LAKE WALES, Fla. — Artificial-intelligence operating-system provider Veritone Inc. is tackling the needle-in-a-haystack problem of real-world data cognition by applying multi-engine “orchestration” to AI, in a bid to streamline the process and increase accuracy.
As shown in this screen shot of the multiline Conductor input screen for a Donald Trump-related search, Conductor can simultaneously seek keywords, video, images, and other cognitive classes.
(Source: Veritone)
The latest iteration of the Conductor layer in Veritone’s aiWare platform can sift through thousands of hours’ worth of audio and video to find the desired clip in a fraction of the time a human would require, according to Veritone (Costa Mesa, Calif.). Like a human orchestra conductor marshaling an assemblage of specialists to produce a cohesive musical performance, the Conductor software intelligently orchestrates multiple, independent cognitive engines, selecting from among roughly 112 third-party engines and cueing the ones most suited to solving various aspects of the cognition problem at hand.
Conductor searches through video for broadcasts by Emily Chang (bottom) from an online photo of her (upper left).
(Source: Veritone)
For example, Veritone claims to have recently sifted through 33,000 hours of recorded conversations and found the relevant clips in just 140 hours, saving the client millions compared with the outlay for real-time playback and analysis by humans.
Conductor uses its Transcription and Sentiment engines to analyze a media clip of a speech for public acceptance.
(Source: Veritone)
“Our complete platform is aiWare, with a suite of applications that sit on top of the stack [and] can ingest any kind of structured or unstructured data, [from] webcams to cloud connections to sensor streams of all kinds,” Tyler Schulze, vice president and general manager of the Veritone partner ecosystem, told EE Times. “Conductor chooses which engine from the cognition layer to use, and the temporal elastic database handles time correlation among the various speeds of data streams. The user just sees the top application layer, which allows the search and discovery of any number of simultaneous cognitive classes.”
Conductor analyzes police bodycam footage showing a man�s tattooed arm holding semiautomatic pistol.
(Source: Veritone)
The aiWare operating system, which Veritone claims is the first available OS platform for artificial intelligence, works essentially in two passes. First, it samples audio (such as from recorded conversations or wiretaps), video (for example, from closed-circuit television feeds, on-dash recorders, or police body cams), or other data streams. Conductor then does a preliminary analysis as to which AI engine or engines are most appropriate for solving the cognition problem. Next, it submits the complete stream file to the selected AI engine or engines, which extract the desired clips. Sometimes, to provide more actionable insights, it runs multiple passes using different specialized AI engines to fill in the gaps and achieve the client’s desired level of accuracy.
The company says its orchestrated, multi-engine approach yields a 7 percent increase in accuracy, along with 99 percent increases in speed and convenience, over the results obtainable using a single AI engine. Veritone’s original business model for Conductor was limited to software-as-a-service, but depending on the cognitive engines the user needs, it can also be run in-house or even on a laptop under certain circumstances.
Meanwhile, Veritone is upgrading Conductor to run multiple AI engines simultaneously on different elements of a single data stream, taking into account the elements’ divergent qualities and accuracy requirements. It has already added AI engines that can recognize human sentiments and create audio/video “signatures.” For example, Conductor can use police officers’ body-camera footage to help identify suspects and ensure accountability.
The company is also expanding its platform’s reach by layering on cognitive engines that can handle contextual data, optical character recognition, license plate recognition, logo recognition, and other specialized cognition problems, with the hope of achieving double-digit increases in accuracy over the results obtainable using a single AI engine.
Veritone was founded in 2014 and went public this spring, trading on the Nasdaq.
— R. Colin Johnson, Advanced Technology Editor, EE Times 
Related articles:
- Cray Rolls Clustered Supercomputers for AI
- IBM Claims 5nm Nanosheet Breakthrough
- Cray Moves to Lasso 'Big Data Deluge'
- IBM Bets on Cognitive Computing
- Intel Banks on Artificial Intelligence