A new accelerator chip called “Hiddenite” that can achieve state-of-the-art accuracy in the calculation of sparse “hidden neural networks” with lower computational burdens has now been developed by Tokyo Tech researchers. By employing the proposed on-chip model construction, which is the combination of weight generation and “supermask” expansion, the Hiddenite chip drastically reduces external memory access for enhanced computational efficiency.

Deep neural networks (DNNs) are a complex piece of machine learning architecture for AI (artificial learning) that require numerous parameters to learn to predict outputs. DNNs can, however, be “pruned,” thereby reducing the computational burden and model size. A few years ago, the “lottery ticket hypothesis” took the machine learning world by storm. The hypothesis stated that a randomly initialized DNN contains subnetworks that achieve accuracy equivalent to the original DNN after training. The larger the network, the more “lottery tickets” for successful optimization. These lottery tickets thus allow “pruned” sparse neural networks to achieve accuracies equivalent to more complex, “dense” networks, thereby reducing overall computational burdens and power consumptions.

Figure 1. HNNs find sparse subnetworks which achieve equivalent accuracy to the original dense trained model. Credit: Masato Motomura from Tokyo Tech

One technique to find such subnetworks is the hidden neural network (HNN) algorithm, which uses AND logic (where the output is only high when all the inputs are high) on the initialized random weights and a “binary mask” called a “supermask” (Fig. 1). The supermask, defined by the top-k% highest scores, denotes the unselected and selected connections as 0 and 1, respectively. The HNN helps reduce computational efficiency from the software side. However, the computation of neural networks also requires improvements in the hardware components.

Traditional DNN accelerators offer high performance, but they do not consider the power consumption caused by external memory access. Now, researchers from Tokyo Institute of Technology (Tokyo Tech), led by Professors Jaehoon Yu and Masato Motomura, have developed a new accelerator chip called “Hiddenite,” which can calculate hidden neural networks with drastically improved power consumption. “Reducing the external memory access is the key to reducing power consumption. Currently, achieving high inference accuracy requires large models. But this increases external memory access to load model parameters. Our main motivation behind the development of Hiddenite was to reduce this external memory access,” explains Prof. Motomura. Their study will feature in the upcoming International Solid-State Circuits Conference (ISSCC) 2022, a prestigious international conference showcasing the pinnacles of achievement in integrated circuits.

Figure 2. The new Hiddenite chip offers on-chip weight generation and on-chip “supermask expansion” to reduce external memory access for loading model parameters. Credit: Masato Motomura from Tokyo Tech

“Hiddenite” stands for Hidden Neural Network Inference Tensor Engine and is the first HNN inference chip. The Hiddenite architecture (Fig. 2) offers three-fold benefits to reduce external memory access and achieve high energy efficiency. The first is that it offers the on-chip weight generation for re-generating weights by using a random number generator. This eliminates the need to access the external memory and store the weights. The second benefit is the provision of the “on-chip supermask expansion,” which reduces the number of supermasks that need to be loaded by the accelerator. The third improvement offered by the Hiddenite chip is the high-density four-dimensional (4D) parallel processor that maximizes data re-use during the computational process, thereby improving efficiency.

Figure 3. Fabricated using 40nm technology, the core of the chip area is only 4.36 square millimeters. Credit: Masato Motomura from Tokyo Tech

“The first two factors are what set the Hiddenite chip apart from existing DNN inference accelerators,” reveals Prof. Motomura. “Moreover, we also introduced a new training method for hidden neural networks, called ‘score distillation,’ in which the conventional knowledge distillation weights are distilled into the scores because hidden neural networks never update the weights. The accuracy using score distillation is comparable to the binary model while being half the size of the binary model.”

Based on the hiddenite architecture, the team has designed, fabricated, and measured a prototype chip with Taiwan Semiconductor Manufacturing Company’s (TSMC) 40nm process (Fig. 3). The chip is only 3mm x 3mm and handles 4,096 MAC (multiply-and-accumulate) operations at once. It achieves a state-of-the-art level of computational efficiency, up to 34.8 trillion or tera operations per second (TOPS) per Watt of power, while reducing the amount of model transfer to half that of binarized networks.

These findings and their successful exhibition in a real silicon chip are sure to cause another paradigm shift in the world of machine learning, paving the way for faster, more efficient, and ultimately more environment-friendly computing.

Companies with technologies that may advance exploration but need a little extra push to finalize development have two new opportunities to partner with NASA to make it over the finish line.

Through Tipping Point, NASA seeks to support space technologies that can foster the growth of commercial space capabilities and benefit future agency missions. NASA is also offering businesses a chance to work with agency experts or use facilities to complete their work through a separate Announcement of Collaboration Opportunity.

“NASA’s investment and support at this pivotal stage in development can be the key to ultimately bringing new technologies to market,” said NASA Administrator Bill Nelson. “Public-private partnerships established through these opportunities will combine agency resources with industry contributions, a benefit that will unlock new ideas to advance how we live and work in space while providing greater value to the American people.”

These opportunities focus on technology development for space infrastructure and capabilities for the Moon and near-Earth space. Selected proposals for working on and near the Moon could include infrastructure for power distribution on the lunar surface, solutions for using lunar resources, or autonomous construction – key components for long-term lunar exploration under Artemis. NASA will also consider proposals for infrastructure and capabilities in Earth orbit – which could range from climate research tools to in-space manufacturing and advanced propulsion.

“Forging strong partnerships with the commercial space economy is critical for driving our nation’s space exploration further and faster,” said Jim Reuter, associate administrator for the Space Technology Mission Directorate (STMD) at NASA’s Headquarters in Washington, which manages the new opportunities. “I look forward to the new slate of public-private partnerships that will come from these new opportunities.”

Under the Tipping Point opportunity, NASA will award a total of up to $200 million to multiple companies using funded Space Act Agreements. The opportunity also includes incentives for small businesses, allowing companies with fewer than 500 employees to contribute less to the cost of the technology development.

“A funded Space Act Agreement provides more flexibility for commercial partners in intellectual property, private sector contribution, and accounting requirements,” said LK Kubendran, STMD lead on Tipping Point and Announcement of Collaboration Opportunity.

More than half a billion dollars have been awarded to 50 projects since NASA announced the first Tipping Point opportunity in 2015. Space technologies advanced through this funding are now part of current and future mission plans. They include:

• Maxar’s Space Infrastructure Dexterous Robot aboard NASA’s On-orbit Servicing, Assembly and Manufacturing Mission-1 mission will demonstrate in-space assembly to form a communications antenna following OSAM-1’s launch
• Tethers Unlimited’s Hydros thruster flew on NASA’s first Pathfinder Technology Demonstrator CubeSat mission to demonstrate the use of liquid water as fuel in space
• Two technologies will be aboard the second Intuitive Machines flight to the Moon under the Commercial Lunar Payload Services initiative: Nokia of America Corporation’s lunar LTE/4G communications system and Intuitive Machines’ hopper robot

The Announcement of Collaboration Opportunity allows companies to use NASA facilities and expertise to aid in technology development without a monetary award.

More information about NASA’s technology investments is available online.

The deadline to submit initial proposals is March 31, 2022. Companies interested in submitting proposals can also attend an industry information session on Feb. 28. Details on the event are available online.

Astronomers are increasingly drawing back the curtains on black holes. In the past few years, we have finally captured actual photos of these fearsome creatures and measured the gravitational waves – ripples in spacetime – that they create when colliding. But there’s still a lot we don’t know about black holes. One of the biggest enigmas is exactly how they form in the first place.

My colleagues and I now believe we have observed this process, providing some of the best indications yet of exactly what happens when a black hole forms. Our results are published in two papers in Nature and the Astrophysical Journal.

Astronomers believe, on both observational and theoretical grounds, that most black holes form when the center of a massive star collapses at the end of its life. The star’s core normally provides pressure, or support, using heat from intense nuclear reactions. But once such a star’s fuel is exhausted and nuclear reactions stop, the inner layers of the star collapse inward under gravity, crushing down to extraordinary densities.

First image of a black hole. Credit: EHT

Most of the time, this catastrophic collapse is halted when the star’s core condenses into a solid sphere of matter, rich in particles called neutrons. This leads to a powerful rebound explosion that destroys the star (a supernova), and leaves behind an exotic object known as a neutron star. But models of dying stars show that if the original star is massive enough (40-50 times the mass of the Sun), the collapse will simply continue unabated until the star is crushed down into a gravitational singularity – a black hole.

Explosive theories

While stars collapsing to form neutron stars are now routinely observed throughout the universe (supernova surveys find dozens of new ones every night), astronomers are not yet entirely sure what happens during the collapse to a black hole. Some pessimistic models suggest the entire star would be swallowed up without much of a trace. Others propose that the collapse to a black hole would produce some other kind of explosion.

For example, if the star is rotating at the time of collapse, some of the infalling material may be focused into jets that escape the star at high velocity. While these jets wouldn’t contain much mass, they’d pack a big punch: if they slammed into something, the effects might be quite dramatic in terms of the energy released.

Up until now, the best candidate for an explosion from the birth of a black hole has been the strange phenomenon known as long-duration gamma-ray bursts. First discovered in the 1960s by military satellites, these events have been hypothesized to result from jets accelerated to mindboggling speeds by newly formed black holes in collapsing stars. However, a longstanding problem with this scenario is that gamma-ray bursts also expel abundant radioactive debris that continues to shine for months. This suggests most of the star exploded outward into space (as in an ordinary supernova), instead of collapsing inward to a black hole.

While this doesn’t mean a black hole can’t have been formed in such an explosion, some have concluded that other models provide a more natural explanation for gamma ray bursts than a black hole forming. For example, a super-magnetized neutron star could form in such an explosion and produce powerful jets of its own.

Mystery solved?

My colleagues and I, however, recently uncovered a new and (in our view) much better candidate event for creating a black hole. On two separate occasions in the past three years – once in 2019 and once in 2021 – we witnessed an exceptionally fast and fleeting type of explosion that, much like in gamma-ray bursts, originated from a small amount of very fast-moving material slamming into gas in its immediate environment.

By using spectroscopy – a technique that breaks down light into different wavelengths – we could infer the composition of the star that exploded for each of these events. We discovered that the spectrum was very similar to so-called “Wolf-Rayet stars” – a very massive and highly-evolved type of star, named after the two astronomers, Charles Wolf and Georges Rayet, that first detected them. Excitingly, we were even able to rule out a “normal” supernova explosion. As soon as the collision between the fast material and its environment ceased, the source practically vanished – rather than glowing for a long time.

This is exactly what you would expect if, during the collapse of its core, the star ejected only a small amount of material with the rest of the object collapsing downward into an enormous black hole.

The new study observed two events that may belong to third type of explosion, lasting only a short time. Credit: Bill Saxton, NRAO/AUI/NSF

While this is our favoured interpretation, it’s not the only possibility. The most prosaic one is that it was a normal supernova explosion, but that a vast shell of dust formed in the collision, concealing the radioactive debris from view. It’s also possible that the explosion is of a new and unfamiliar type, originating from a star we’re not familiar with.

To answer these questions, we will need to search for more such objects. Until now these kinds of explosions have been difficult to study because they are fleeting and hard to find. We had to use several observatories together in quick succession to characterize these explosions: the Zwicky Transient Facility to discover them, the Liverpool Telescope and the Nordic Optical Telescope to confirm their nature, and large high-resolution observatories (the Hubble Space Telescope, Gemini Observatory, and the Very Large Telescope) to analyze their composition.

While we didn’t initially know exactly what we were seeing when we first discovered these events, we now have a clear hypothesis: the birth of a black hole.

More data from similar events may soon be able to help us verify or falsify this hypothesis and establish the link to other types of unusual, fast explosions that our team and others have been finding. Either way, it seems this truly is the decade we crack the mysteries of black holes.

Written by Daniel Perley, Reader of Astrophysics, Liverpool John Moores University.

This article was first published in The Conversation.