PORTLAND, Ore. Intel Corp. unveils a do-everything machine learning package this week capable of fusing separate streams from real-time sensors as easily as spotting and identifying objects and conditions.
Released as "open source" software, the downloadable machine learning suite will be described at the Neural Information Processing Systems (NIPS) conference this week in Vancouver, B.C.
For the first time ever, according to Intel, its probabalistic network library will enable causal relations to be easily cast into control programs that monitor sensor networks. Prior to probabalistic networks, statistical methods could only categorize correlations, which could relate, for example, a wet lawn to rain, but not tell which caused which.
By adding directed graphs, which show the direction of causality, large numbers of incoming data streams can be tamed down to the conclusions that should be drawn from them by merely "following the graph."
"The probabalistic graphical models in PNL provide a formal way, for the first time in science, of describing causality," said Intel researcher Gary Bradski. "The arrows in a Baysean networks don't show direction, only correlations, so people used to have to do a lot of hand-waving to justify their conclusions.
"But in our causal models the arrows do show causal directions, or if not causality, then they show a generative relationship."
Instead of housing a conglomeration of diverse techniques, Intel chose to combine an one-size-fits-all core inference engine with a host of enhancements to fill in the cracks. The initial release is the beta version, with a "gold" release dur next summer.
"The proper terminology is that this is a probabalistic graphical model, and these techniques cover a huge range of formal techniques everything from principle component analysis to Kalman filters to Markov models" used in speech and tracking, said Bradski. "You can describe a lot of neural networks as graphical models, you can also use a neural network as an input observation. PNL is a much a better formalism for wrapping a lot of algorithms together."
In addition to the Baysean networks, here called a directed acyclic graph, or DAG, the PNL also offers a wide range of undirected models and some mixed types. "In undirected models, you typically call a random Markov field, and often these are grid-like as in physics' spin-glass problems. All these and more can be subsumed under probabalistic graphical models,"
Bradski said. "There are also combinations of directed and undirected graphs, called chain graphs. We still have a few holes in the library now, but those should be filled in by the gold release."
At the NIPS conference, Intel will compare nine other similar machine learning software packages, only three of which (including Intel) come with the source code. Of the remaining two, only Intel's library in C++, the other two being in MatLab and Java.