Embedded AI Chips for Robust Deep Image Classification and its Applications Opportunity and Significance The aim of this research is to develop novel algorithms to meet the forthcoming challenges in the area of machine intelligence and computer vision on FPGA chips. This area is having high potential for addressing human machine semantic gaps and will bring a revolution to the techniques for indexing, retrieving and interacting video signals. The project aims that the deep and distributed learning. The research involves novel methods for economically, medically and scientifically big data problems and will lead to new discoveries in the domains of Biomedical, Defense, Automotive and Cyber Security Sector. Next Steps for Development and Test Our next step is to provide the fast software implementation of the proposed approach, therefore can be embedded onto cheap hardware for automotive industry applications. Technical Objectives 1) To develop low cost, less energy embedded Artificial Intelligence FPGA chips to monitor health, decode Genome and to embedded to Automobile. 2) Developing AI based techniques for video classification. 3) Investigate the theoretical aspects of convergence, covariance and scalability. Related Work and State of Practice Performance of Deep learning architecture is exponentially better and can be implemented on FPGAs along with this we need algorithms hat can deal with different types of noise at different levels. Commercialization Plan & Partners With the partners in automotive industries and medical industry, tailor the model to data acquired from automotive sensors for robust classification of objects and hazards References [1] S. Reed, H. Lee, D. Anguelov, C. Szegedy, D. Erhan, and A. Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014. [2] I. Jindal, M. Nokleby, and X. Chen. Learning deep networks from noisy labels with dropout regularization. In Data Mining (ICDM), 2016 IEEE [3] S. Sukhbaatar, J. Bruna, M. Paluri, L. Bourdev, and R. Fergus.Training convolutional networks with noisy labels.arXiv preprint 2014. Ishan Jindal Electrical and Computer Engineering Deep Neural Networks Given datasets Train Neural Network offline Output Label Using LEDs LED 9 will turn ON I II III Theorem: For any K-S classification problem, the diversity order is 32x32 [SSL] Standard Subspace Learning [K-SLD 2 ] Proposed Approach
Welcome message from author
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Embedded AI Chips for Robust Deep Image Classification and its Applications
Opportunity and SignificanceThe aim of this research is to develop novel algorithms to meet the forthcoming challenges in the area of machine
intelligence and computer vision on FPGA chips. This area is having high potential for addressing human machine
semantic gaps and will bring a revolution to the techniques for indexing, retrieving and interacting video signals. The project aims that the deep and distributed
learning. The research involves novel methods for economically, medically and scientifically big data
problems and will lead to new discoveries in the domains of Biomedical, Defense, Automotive and Cyber Security
Sector.
Next Steps for Development and Test
Our next step is to provide the fast software implementation of the proposed approach,
therefore can be embedded onto cheap hardware for automotive industry applications.
Technical Objectives
1) To develop low cost, less energy embedded Artificial Intelligence FPGA chips to monitor health, decode Genome and to embedded to Automobile.
2) Developing AI based techniques for video classification.3) Investigate the theoretical aspects of convergence,
covariance and scalability.
Related Work and State of PracticePerformance of Deep learning architecture is exponentially
better and can be implemented on FPGAs along with this we need algorithms hat can deal with different types of noise at
different levels.
Commercialization Plan & Partners
With the partners in automotive industries and medical industry, tailor the model to data acquired from automotive sensors for robust classification of
objects and hazards
References[1] S. Reed, H. Lee, D. Anguelov, C. Szegedy, D. Erhan, and A. Rabinovich. Training deep neural networks on noisy labels with bootstrapping. arXiv preprint arXiv:1412.6596, 2014.[2] I. Jindal, M. Nokleby, and X. Chen. Learning deep networks from noisy labels with dropout regularization. In Data Mining (ICDM), 2016 IEEE[3] S. Sukhbaatar, J. Bruna, M. Paluri, L. Bourdev, and R. Fergus.Training convolutional networks with noisy labels.arXiv preprint 2014.
Ishan JindalElectrical and Computer Engineering
Deep Neural Networks
Given datasets Train Neural Network offline
Output LabelUsing LEDs
LED 9 will turn ON
I II III
Theorem: For any K-S classification problem, the diversity order is
32x32
[SSL] Standard Subspace Learning[K -S L D 2 ] Proposed Approach
Related Documents
