One of the exciting aspects of statistics is how rapidly artificial intelligence (AI) is being explored for traditional human easy/computer hard tasks. Image recognition using "deep learning" algorithms (i.e, convolutional neural networks) helped pique my interest in application to health care data. I'm not alone with this recognition. Actually, the application of deep learning techniques feels like the next space race.
Part of this space race is having a computer system that can be used to explore the applications of deep learning. The software, mainly Keras and TensorFlow running on top of Python, provide the statistical modeling framework. The hardware requires a graphical processing unit (GPU) to be efficient in training (estimating) the models. A practical problem for me at my work is the need to work off a managed computer system. This limits the access to what I like to call 'rogue' UNIX boxes. Windows systems do not work as well with Python as needed. So, that leaves Mac systems.
That's great, problem solved? Not quite, NVIDIA's CUDA deep learning libraries require NVIDIA GPUs and Apple and NVIDIA parted ways many moons ago. Fortunately, one can more or less hack the mac to allow it to use an external GPU (eGPU). This is not standard use nor is it really within the warranty specifications of Apple. Nonetheless, there's a community of people rapidly exploring how to translate a run of a mill Mac computer into a deep learning beast. I've been involved with this movement for almost a year now. It's fascinating. This is the first of a series of posts describing how I built two GPU-powered Mac systems. There's hope that the next MacOS release will bring native support, but until then, one needs to wade deeply into the bowels of the internet to get the systems up and running.
This was my first video made of a repeatably usable Mac eGPU system. The GPU is a Titan Xp enclosed in a Akitio Node eGPU housing. Roughly decreased training time on grossly mismatched CPU systems by 20x or so! That moves deep learning training into something can be realistically performed in practice.
One thing that video shows is that I'm not really using Python. I'm using an interface to Keras built by RStudio. This RStudio Keras allows me to work proficiently in R, a language I know well, but can harness the power of TensorFlow. Thank you RStudio!
This is another video that I've made showing a "real" machine learning task. This is a problem I'm working on at work. This is much hard than simple digit recognition. We'll see how we advance the modeling with time.
Coming up next, I'll have a post using the RStudio tfruns package. This package provides an eloquent way to train a deep learning algorithm by brute force. Fun! So until next time. Thanks for reading my content.
~Rickey