With next-generation neural processing chips becoming more widely available, Ooblex saw an opportunity to help clients shift their focus from traditional predictive AI to a more reactive, real-time approach.
Ooblex has developed a new type of distributed processing node that can perform hundreds of AI inferences per second—unlocking the potential of real-time AI processing in IBM® Watson® Studio.
Opens upnew possibilities for facial recognition and fraud detection use cases
Reducestime-to-market with seamless model design, training and deployment
Enablesdeployment across a wide range of architectures and devices via Intel OpenVINO
Business challenge story
Creating transformative AI
Over the past decade, artificial intelligence (AI) has begun delivering on its promises, not just as a high-profile showcase for state-of-the-art research, but as a practical tool to automate business processes and disrupt inefficient business models. Many companies are now building up their data science capabilities and working on integrating machine learning models and deep neural networks into their applications.
However, most of these first-generation AI solutions have certain limitations. Training a new model is computationally intensive, and tends to require large farms of power-hungry graphical processing units (GPUs) to deliver results in a reasonable amount of time. Models are typically deployed in the cloud too, which means data needs to travel from an end-user’s device to the cloud and back again to give the user a response. The latency and unreliability of network connections can be a major problem for some use cases—for example, drivers might not find it easy to trust a self-driving car that can’t make decisions unless it is connected to the cloud.
That’s why the arrival of a new generation of neural processing chips is exciting news: it opens up more possibilities for AI training and inference processing to take place on devices at the network edge, from autonomous vehicles to smartphones and even Internet of Things sensors.
Ooblex was created to help organizations take advantage of this new technology, and the ecosystem has a vision for the future of AI that goes far beyond most current use cases, as creator Stephen Seguin explains.
“Today, most AIs are relatively static; they’re set up to address a particular problem and make predictions to help guide decision-making. But if the problem domain evolves—which it almost always does—then the model has to be taken back into the lab, retrained and redeployed.
“With Ooblex, we want to change that paradigm. Instead of having to plan ahead and predict everything in advance, we see AI becoming much more reactive. It should be able to respond to a changing situation in real time and transform its approach to meet new challenges as they arise. Now that the hardware exists to push model training out to devices at the edge of the network, this has become much more possible.”
To turn this vision of transformative AI into reality, Ooblex realized that data scientists and application developers would need better tools. The company began looking for a data science platform that could augment its own efforts and provide a seamless experience for model design, training and deployment.
Building on strong foundations
Ooblex has developed a new type of distributed processing node, called a Tensor Thread. Each of these nodes is capable of supporting hundreds of AI inferences per second, with no model-switching delay. This enables multi-stage inference processing at much higher speeds than a traditional approach—making the Tensor Threads ideal for handling a new generation of AI solutions that are capable of performing well on edge devices.
When a data scientist builds a model, they can use common frameworks such as Caffe, Keras and Tensorflow, as well as outputting the resulting model in an open graph format. This allows users to harness Intel OpenVINO to make the model fast and portable across a wide range of platforms, including Intel CPUs, integrated graphics processors, FPGAs and Movidious devices.
To demonstrate the power of Tensor Threads, Ooblex has also built a service that puts them to good use: the Ooblex Media Gateway. The gateway acts as a real-time video injection platform that provides on-demand access to live media, and eliminates many of the complexities around real-time streaming, video processing and video capture.
Ooblex has open-sourced these technologies, so they are freely available for any organization to leverage—but the company realized that for maximum impact, it needed to frame the new tools in an environment that would make them easy to adopt.
Stephen Seguin emailed: “While it’s possible to cobble together a tool-chain for AI from different open-source projects, we believe that scaling AI to the enterprise level will take more than that. Data science teams need a coherent, well-governed environment to help them design, develop, train and deploy models.
“When we saw IBM Watson Studio, we switched from our existing cloud provider and moved everything to the IBM Cloud™. Of all the data science platforms we saw on the market, Watson Studio was by far the most impressive—not just in terms of its intuitive user interface [UI], but also the way IBM has engineered it to support open standards and integrate open-source tools. It really fits with our vision of an open medium for AI.”
Both Ooblex Media Gateway and the Tensor Threads are now available for IBM Watson Studio, and can be deployed to an IBM Cloud server with just a few clicks. This makes it easy for data scientists to start experimenting with real-time transformative AI models and the open-source Ooblex toolkit.
“With IBM Watson Studio, we can fit Ooblex’s revolutionary technology seamlessly into data scientists’ existing workflows,” says Stephen Seguin. “We’re excited to see the solutions that people will build when they unlock the power of transformative AI.”
Solving new types of problems
To date, Ooblex has been focusing mainly on developing the Media Gateway and Tensor Thread technologies, but the company’s real goal is to help organizations across all industries solve business problems using transformative AI. Ooblex has already built relationships with a number of clients who have identified use cases for the technology, especially in areas such as facial recognition and computer vision.
“Much of our work with clients is confidential at the moment,” says Stephen Seguin. “But we have been doing some public demos to help clients understand what kinds of solutions they can build with our technology.”
For example, one demo focuses on how Ooblex technology can counteract the phenomenon of “deepfaking”, where deep neural networks are used to manipulate video content to superimpose the features of celebrities or politicians on existing videos. As deepfaking techniques continue to evolve, it is becoming more and more difficult for human viewers to tell the difference between real and doctored footage—making it easier for malicious users to create hoaxes and spread fake news, or potentially even gain access to devices that use facial recognition as a security mechanism.
The Ooblex demo acts as a showcase of real-time deepfaking, taking a live video stream and outputting a new stream that replaces the facial features of a presenter with those of a famous politician. The point of the demo is not to show that Ooblex can create better deepfakes, but that Ooblex’s AI technology can detect whether a given video stream has been tampered with.
Stephen Seguin comments: “From smartphone vendors to banks, many organizations are exploring the use of facial recognition as a secure, convenient way to authenticate users who log into their systems. But as deepfaking techniques become more sophisticated, these systems won’t be secure unless they can learn and adapt to the new tricks that fakers are using. Ooblex technology will help our clients set AI against AI to detect even the most advanced fraud attempts in real time.”
He concludes: “By building on our close relationship with IBM, Ooblex will be able to bring transformative AI to a much wider market. IBM Watson Studio and the IBM Cloud offer a powerful platform for harnessing Tensor Threads to build the next generation of AI-infused applications.”
Ooblex is building an open-source ecosystem for transformative artificial intelligence—shifting the AI paradigm from predictive models to reactive applications that can transform the way our world works in real time.
Take the next step
IBM Analytics offers one of the world's deepest and broadest analytics platform, domain and industry solutions that deliver new value to businesses, governments and individuals. For more information about how IBM Analytics helps to transform industries and professions with data, visit ibm.com/analytics. Follow us on Twitter at @IBMAnalytics, on our blog at ibmbigdatahub.com and join the conversation #IBMAnalytics.