Upgrade 2021, the NTT Research Summit, is officially in the books. The theme of this year’s event was Spark Curiosity and the sessions on both days did just that. This blog post reviews the talks on Day 1 that featured demos of ongoing projects. In the first, Medical & Health Informatics (MEI) Lab Senior Research Scientist Jon Peterson introduced attendees to “Otto,” the MEI Lab’s cardiovascular bio digital twin (CV BioDT) model. In the second, Physics & Informatics (PHI) Lab Senior Research Scientist Satoshi Kako illustrated the problem-solving potential of a coherent Ising machine (CIM). In the third, NTT Research VP of Strategy Kei Karasawa demonstrated the interface to a product based on attribute-based encryption. Here is more detail on each of these three demos.
Like any research organization, the MEI Lab is building upon prior findings, including those of Dr. Prof. Otto Frank, a German physiologist who pioneered cardiovascular performance metrics in the late 19th and early 20th centuries. Named “Otto” in his honor, the MEI Lab’s model represents the cardiovascular system’s fluid-based components in mechanistic terms. With its MATLAB-based front end, Dr. Peterson showed how the system works by filling the “body” virtually with fluid and then displaying heart rates, blood pressure, cardiac output, pressure/volume plots, and different phases of contraction and relaxation.
Moving towards an individual digital twin requires building out several groups of parameters. By collaborating with the National Cerebral and Cardiovascular Center (NCVC) in Japan, the MEI Lab is more fully defining values for the model’s mechanistic components that comprise its hypothesis space. Medical events, the effects of drugs and other such factors make up the model’s scenario space. Over time, these variables will become more complex. “Imagine a 20-dimensional hypercube that would contain all of the parameters that we’re going to need to search through if we want to find the right one for you,” Dr. Peterson said.
In the next step, the MEI Lab runs simulations against several million combinations of values in the hypothesis and scenario spaces to produce real-world correlates. These outputs represent the model’s measurement space. A partnership with another Silicon Valley company, Embody Bio, enables the MEI Lab to take the next step of establishing probabilistic connections between the model’s measurement, scenario and hypothesis realms to introduce the actual CV BioDT. This twin is what will eventually enable clinicians to play what-if scenarios, quickly examine a wide range of options and help determine the best treatment for a particular patient.
One goal going forward is to incorporate real-time patient data and create a closed-loop therapeutic system, guided by one’s physician but informed by the CV BioDT. Over time, the MEI Lab also expects to move from acute to chronic scenarios and eventually, with improvements to the model and greater access to remote sensors, to wellness and proactive care.
The PHI Lab has simulated the physical CIM, which is a network of optical parametric oscillators, by implementing an optimized CIM algorithm on a field-programmable gate array (FPGA)-based silicon chip. The same principles apply to both CIMs. “Once we input energy via internal pulses or injection coupler, then the network finds a solution to the Ising model,” said Dr. Kako. “And once we convert a combinatorial problem to the Ising model, we can solve using the CIM.”
To demonstrate the power of the cyber CIM, Dr. Kako used a Max Cut problem, one of those problems whose solution requires sorting through an exponentially large number of combinations. In Max Cut, when n = 5, the total number of combinations is 32. But when n = 30, there are more than 1 billion, and when n = 100, the total is an astronomically high 10301. In this demo, Dr. Kako went even higher, to n = 1000.
Physicists and other scientists are exploring several ways to solve these types of computationally hard problems. In this demonstration, Dr. Kako compared performance of the cyber CIM to that of quantum annealing (QA). Running on an Intel i9 16 core processor, the QA algorithm attempted to solve five instances of the Max Cut problem, with n = 1,000 and 100 trials per instance. The result? In only one of the five instances, which took ten seconds each, was the QA algorithm able to find a solution in every trial. On the other hand, the cyber CIM found solutions in all of the five instances in one tenth of the time. Moreover, these instances each contained 1,000 trials. The net output, therefore, was a 100 times speedup in cyber CIM, as compared to simulated annealing.
These benchmarking results are important, but more will follow. To address a host of real-world applications, it will be necessary to increase the number of variables from 1,000 to 10,000 to 100,000 to 1 million. “This demonstration today is just the first step toward that future progress,” Dr. Kako said. Through the NTT Research Industry/Academia Open Laboratory strategy, which now involves multiple partnerships, the PHI Lab anticipates both scaling and improvements in CIM architecture and algorithm.
In contrast with today’s current encryption schemes, attribute-based encryption (ABE) is much more finely tuned. Instead of an all-or-nothing functional model, ABE can grant prescribed access to someone who has a set of matching traits. Introduced in a technical paper co-authored by CIS Lab Senior Scientist Brent Waters 16 years ago, ABE is now on the verge of becoming productized.
Dr. Karasawa demonstrated three ABE use cases involving access control of various elements in a user’s data wallet. First, at the airport, where someone needs to show a driver’s license, ABE can issue a decryption key, which is then stored in the airport system. Thereafter, the system will be able to read the one line in the wallet that identifies the traveler. ABE could also facilitate credit-card reloading of a transit pass, by creating an encryption key with simultaneous access to both lines (the pass and the card). In a third example, involving a crowded sports stadium, a data wallet may need to provide both a photo ID and evidence of a current vaccination. An ABE system requires a decryption key embedded with an access policy for data having those specific attributes. Notably, this technology will work even when a security system is offline, enabling high performance in such cases. While these examples are geared toward individuals, ABE is also applicable to data lakes containing corporate information as well as wider-scale initiatives, such as those involving IoT and Smart Cities.
“Our future is supporting the data holders, like the wallet holder, or company CISO or government CISO,” said Dr. Karasawa. “They decide security policies, for example, the security PAP (policy administration point) which supports policy settings like those seen in the demonstrations. Once an access policy is set, a key generation server (KGS) provided by NTT distributes an encryption key and multiple decryption keys so that a user can encrypt data and enforce a security policy at a data layer.”
One reason for the arrival of ABE is the emergence of global standards. The U.S. National Institute of Standards and Technology (NIST) issued a specification for attribute-based access control (ABAC) in 2014, and the European Telecommunications Standards Institute (ETSI) published related specifications in 2018, which it updated in May 2021.