Dylan Cashman

Ashley Suh, Gabriel Appleby, Erik W. Anderson et al.

Presenting a predictive model's performance is a communication bottleneck that threatens collaborations between data scientists and subject matter experts. Accuracy and error metrics alone fail to tell the whole story of a model - its risks, strengths, and limitations - making it difficult for subject matter experts to feel confident in their decision to use a model. As a result, models may fail in unexpected ways or go entirely unused, as subject matter experts disregard poorly presented models in favor of familiar, yet arguably substandard methods. In this paper, we describe an iterative study conducted with both subject matter experts and data scientists to understand the gaps in communication between these two groups. We find that, while the two groups share common goals of understanding the data and predictions of the model, friction can stem from unfamiliar terms, metrics, and visualizations - limiting the transfer of knowledge to SMEs and discouraging clarifying questions being asked during presentations. Based on our findings, we derive a set of communication guidelines that use visualization as a common medium for communicating the strengths and weaknesses of a model. We provide a demonstration of our guidelines in a regression modeling scenario and elicit feedback on their use from subject matter experts. From our demonstration, subject matter experts were more comfortable discussing a model's performance, more aware of the trade-offs for the presented model, and better equipped to assess the model's risks - ultimately informing and contextualizing the model's use beyond text and numbers.

Sonny George, Chris Sypherd, Dylan Cashman

Large language model (LLM) agents show promise in an increasing number of domains. In many proposed applications, it is expected that the agent reasons over accumulated experience presented in an input prompt. We propose the OEDD (Operationalize Experience Despite Distraction) corpus, a human-annotator-validated body of scenarios with pre-scripted agent histories where the agent must make a decision based on disparate experiential information in the presence of a distractor. We evaluate three state-of-the-art LLMs (GPT-3.5 Turbo, GPT-4o, and Gemini 1.5 Pro) using a minimal chain-of-thought prompting strategy and observe that when (1) the input context contains over 1,615 tokens of historical interactions, (2) a crucially decision-informing premise is the rightful conclusion over two disparate environment premises, and (3) a trivial, but distracting red herring fact follows, all LLMs perform worse than random choice at selecting the better of two actions. Our code and test corpus are publicly available at: https://github.com/sonnygeorge/OEDD .

James Petullo, Sonny George, Dylan Cashman et al.

A standard technique for scaling inference-time reasoning is Self-Consistency, whereby multiple candidate answers are sampled from an LLM and the most common answer is selected. More recently, it has been shown that weighted majority voting (e.g. Confidence-Informed Self Consistency (CISC)), which assigns a confidence value to each candidate answer and chooses the answer with the largest accumulated score, tends to be more accurate on a wide range of popular benchmarks. In practice, weighted majority voting necessitates calling a critic LLM on each candidate's reasoning trace to produce the answer's confidence score. This secondary series of LLM calls greatly increases the overhead and cost of weighted majority voting, despite its potential performance benefits. To reduce this expense, we propose VecCISC, a lightweight, adaptive framework that uses a measure of semantic similarity to filter reasoning traces that are semantically equivalent to others, degenerate, or hallucinated, thus decreasing the number of candidate answers that must be evaluated by the critic. To ensure adequate experimental thoroughness, we evaluate VecCISC on five challenging, widely-adopted datasets spanning the domains of mathematics, chemistry, biology, commonsense reasoning, and the humanities. Our results demonstrate that VecCISC reduces the total token usage by 47%, while maintaining or exceeding the accuracy of CISC.

Mateus Espadoto, Gabriel Appleby, Ashley Suh et al.

Projection techniques are often used to visualize high-dimensional data, allowing users to better understand the overall structure of multi-dimensional spaces on a 2D screen. Although many such methods exist, comparably little work has been done on generalizable methods of inverse-projection -- the process of mapping the projected points, or more generally, the projection space back to the original high-dimensional space. In this paper we present NNInv, a deep learning technique with the ability to approximate the inverse of any projection or mapping. NNInv learns to reconstruct high-dimensional data from any arbitrary point on a 2D projection space, giving users the ability to interact with the learned high-dimensional representation in a visual analytics system. We provide an analysis of the parameter space of NNInv, and offer guidance in selecting these parameters. We extend validation of the effectiveness of NNInv through a series of quantitative and qualitative analyses. We then demonstrate the method's utility by applying it to three visualization tasks: interactive instance interpolation, classifier agreement, and gradient visualization.

Dylan Cashman, Shenyu Xu, Subhajit Das et al.

Most visual analytics systems assume that all foraging for data happens before the analytics process; once analysis begins, the set of data attributes considered is fixed. Such separation of data construction from analysis precludes iteration that can enable foraging informed by the needs that arise in-situ during the analysis. The separation of the foraging loop from the data analysis tasks can limit the pace and scope of analysis. In this paper, we present CAVA, a system that integrates data curation and data augmentation with the traditional data exploration and analysis tasks, enabling information foraging in-situ during analysis. Identifying attributes to add to the dataset is difficult because it requires human knowledge to determine which available attributes will be helpful for the ensuing analytical tasks. CAVA crawls knowledge graphs to provide users with a a broad set of attributes drawn from external data to choose from. Users can then specify complex operations on knowledge graphs to construct additional attributes. CAVA shows how visual analytics can help users forage for attributes by letting users visually explore the set of available data, and by serving as an interface for query construction. It also provides visualizations of the knowledge graph itself to help users understand complex joins such as multi-hop aggregations. We assess the ability of our system to enable users to perform complex data combinations without programming in a user study over two datasets. We then demonstrate the generalizability of CAVA through two additional usage scenarios. The results of the evaluation confirm that CAVA is effective in helping the user perform data foraging that leads to improved analysis outcomes, and offer evidence in support of integrating data augmentation as a part of the visual analytics pipeline.

Dylan Cashman, Adam Perer, Remco Chang et al.

Deep learning models require the configuration of many layers and parameters in order to get good results. However, there are currently few systematic guidelines for how to configure a successful model. This means model builders often have to experiment with different configurations by manually programming different architectures (which is tedious and time consuming) or rely on purely automated approaches to generate and train the architectures (which is expensive). In this paper, we present Rapid Exploration of Model Architectures and Parameters, or REMAP, a visual analytics tool that allows a model builder to discover a deep learning model quickly via exploration and rapid experimentation of neural network architectures. In REMAP, the user explores the large and complex parameter space for neural network architectures using a combination of global inspection and local experimentation. Through a visual overview of a set of models, the user identifies interesting clusters of architectures. Based on their findings, the user can run ablation and variation experiments to identify the effects of adding, removing, or replacing layers in a given architecture and generate new models accordingly. They can also handcraft new models using a simple graphical interface. As a result, a model builder can build deep learning models quickly, efficiently, and without manual programming. We inform the design of REMAP through a design study with four deep learning model builders. Through a use case, we demonstrate that REMAP allows users to discover performant neural network architectures efficiently using visual exploration and user-defined semi-automated searches through the model space.

Dylan Cashman, Genevieve Patterson, Abigail Mosca et al.

We present RNNbow, an interactive tool for visualizing the gradient flow during backpropagation training in recurrent neural networks. RNNbow is a web application that displays the relative gradient contributions from Recurrent Neural Network (RNN) cells in a neighborhood of an element of a sequence. We describe the calculation of backpropagation through time (BPTT) that keeps track of itemized gradients, or gradient contributions from one element of a sequence to previous elements of a sequence. By visualizing the gradient, as opposed to activations, RNNbow offers insight into how the network is learning. We use it to explore the learning of an RNN that is trained to generate code in the C programming language. We show how it uncovers insights into the vanishing gradient as well as the evolution of training as the RNN works its way through a corpus.

Dylan Cashman, Shah Rukh Humayoun, Florian Heimerl et al.

Many visual analytics systems allow users to interact with machine learning models towards the goals of data exploration and insight generation on a given dataset. However, in some situations, insights may be less important than the production of an accurate predictive model for future use. In that case, users are more interested in generating of diverse and robust predictive models, verifying their performance on holdout data, and selecting the most suitable model for their usage scenario. In this paper, we consider the concept of Exploratory Model Analysis (EMA), which is defined as the process of discovering and selecting relevant models that can be used to make predictions on a data source. We delineate the differences between EMA and the well-known term exploratory data analysis in terms of the desired outcome of the analytic process: insights into the data or a set of deployable models. The contributions of this work are a visual analytics system workflow for EMA, a user study, and two use cases validating the effectiveness of the workflow. We found that our system workflow enabled users to generate complex models, to assess them for various qualities, and to select the most relevant model for their task.