, Roland Fernandez, Caitlin Smith, Mohit Bansal, Jianfeng Gao. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Transformers have increasingly outperformed gated RNNs in obtaining new state-of-the-art results on supervised tasks involving text sequences. Inspired by this trend, we study the question of how Transformer-based models can improve the performance of sequential decision-making agents. We present the Working Memory Graph (WMG), an agent that employs multi-head self-attention to reason over a dynamic set of vectors representing observed and recurrent state. We evaluate WMG in three environments featuring factored observation spaces: a Pathfinding environment that requires complex reasoning over past observations, BabyAI gridworld levels that involve variable goals, and Sokoban which emphasizes future planning. We find that the combination of WMG’s Transformer-based architecture with factored observation spaces leads to significant gains in learning efficiency compared to baseline architectures across all tasks. WMG demonstrates how Transformer-based models can dramatically boost sample efficiency in RL environments for which observations can be factored.

Unit visualizations are a family of visualizations where every data item is represented by a unique visual mark-a visual unit-during visual encoding. For certain datasets and tasks, unit visualizations can provide more information, better match the user's mental model, and enable novel interactions compared to traditional aggregated visualizations. Current visualization grammars cannot fully describe the unit visualization family. In this paper, we characterize the design space of unit visualizations to derive a grammar that can express them. The resulting grammar is called ATOM, and is based on passing data through a series of layout operations that divide the output of previous operations recursively until the size and position of every data point can be determined. We evaluate the expressive power of the grammar by both using it to describe existing unit visualizations, as well as to suggest new unit visualizations.

Abstract Browsing is a fundamental aspect of exploratory information‐seeking. Associative browsing represents a common and intuitive set of exploratory strategies in which users step iteratively from familiar to novel bits of information. In this paper, we examine associative browsing as a strategy for bottom‐up exploration of large, heterogeneous networks. We present Refinery, an interactive visualization system informed by guidelines for associative browsing drawn from literature on exploratory information‐seeking. These guidelines motivate Refinery's query model, which allows users to simply and expressively construct queries using heterogeneous sets of nodes. This system computes degree‐of‐interest scores for associated content using a fast, random‐walk algorithm. Refinery visualizes query nodes within a subgraph of results, providing explanatory context, facilitating serendipitous discovery, and stimulating continued exploration. A study of 12 academic researchers using Refinery to browse publication data demonstrates how the system enables discovery of valuable new content, even within existing areas of expertise.

In order to use new visualizations, most toolkits require application developers to rebuild their applications and distribute new versions to users. The WebCharts Framework take a different approach by hosting JavaScript from within an application and providing a standard data and events interchange. In this way, applications can be extended dynamically, with a wide variety of visualizations. We discuss the benefits of this architectural approach, contrast it to existing techniques, and give a variety of examples and extensions of the basic system.

Designers' extensive software needs have not been adequately documented in the research literature, and are poorly supported by software. Without appropriate tools to support their needs, designers have difficulty knowing the best way to evolve the look and feel of interactive applications they are designing. In order to inform the design of new tools for interactive application design, we used a grounded theory approach to find out what designers' needs are when designing such applications. This paper reports our findings (20 designer needs) from content analysis of five types of artifacts: surveys, blend discussion list emails, dreamweaver forum entries, flash forum entries, and interviews with ten designers. These 20 needs were then validated in follow-up interviews and focus group sessions. The results of this work revealed trends regarding the importance of each need and show that flow is one of the most important needs.

The difference between using one mobile phone and another can feel like learning a new language based on our extensive experience designing mobile applications for spatial data navigation, faceted search, and glanceable information, we have developed design principles for unifying the various aspects of the internet connected mobile phone ("smartphone") user experience.

Animation has been used to show trends in multi-dimensional data. This technique has recently gained new prominence for presentations, most notably with Gapminder Trendalyzer. In Trendalyzer, animation together with interesting data and an engaging presenter helps the audience understand the results of an analysis of the data. It is less clear whether trend animation is effective for analysis. This paper proposes two alternative trend visualizations that use static depictions of trends: one which shows traces of all trends overlaid simultaneously in one display and a second that uses a small multiples display to show the trend traces side-by-side. The paper evaluates the three visualizations for both analysis and presentation. Results indicate that trend animation can be challenging to use even for presentations; while it is the fastest technique for presentation and participants find it enjoyable and exciting, it does lead to many participant errors. Animation is the least effective form for analysis; both static depictions of trends are significantly faster than animation, and the small multiples display is more accurate.