Posts by Collection

portfolio

Poster 1

Published:

The poster presents a study where task-optimized convolutional neural networks (CNNs) challenge the expertise hypothesis, suggesting that systems broadly optimized for object recognition provide a better foundation for learning fine-grained tasks like car discrimination than systems optimized for face recognition, thus questioning the computational viability of the expertise hypothesis.

Poster 2

Published:

The study explores face pareidolia, where humans see faces in random stimuli, using the DeepGaze model to compare its detection abilities with human gaze patterns, revealing DeepGaze’s potential in recognizing face-like patterns but also its limitations in fully capturing the nuances of human gaze behavior in face pareidolia.

publications

Digital image processing with deep learning for automated cutting tool wear detection

Published in Procedia Manufacturing, 2020

This study explores the application of deep learning in digital image processing for the detection of wear on cutting tools, with a focus on detection. Measurement is considered in the next paper.

Recommended citation: Bergs, T., Holst, C., Gupta, P., & Augspurger, T. (2020). "Digital image processing with deep learning for automated cutting tool wear detection." JProcedia Manufacturing, 48, 947–958.

Deep learning and rule-based image processing pipeline for automated metal cutting tool wear detection and measurement

Published in IFAC-PapersOnLine, 2022

This paper presents a digital and big data analytics approach to quantify metal cutting tool wear, employing a pipeline of deep learning for processing images and a rule-based method for measuring wear along the cutting edge. The automated system enables inline tool wear detection and measurement within CNC machining applications.

Recommended citation: Holst, C., Yavuz, T. B., Gupta, P., Ganser, P., & Bergs, T. (2022). "Deep learning and rule-based image processing pipeline for automated metal cutting tool wear detection and measurement." IFAC-PapersOnLine, 55(2), 534–539.

CNNs reveal the computational implausibility of the expertise hypothesis

Published in Iscience, 2023

This study challenges the expertise hypothesis suggesting face-specific brain mechanisms are domain-general, showing neural networks optimized for generic object categorization outperform those for face recognition in expert object discrimination. It highlights the computational implausibility of domain-general mechanisms being as effective as face-specific ones in specialized tasks.

Recommended citation: Kanwisher, N., Gupta, P., & Dobs, K. (2023). "CNNs reveal the computational implausibility of the expertise hypothesis." Iscience, 26(2).

Human-like face pareidolia emerges in deep neural networks optimized for face and object recognition [Under Review]

Published in PLoS computational biology [Under Review], 2023

Using deep convolutional neural networks (CNNs) and magnetoencephalography (MEG), this study investigates the neural basis of face pareidolia, showing that initial misidentification of faces in inanimate objects is a byproduct of the brain’s optimization for face and object recognition. The research reveals that while early stages of processing mistake pareidolia for real faces, this error is corrected in later stages through specialized face recognition optimization.

Recommended citation: Gupta, P., & Dobs, K. (2023). "Human-like face pareidolia emerges in deep neural networks optimized for face and object recognition [Under Review]."

Investigating face pareidolia using DeepGaze: Bridging human and artificial perception [In Preparation]

Published in [In Preparation], 2024

This study employs DeepGaze models to investigate face pareidolia, revealing their superior ability to detect face-like patterns over standard models and highlighting challenges in explaining gaze prediction complexity. Findings underscore the importance of dataset diversity and reveal nuances in modeling individual versus collective gaze patterns in understanding human visual perception.

Recommended citation: Gupta, P., & Dobs, K. (2024). "Investigating face pareidolia using deepgaze: Bridging human and artificial perception [In Preparation]."

talks

CNNs reveal the computational implausibility of the expertise hypothesis

Published:

In this workshop poster presentation, I discussed the use of convolutional neural networks (CNNs) to test the computational plausibility of the expertise hypothesis in visual recognition processes. The session included an in-depth analysis of how deep learning models can inform our understanding of visual cognition, emphasizing the parallels and distinctions between artificial and human perceptual capabilities.

CNNs reveal the computational implausibility of the expertise hypothesis

Published:

This poster presentation provided groundbreaking research using convolutional neural networks (CNNs) to challenge the expertise hypothesis within the field of visual perception, offering critical insights into the computational limits and possibilities of Fusiform Face Area (FFA).

teaching

Mentoring experience 1

Master Thesis, Justus Liebig University, FB-06 Department, 2022

Samuel Sander’s Master Thesis explores the inversion effects in humans and deep neural networks, examining how orientation affects object recognition in both. By comparing human performance with that of deep neural networks across various visual tasks, this work seeks to understand if neural networks can predict inversion effects in humans. Through methodological approaches involving the Ecoset dataset and different network architectures, the thesis finds significant inversion effects in both humans and neural networks, suggesting similarities in classification behaviors despite differences in error distributions under increased image distortion.

Mentoring experience 2

Bachelor Thesis, Justus Liebig University, FB-06 Department, 2023

Christine Huschens’ Bachelor Thesis focuses on a study of the inversion effects in the perception of faces, pareidolias, and objects in both biological and artificial networks, employing eye-tracking and DeepGaze saliency maps for comparison. The thesis covers the phenomenon of pareidolia—the tendency to perceive faces in everyday objects—and investigates how this phenomenon and face recognition are affected when images are inverted. It explores the neural processes involved in detecting faces and pareidolias, comparing human and artificial neural network responses to these stimuli. The study aims to understand the influence of image inversion on fixation patterns in humans and artificial networks, and how context (art vs. real objects) affects these patterns.