A5101754207- Visual perception and processing mechanisms
- Neural dynamics and brain function
- Advanced Algorithms and Applications
- Neural Networks and Applications
- Distributed systems and fault tolerance
- Advanced Neural Network Applications
- Handwritten Text Recognition Techniques
- Remote Sensing and Land Use
- Machine Learning and Data Classification
- Embedded Systems and FPGA Design
- Time Series Analysis and Forecasting
- Cancer Cells and Metastasis
- Explainable Artificial Intelligence (XAI)
- Topic Modeling
- Visual Attention and Saliency Detection
- Advanced Data Compression Techniques
- Advanced Sensor and Control Systems
- Pancreatic and Hepatic Oncology Research
- Spectroscopy and Chemometric Analyses
- Image Retrieval and Classification Techniques
- Telecommunications and Broadcasting Technologies
- Vehicle License Plate Recognition
- Adversarial Robustness in Machine Learning
- Image and Signal Denoising Methods
- Face Recognition and Perception
Huazhong University of Science and Technology
2021
Shanghai Jiao Tong University
2006-2020
Carnegie Mellon University
2018-2019
Center for the Neural Basis of Cognition
2018
Hefei University
2013
This paper presents a method to pursue semantic and quantitative explanation for the knowledge encoded in convolutional neural network (CNN). The estimation of specific rationale each prediction made by CNN key issue understanding networks, it is significant values real applications. In this study, we propose distill from into an explainable additive model, which explains quantitatively. We discuss problem biased interpretation predictions. To overcome interpretation, develop prior losses...
Neurons in macaque inferotemporal cortex (ITC) respond less strongly to familiar than novel images. It is commonly assumed that this effect arises within ITC because its neurons selectively complex images and thus encode an explicit form information sufficient for identifying a particular image as familiar. However, no prior study has examined whether low-order visual areas selective local features also exhibit familiarity suppression. To address issue, we recorded from area V2 with...
In this paper, we present a method to mine object-part patterns from conv-layers of pre-trained convolutional neural network (CNN). The mined are organized by an And-Or graph (AOG). This interpretable AOG representation consists four-layer semantic hierarchy, i.e., parts, part templates, latent patterns, and units. associates each object with certain units in feature maps conv-layers. is constructed very few annotations (e.g., 3–20) parts. We develop question-answering (QA) that uses active...
In this paper we developed a hierarchical network model, called Hierarchical Prediction Network (HPNet), to understand how spatiotemporal memories might be learned and encoded in the recurrent circuits visual cortical hierarchy for predicting future video frames. This neurally inspired model operates analysis-by-synthesis framework. It contains feed-forward path that computes encodes features of successive complexity feedback levels project their interpretations level below. Within each...
Scene text detection plays an important role in terms of its position as the first step many methods such recognition. However, despite tremendous improvement made for scene detection, context this deep learning era, there still remains challenges variety scales and distinguish difficulty due to disturbance background. This paper introduces attention guided multi-scale regression method detecting text, which achieves promising performance. The pipeline gathers feature maps from two different...
<abstract> <b>Abstract.</b> Size and color are two important indexes to measure the external quality of apples. The current apple size detection methods mostly designed on PC platform, which difficult be popularized smaller companies growers for high expense. This research implemented a novel system based embedded platform that covers full function but with much lower cost. hardware was mainly composed ARM11 processor, CMOS camera stepping motor. real-time image acquired by diameter value...
The overhead brings by metadata journaling is extra space and performance degrade caused frequent journal data flush. A file system based remote scheme was designed implementation. removes the I/O activities from local disk to server. According experiments in this paper, increases about 8% 19% performance, but penalty light. Although does need more CPU time for network transfer, less than 8%. And bandwidth taken 6% bound workload.
Feedforward CNN models have proven themselves in recent years as state-of-the-art for predicting single-neuron responses to natural images early visual cortical neurons. In this paper, we extend these with recurrent convolutional layers, reflecting the well-known massive recurrence cortex, and show robust increases predictive performance over feedforward across thousands of hyperparameter combinations three datasets macaque V1 V2 responses. We propose circuit can be conceptualized a form...