Deep Learning on GPGPUs
We introduce the most important applications of deep learning, using CAFFE and TensorFlow platforms.
We focus on performance considerations - for the training phase using multicore CPUs, distributed systems (cluster, cloud) and Nvidia GPU accelerators using NVIDIA CUDA compatible frameworks (NVIDIA Tesla / NVIDIA GTX), as well as the deployment phase on various platforms including embeddable boards such as NVIDIA Jetson TK1. Trainees will have deep understanding of Deep Learning frameworks and their real world usage for developing production applications.
Deep learning on GPGPUs training (3 days)
Objectives:
|
Day 1
Morning (9AM-12PM) - Introduction & Demo
• Deep Learning & Machine Learning • Usages & applications • Supervised & Unsupervised Learning • Learning strategies • Neural Network history and classification
|
Day 2
Morning (9AM-12PM) - Supervised Learning Fundamentals: Theory and Practice with TensorFlow / Caffe (2/2)
• Convolutional neural networks • Image classification using convolutional filtering • BackPropagation and Recurrent Neural Network |
Afternoon (1PM-5PM) - Supervised Learning Fundamentals: Theory and Practice with TensorFlow / Caffe (1/2)
• Deep Learning available Frameworks overview • Supervised Learning Fundamental 1: feed-forward network, stochastic gradient descent, testing and validation, regularization • non-linear classifier using TensorFlow / Caffe - define network, training model, refine accuracy, visualize network, deploy |
Afternoon (1PM-5PM) - Caffe and TensorFlow comparison, Supervised Learning Application
• Introduction • cuDNN & Caffe, cuDNN and TensorFlow
|
|
|
Day 3
Morning (9AM-12PM) - Unsupervised Learning Theory and Practice
• Auto-encoders • Dynamic Networks • Competitive Networks
|
Afternoon (1PM-5PM) - Practical Issues
• Data Preprocessing • Training algorithms • Training initialization and Stop Conditions • Post-Training Analysis and case Studies: Fitting, Pattern Recognition, Prediction, Sensitivity analysis, Overfitting, Clustering |