Machine learning (ML) has become one of the hottest areas in data, with computational systems now able to learn patterns in data and act on that information. The applications are wide-ranging: from autonomous robots, to image recognition, drug discovery, fraud detection, etc.
At the cutting edge is deep learning, which draws its inspiration from the networks of neurons that comprise the cerebral cortex. These networks are massively parallel. As such, it’s no surprise that an increasing number of ML approaches are turning to graphical processing units (GPUs)—a key hardware component for general-purpose parallel computation.
Kinetica has been leveraging GPUs for massively parallel data analysis since 2012. As an in-memory analytical database, Kinetica is able to utilize multiple GPUs across many nodes to perform massively parallel statistical and analytical queries. Users can also apply custom code for analytical processing by leveraging user-defined functions, allowing Kinetica to integrate with a growing number of GPU-accelerated ML libraries, such as TensorFlow, Caffe, Torch, and BIDMach.
But this raises the question: if your ML library is already leveraging GPUs, what does Kinetica add to the ML stack?
Kinetica is tried and tested in large-scale enterprises, with production clusters deployed over dozens of nodes. At this scale most ML models are trained on subsets of the raw data, and most do not actually retain this raw data. Instead, they use the raw data to learn a state (e.g., the strengths of various network connections) before disposing of it—or siloing it in a data warehouse, never to be seen again.
With Kinetica, data can be stored in-memory and be rapidly accessed by the ML model as necessary. One key advantage to having the data closely integrated means that the user can always go back and fit their model as necessary.
Consider an example using time series data. It turns out that by learning the data in two stages—first forwards in real time but then again backwards—you will generally achieve a better overall fit to the entire dataset (i.e., Kalman smoothing vs. Kalman filtering).
To return to the neuroscience analogy, there is a close parallel to wake-sleep cycle animals. The networks of the brain are thought to learn online throughout the course of the day but require a period of sleep in which these model are re-fit to stored memories, most famously in the auto-associative networks of the hippocampus.
Theorists in machine learning have long been aware of the No Free Lunch Theorem. Simply put, there is no magic algorithm that can perform any better than any other in general — that is, when averaged over all conceivable inputs. What this means is that ML models can only succeed to the extent they are well-constructed for the problem at hand. A model that has been developed for image recognition is unlikely to do well when applied to credit card fraud.
This is true even with deep learning. It is often asserted that deep learning is a fundamentally new innovation that solves the feature selection problem—that is, deep learning will learn features from raw data obviating the need for feature selection. Unfortunately, there is no getting around the No Free Lunch Theorem.
Let’s again consider the cerebral cortex. It is certainly true that the cortex is capable of selecting and refining features via feedback, such as in the the early visual cortex. But note that before even arriving in the cortex, visual information has been extensively filtered, such as in the complex circuitry of the human retina. And most of this is fairly hard-wired: if the rules of physics suddenly changed, your eyes would probably not be of much use.
What this means for ML is that models can benefit enormously from incorporating field expertise and the discovered insights of data scientists.
Here Kinetica is an invaluable addition to your machine learning stack.
