ML-o-Scope: A Diagnostic Visualization System for Deep Machine Learning Pipelines

Abstract

We thank Evan Sparks and Josh Rosen for their many contributions to deepViz, a project that served as the basis for development of ML-o- AbstractThe recent success of deep learning is driving a trend towards structurally complex computer vision models that combine feature extraction with predictive elements into integrated pipelines. While some of these models have achieved breakthrough results in applications like object recognition, they are difficult to design and tune, impeding progress. We feel that visual analysis can be a powerful tool to aid iterative development of deep model pipelines. Building on feature evaluation work in the computer vision community, we introduce ML-o-scope, an interactive visualization system for exploratory analysis of convolutional neural networks, a prominent type of pipelined model. We present ML-o-scope's time-lapse engine that provides views into model dynamics during training, and evaluate the system as a support for tuning large scale object-classification pipelines.

Visual Analytics in Deep Learning: An Interrogative Survey for the Next Frontiers

“…Some tools also address the inherent iterative nature of training neural networks. For example, ML-o-scope [31] utilizes a time-lapse engine to inspect a model's training dynamics to better tune hyperparameters, while work by Chae et al [34] visualizes classification results during training and suggests potential directions to improve performance in the model building pipeline. Lastly, visual analytics tools are beginning to be built for expert users who wish to use models that are more challenging to work with.…”

“…It is common to see example use cases or illustrative usage scenarios that demonstrate the capabilities of the interactive systems. Some works go beyond these and conduct user studies to evaluate utility and usability [31]. In the AI communities, most works do not include user studies.…”

Caffe2Unity: Immersive Visualization and Interpretation of Deep Neural Networks

“…In this category, much work has been carried out on the 2D visualization of images. In [7,13], interactive software tools were used for understanding, and an exploratory analysis was done on how we can visualize the activations produced at each layer in response to user input. In [14], layer-wise relevance propagation (LRP)-based heatmaps were used for region perturbation and for highlighting the important parts in an image during a classification task, while in [8,15], it was demonstrated that a single neuron can represent multiple facets and be visualized via a synthetically generated image, highlighting the specific areas in natural images that best activate the neurons.…”

HyperTendril: Visual Analytics for User-Driven Hyperparameter Optimization of Deep Neural Networks

“…TensorFlow Playground [34], ShapeShop [14], and ReVACNN [5] provide a straightforward analysis system for neural network improvement by allowing the user to directly select the hyperparameters corresponding to the model training preparation stage. ML-o-scope [4], explAIner [35],…”