A Vanderbilt engineering professor is leading part of an international initiative to create advanced artificial intelligence programs that will enable machines to learn progressively over a lifetime and share those experiences. Researchers hope the technology will allow machines to reuse information, adapt quickly to new conditions, and collaborate by sharing information. 

Soheil Kolouri, assistant professor of computer science, in partnership with Hamed Pirsiavash, associate professor of computer science at the University of California, Davis, will lead a research team focusing on continual machine learning mechanisms.

The prototype project, “Information Distillation for Embodied and Articulate Lifelong Learners,” or IDEALL, has received a $1M award from the Defense Advanced Research Projects Agency as part of the agency’s Shared-Experience Lifelong Learning (ShELL) initiative. DARPA wants to develop AI agents that share their experiences and is seeking innovative basic or applied research concepts in lifelong learning.

In addition to leading IDEALL, Kolouri’s team has partnered with Andrea Soltoggio, associate professor of computer science, University of Loughborough, UK, to develop a theoretical framework that allows AI agents to measure tasks’ similarities and continually learn by analogies. Cong Liu, associate professor of computer science at, University of Texas, Dallas, is a member of the Loughborough team.  

The Vanderbilt-UC, Davis team will concentrate on the algorithmic theory and statistical foundation of the learning mechanisms. The UK team will focus on novel bio-inspired neural networks that learn shareable knowledge exploiting neuromodulation and synaptic consolidation mechanisms, and the Texas researchers will focus on the hardware integration and deployment for potential transition to industrial and real-world applications.

The real-world uses of this new technology could include cooperating self-learning autonomous vehicles such as self-driving cars, robotic rescue and exploration systems, distributed monitoring systems to detect emergencies, or cyber security systems of agents that monitor large networks.

Lifelong Learning is a relatively new area of machine learning research in which agents continually learn as they encounter varying conditions and tasks while deployed in the field, acquiring experience and knowledge and improving performance on both novel and previous tasks. This differs from the train-then-deploy process for typical ML systems.

LL is an emerging area of machine learning that differs from the traditional train and then deploy process. In LL, an AI agent must continually learn from the input data stream while preserving and improving its previously acquired knowledge.

“Lifelong learning from the never-ending stream of everchanging data is the key to scaling up AI systems,” said Kolouri. “One of the major roadblocks in achieving LL is the so-called plasticity-stability trade-off, where plasticity refers to the ability to learn from new data, and stability refers to retaining the previously learned knowledge.” A team of undergraduate and graduate students at the Machine Intelligence and Neural Technologies (MINT) Lab directed by Kolouri, in collaboration with computer science associate professor Vladimir Braverman’s group at Johns Hopkins University, is currently studying this phenomenon.

“Today we know the importance of social interactions in the evolution of human intelligence. Artificial General Intelligence (AGI) could not be realized with a single AI agent. Similar to the cognitive revolution in sapiens, a transition is needed from our current single-agent LL to articulate LL machines that can encode information about their surroundings into a compact compositional language and use it for machine-to-machine communication and, maybe more importantly, for thinking, which is a form of self-communication!” said Kolouri. “The ShELL program aims to develop such communicative LL agents that continually learn from their collective experiences.”

“We are very excited to be part of this fast-paced, innovative program and look forward to transitioning our developed tools into medical applications,” Kolouri said.

Superconducting coupling between two regions separated by a one-micron wide ferromagnetic compound has been proved by an international team. This macroscopic quantum effect, known as Josephson effect, generates an electrical current within the ferromagnetic compound made of superconducting Cooper-pairs. Magnetic imaging of the ferromagnetic region at BESSY II has contributed to demonstrating that the spin of the electrons forming the Cooper pairs is equal. These results pave the way for low-power consumption superconducting spintronic-applications where spin-polarized currents can be protected by quantum coherence. 

When two superconducting regions are separated by a strip of non-superconducting material, a special quantum effect can occur, coupling both regions: The Josephson effect. If the spacer material is a half-metal ferromagnet novel implications for spintronic applications arise. An international team has now for the first time designed a material system that exhibits an unusually long-range Josephson effect: Here, regions of superconducting YBa2Cu3O7 are separated by a region of half-metallic, ferromagnetic manganite (La2/3Sr1/3MnO3) one micron wide.

With the help of magneto-transport measurements, the researchers were able to demonstrate the presence of a supercurrent circulating through the manganite – this supercurrent is arising from the superconducting coupling between both superconducting regions, and thus a manifestation of a Josephson effect with a macroscopic long range.

Extremely rare: Triplett superconductivity

In addition, the scientists explored another interesting property with profound consequences for spintronic applications. In superconductors electrons pair together in so-called Cooper pairs. In the vast majority of superconducting materials, these pairs are composed of electrons with opposite spins to minimize the magnetic exchange field which is detrimental for the stabilization of superconductivity. The ferromagnet used by the international team has been a half-ferromagnet for which only one spin-type electron is allowed to circulate. The fact that a supercurrent has been detected within this material, implies that the Cooper pairs of this supercurrent must be composed of electrons having the same spin. This so-called “triplet” superconductivity is extremely rare.

Mapping magnetic domains at BESSY II

"At the XMCD-PEEM station at BESSY II, we mapped and measured the magnetic domains within the manganite spacer. We observed wide regions homogeneously magnetized and connecting the superconducting regions. Triplet spin pairs can propagate freely in these,” explains Dr. Sergio Valencia Molina, HZB physicist, who supervised the measurements at BESSY II. 

Superconducting currents flow without resistance which makes them very appealing for low-power consumption applications. In the present case, this current is made of electrons with equal spins. Such spin-polarized currents could be used in novel superconducting spintronic applications for the transport (over long distances) and reading/writing of information while profiting from the stability imposed by the macroscopic quantum coherence of the Josephson effect.

The new device made of the superconducting and ferromagnetic components, therefore, opens up opportunities for superconducting spintronics and new perspectives for quantum supercomputing.

An international team of scientists believes they may have found a molecular mechanism behind the extremely rare blood clots linked to adenovirus COVID-19 vaccines.

Scientists led by a team from Arizona State University, Cardiff University, and others worked with AstraZeneca to investigate vaccine-induced immune thrombotic thrombocytopenia (VITT), also known as thrombosis with thrombocytopenia syndrome (TTS), a life-threatening condition seen in a very small number of people after receiving the Oxford-AstraZeneca or Johnson & Johnson vaccines. 

“The mechanism which results in this condition, termed vaccine-induced immune thrombotic thrombocytopenia (VITT), was unknown,” said Abhishek Singharoy, an Arizona State University scientist and corresponding author of the study who teamed up to lead an international effort to tease out the details. So, a team quickly assembled to try to understand the problem more clearly. 

Together, they worked to solve the structural biology of the vaccine and see the molecular details that may be at play, utilizing ASU’s new cryo-EM facilities, and a state-of-the-art Titan Krios machine at ASU’s Eyring Materials Center at Arizona State University.

ASU scientists included several from the School of Molecular Sciences and Biodesign Institute: Ryan J. Boyd, Daipayan Sarkar, John Vant, Eric Wilson, Chloe D. Truong, Petra Fromme, Po-Lin Chiu, Dewight Williams, and Josh Vermaas (ASU alumnus now at Michigan State University). Mitesh Borad, Bolni M. Nagalo, and Alexander T. Baker were also part of the Arizona-based team.

The global team used state-of-the-art cryo-EM technology to analyze the AstraZeneca vaccine in minute detail to understand whether the ultra-rare side effect could be linked to the viral vector which is used in many vaccines, including those from Oxford/AstraZeneca and Johnson & Johnson.

Their findings suggest it is the viral vector – in this case, an adenovirus used to shuttle the coronavirus’ genetic material into cells – and the way it binds to platelet factor 4 (PF4) once injected that could be the potential mechanism.

In very rare cases, the scientists suggest, the viral vector may enter the bloodstream and bind to PF4, where the immune system then views this complex as foreign. They believe this misplaced immunity could result in the release of antibodies against PF4, which bind to and activate platelets, causing them to cluster together and triggering blood clots in a very small number of people after the vaccine is administered.

“It’s really critical to fully investigate the vector-host interactions of the vaccine at a mechanistic level,” said Singharoy. “This will assist in understanding both how the vaccine generates immunity, and how it may lead to any rare adverse events, such as VITT.”

Their findings are published today in the international journal Science Advances.

Professor Alan Parker, an expert in the use of adenoviruses for medical applications from Cardiff University’s School of Medicine, said: “VITT only happens in extremely rare cases because a chain of complex events needs to take place to trigger this ultra-rare side effect. Our data confirms PF4 can bind to adenoviruses, an important step in unraveling the mechanism underlying VITT. Establishing a mechanism could help to prevent and treat this disorder.”

“We hope our findings can be used to better understand the rare side effects of these new vaccines – and potentially to design new and improved vaccines to turn the tide on this global pandemic.”

Both the AstraZeneca and Johnson & Johnson vaccines use an adenovirus to carry spike proteins from the coronavirus into people to trigger a protective immune response.

When both vaccines showed the ultra-rare side effect of VITT, scientists wondered whether the viral vector had some part to play. Another important clue was that neither the Moderna nor Pfizer vaccines, made from an entirely different technology called mRNA vaccines, showed this effect.

The team used cryo-EM technology to flash-freeze preparations of ChAdOx1, the adenovirus used in the AstraZeneca vaccine, and bombard them with electrons to produce microscopic images of the vaccine components.

They were then able to look at the atomic level at the structure of the outer protein cage of the virus – the viral capsid – and other critical proteins that allow entry of the virus into the cell.

In particular, the team outlined the details for the structure and receptor of ChAdOx1, which is adapted from chimpanzee adenovirus Y25 – and how it interacts with PF4. They believe it is this specific interaction – and how it is then presented to the immune system – that could prompt the body’s defenses to view it as foreign and release antibodies against this self-protein.

The research team also used the computational models of Singharoy to show that one of the ways the two molecules tightly bind is via electrostatic interactions. The group showed that ChAdOx1 is mostly electronegative. This makes the protein act like the negative end of a battery terminal and could attract other positively charged molecules to its surface.

The first author of the study Dr. Alexander Baker, said: “We found that ChAdOx1 has a strong negative charge. This means the viral vector can act as a magnet and attract proteins with the opposite, positive charge, like PF4.” Baker is a member of ASU’s Biodesign Center for Applied Structural Discovery and an Honorary Research Fellow at Cardiff University School of Medicine.

“We then found that PF4 is just the right size and shape that when it gets close to ChAdOx1 it could bind in between the negatively charged parts of ChAdOx1’s surface, called hexons.”

The research team is hopeful that armed with a better understanding of what may be causing rare VITT they can provide further insights into how vaccines and other therapies, which rely on the same technology, might be altered in the development of the next-generation vaccines and therapies.

“With a better understanding of the mechanism by which PF4 and adenoviruses interact there is an opportunity to engineer the shell of the vaccine, the capsid, to prevent this interaction with PF4. Modifying ChAdOx1 to reduce the negative charge may reduce the chance of causing thrombosis with thrombocytopenia syndrome,” said Baker.

The team likens it to the two birds, one stone effect. The key contacts of individual amino acids that are essential to the capsid protein’s proteins interaction with PF4 can be removed or substituted.

“The modification of the ChAdOx1 hexons to reduce their electronegativity may solve two problems simultaneously: reduce the propensity to cause VITT to even lower levels, and reduce the levels of pre-existing immunity, thus helping to maximize the opportunity to induce robust immune responses, said Singharoy.”

Both the UK-based Medicines and Healthcare products Regulatory Agency (MHRA) and Centers for Disease Control and Prevention (CDC) in the U.S. continue to advise that vaccination is the best way to protect people from COVID-19 and the benefits far outweigh the risk of any known side effects.

Breast cancer is the most common cancer with the highest mortality rate. Swift detection and diagnosis diminish the impact of the disease. However, classifying breast cancer using histopathology images — tissues and cells examined under a microscope — is a challenging task because of bias in the data and the unavailability of annotated data in large quantities. Automatic detection of breast cancer using convolutional neural network (CNN), a machine learning technique, has shown promise — but it is associated with a high risk of false positives and false negatives.

Without any measure of confidence, such false predictions of CNN could lead to catastrophic outcomes. But a new machine learning model developed by Michigan Technological University researchers can evaluate the uncertainty in its predictions as it classifies benign and malignant tumors, helping reduce this risk. 

In their study, mechanical engineering graduate students Ponkrshnan Thiagarajan and Pushkar Kharinar and Susanta Ghosh, assistant professor of mechanical engineering and machine learning expert, outline their novel probabilistic machine learning model, which outperforms similar models.

“Any machine learning algorithm that has been developed so far will have some uncertainty in its prediction,” Thiagarajan said. “There is little way to quantify those uncertainties. Even if an algorithm tells us a person has cancer, we do not know the level of confidence in that prediction.”

From Experience Comes Confidence

In the medical context, not knowing how confident an algorithm is has made it difficult to rely on computer-generated predictions. The present model is an extension of the Bayesian neural network — a machine learning model that can evaluate an image and produce an output. The parameters for this model are treated as random variables that facilitate uncertainty quantification. 

The Michigan Tech model differentiates between negative and positive classes by analyzing the images, which at their most basic level are collections of pixels. In addition to this classification, the model can measure the uncertainty in its predictions.

In a medical laboratory, such a model promises time savings by classifying images faster than a lab tech. And, because the model can evaluate its level of certainty, it can refer the images to a human expert when it is less confident.

But why is a mechanical engineer creating algorithms for the medical community? Thiagarajan’s idea kindled when he started using machine learning to reduce the computational time needed for mechanical engineering problems. Whether a computation evaluates the deformation of building materials or determines whether someone has breast cancer, it’s important to know the uncertainty of that computation — the key ideas remain the same.

“Breast cancer is one of the cancers that have the highest mortality and highest incidence,” Thiagarajan said. “We believe that this is an exciting problem wherein better algorithms can make an impact on people’s lives directly.”

Next Steps

Now that their study has been published, the researchers will extend the model for multiclass classification of breast cancer. They will aim to detect cancer subtypes in addition to classifying benign and malignant tissues. And the model, though developed using breast cancer histopathology images, can also be extended for other medical diagnoses.

“Despite the promise of machine learning-based classification models, their predictions suffer from uncertainties due to the inherent randomness and the bias in the data and the scarcity of large datasets,” Ghosh said. “Our work attempts to address these issues and quantifies, uses, and explains the uncertainty.”

Ultimately, Thiagarajan, Khairnar, and Ghosh’s model itself — which can evaluate whether images have high or low measures of uncertainty and identify when images need the eyes of a medical expert — represents the next steps in the endeavor of machine learning.