Scalable deep learning for the enterprise with DL4J

Description

Deep learning offers a powerful suite of techniques for building predictive models and creating value from large stores of digital data that are increasingly common in domains as diverse as manufacturing, finance, advertising, and healthcare. Deep learning models have achieved near or better than human performance in machine translation, speech recognition, and image classification and recently beat the world’s best human Go players, a feat that was believed to be decades away.

The rise of deep learning has several implications for data mining and analytics. First, building successful predictive models depends less on designing good features than it does upon matching architectures to problems and tuning hyperparameters. These in turn require a new set of skills and intuitions among practitioners. Second, deep learning models are computationally expensive to train and validate. Training even simple models on modestly sized datasets can require specialized hardware like GPUs while state-of-the-art performance featured in popular press often involves training large models with massive datasets on complex distributed compute platforms.

Dave Kale, Susan Eraly, and Josh Patterson provide a practical introduction to training neural networks using Deeplearning4j (DL4J), the open source, Java-based deep learning framework. Through both presented material and hands-on exercises, you’ll learn how to meet the challenges that arise when you apply deep learning to your business problems as Dave, Melanie, Susan, and Josh walk you through using DL4J to build and train several canonical neural net architectures, including a convolutional network for image classification and a recurrent neural net for sequence modeling.

Topics include:

• A review of key machine-learning terminology and fundamentals, including task formulation, principled evaluation, overfitting and regularization, loss functions, performance metrics, etc.
• The building blocks of deep learning: Gradient descent, backpropagation, nonlinearities, dropout, etc.
• Specialized neural net architectures for structured data, such as convolutions and recurrent layers
• An overview of the DL4J ecosystem: ND4J, for NumPy-like arrays and operations, DataVec, for data preparation pipelines, Deeplearning4j for building neural net architectures, RL4J for reinforcement learning, Arbiter for hyperparameter tuning, and ScalNet, a Keras-like Scala API
• Parallel training with multiple GPUs and distributed training via the DL4J Spark API
• Loading pretrained models from other frameworks, such as Keras