Ensembles of Deep LSTM Learners for Activity Recognition using Wearables (1703.09370v1)

Abstract: Recently, deep learning (DL) methods have been introduced very successfully into human activity recognition (HAR) scenarios in ubiquitous and wearable computing. Especially the prospect of overcoming the need for manual feature design combined with superior classification capabilities render deep neural networks very attractive for real-life HAR application. Even though DL-based approaches now outperform the state-of-the-art in a number of recognitions tasks of the field, yet substantial challenges remain. Most prominently, issues with real-life datasets, typically including imbalanced datasets and problematic data quality, still limit the effectiveness of activity recognition using wearables. In this paper we tackle such challenges through Ensembles of deep Long Short Term Memory (LSTM) networks. We have developed modified training procedures for LSTM networks and combine sets of diverse LSTM learners into classifier collectives. We demonstrate, both formally and empirically, that Ensembles of deep LSTM learners outperform the individual LSTM networks. Through an extensive experimental evaluation on three standard benchmarks (Opportunity, PAMAP2, Skoda) we demonstrate the excellent recognition capabilities of our approach and its potential for real-life applications of human activity recognition.

• The paper presents a neural contour perception model that leverages lateral neuron connections to enhance flow visualization in advection tasks.
• The paper introduces a specialized streamline tracing algorithm that integrates bottom-up visual cues with top-down cognitive processing.
• The paper shows that aligned streamlines and LIC methods outperform regular and jittered arrow grids, mirroring human performance.

Neural Modeling of Flow Rendering Effectiveness

The paper "Neural Modeling of Flow Rendering Effectiveness" by Daniel Pineo, Colin Ware, and Sean Fogarty investigates the perceptual capabilities of the human visual cortex to determine the most effective 2D flow visualization methods for advection pathway tasks. The authors propose a model that simulates contour perception based on neurological theories related to the primary visual cortex (V1), emphasizing contour enhancement and streamline tracing.

Core Contributions

The paper makes several significant contributions to the understanding and application of neural modeling and visualization:

1. Contour Perception Model: By leveraging Zhaoping Li's cortical model, the authors offer a two-stage process to model contour enhancement. Through this process, contour enhancement is achieved by lateral connections between neurons, enhancing continuous contours' representation in the neural map.
2. Streamline Tracing: A specialized streamline tracing algorithm is developed to evaluate advection tasks. This algorithm considers both bottom-up processes driven by visual information and top-down processes, which reflect task demands and higher-order cognitive processing.
3. Visualization Method Comparison: Utilizing the advection path-tracing task of Laidlaw et al., the authors compare four flow visualization methods (aligned streamlines, LIC, and regular and jittered grids of arrows) in terms of advection pathway effectiveness. The model demonstrated that both aligned streamlines and LIC were superior to regular or jittered arrow grids, aligning well with human subject performance patterns.
4. Perceptual Modeling Insights: The neural-based model allowed for a rigorous comparison of different visualization strategies, revealing the relative efficiency of visualizing advection pathways. The insights provided directly correlate with human visual performance.

Numerical Results

The authors quantitatively validate their model against human subjects tasked with tracing advection pathways. Both the model and human subjects showed that aligned streamline-based methods were the most effective, with similar performance results for LIC-based methods. However, regular or jittered grids of arrows resulted in reduced efficacy in visualizing flow patterns.

Practical and Theoretical Implications

The practical outcomes of this research underscore the necessity of selecting appropriate visualization techniques depending on the visualization's goals, particularly when high perceptual accuracy is required. Theoretically, the research helps bridge the gap between neurological theories of contour perception and practical applications in data visualization.

Moreover, the paper provides foundational groundwork for future investigations into more complex neural processes involved in contour perception. With further enhancements, this model could expand its application to more advanced visualization tasks or aid in the design of better visualization tools that consider human visual perception capabilities.

Future Work Directions

The paper sets the stage for future explorations into neural processing models by suggesting potential improvements to the model to better simulate human visual perception. Future versions could incorporate more sophisticated models accounting for higher cortical areas, non-uniform receptive field sizes, and multiple spatial frequencies.

Furthermore, given the model's simplifications in cortical functions, subsequent work could improve the model's realism and expand its utility in more varied visualization contexts. Applying these advanced models could improve understanding and design of visualization techniques aligned more closely with human cognitive and perceptual capabilities, leading to enhanced decision-making in complex environments.

In conclusion, the paper's integration of neurological theories with data visualization tasks provides a nuanced look into the perceptual processes involved in flow rendering. The use of a neural-based model for evaluating visualization techniques offers valuable insights that could significantly optimize practical applications in multiple fields.