Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 100 Shares149 Views

      in Data Management & Statistics

      Subtle biases in AI can influence emergency decisions

      by

      It’s no secret that people harbor biases — some unconscious, perhaps, and others painfully overt. The average person might suppose that computers — machines typically made of plastic, steel, glass, silicon, and various metals — are free of prejudice. While that assumption may hold for computer hardware, the same is not always true for computer software, which is programmed by fallible humans and can be fed data that is, itself, compromised in certain respects.

      Artificial intelligence (AI) systems — those based on machine learning, in particular — are seeing increased use in medicine for diagnosing specific diseases, for example, or evaluating X-rays. These systems are also being relied on to support decision-making in other areas of health care. Recent research has shown, however, that machine learning models can encode biases against minority subgroups, and the recommendations they make may consequently reflect those same biases.

      A new study by researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Jameel Clinic, which was published last month in Communications Medicine, assesses the impact that discriminatory AI models can have, especially for systems that are intended to provide advice in urgent situations. “We found that the manner in which the advice is framed can have significant repercussions,” explains the paper’s lead author, Hammaad Adam, a PhD student at MIT’s Institute for Data Systems and Society. “Fortunately, the harm caused by biased models can be limited (though not necessarily eliminated) when the advice is presented in a different way.” The other co-authors of the paper are Aparna Balagopalan and Emily Alsentzer, both PhD students, and the professors Fotini Christia and Marzyeh Ghassemi.

      AI models used in medicine can suffer from inaccuracies and inconsistencies, in part because the data used to train the models are often not representative of real-world settings. Different kinds of X-ray machines, for instance, can record things differently and hence yield different results. Models trained predominately on white people, moreover, may not be as accurate when applied to other groups. The Communications Medicine paper is not focused on issues of that sort but instead addresses problems that stem from biases and on ways to mitigate the adverse consequences.

      A group of 954 people (438 clinicians and 516 nonexperts) took part in an experiment to see how AI biases can affect decision-making. The participants were presented with call summaries from a fictitious crisis hotline, each involving a male individual undergoing a mental health emergency. The summaries contained information as to whether the individual was Caucasian or African American and would also mention his religion if he happened to be Muslim. A typical call summary might describe a circumstance in which an African American man was found at home in a delirious state, indicating that “he has not consumed any drugs or alcohol, as he is a practicing Muslim.” Study participants were instructed to call the police if they thought the patient was likely to turn violent; otherwise, they were encouraged to seek medical help.

      The participants were randomly divided into a control or “baseline” group plus four other groups designed to test responses under slightly different conditions. “We want to understand how biased models can influence decisions, but we first need to understand how human biases can affect the decision-making process,” Adam notes. What they found in their analysis of the baseline group was rather surprising: “In the setting we considered, human participants did not exhibit any biases. That doesn’t mean that humans are not biased, but the way we conveyed information about a person’s race and religion, evidently, was not strong enough to elicit their biases.”

      The other four groups in the experiment were given advice that either came from a biased or unbiased model, and that advice was presented in either a “prescriptive” or a “descriptive” form. A biased model would be more likely to recommend police help in a situation involving an African American or Muslim person than would an unbiased model. Participants in the study, however, did not know which kind of model their advice came from, or even that models delivering the advice could be biased at all. Prescriptive advice spells out what a participant should do in unambiguous terms, telling them they should call the police in one instance or seek medical help in another. Descriptive advice is less direct: A flag is displayed to show that the AI system perceives a risk of violence associated with a particular call; no flag is shown if the threat of violence is deemed small.  

      A key takeaway of the experiment is that participants “were highly influenced by prescriptive recommendations from a biased AI system,” the authors wrote. But they also found that “using descriptive rather than prescriptive recommendations allowed participants to retain their original, unbiased decision-making.” In other words, the bias incorporated within an AI model can be diminished by appropriately framing the advice that’s rendered. Why the different outcomes, depending on how advice is posed? When someone is told to do something, like call the police, that leaves little room for doubt, Adam explains. However, when the situation is merely described — classified with or without the presence of a flag — “that leaves room for a participant’s own interpretation; it allows them to be more flexible and consider the situation for themselves.”

      Second, the researchers found that the language models that are typically used to offer advice are easy to bias. Language models represent a class of machine learning systems that are trained on text, such as the entire contents of Wikipedia and other web material. When these models are “fine-tuned” by relying on a much smaller subset of data for training purposes — just 2,000 sentences, as opposed to 8 million web pages — the resultant models can be readily biased.  

      Third, the MIT team discovered that decision-makers who are themselves unbiased can still be misled by the recommendations provided by biased models. Medical training (or the lack thereof) did not change responses in a discernible way. “Clinicians were influenced by biased models as much as non-experts were,” the authors stated.

      “These findings could be applicable to other settings,” Adam says, and are not necessarily restricted to health care situations. When it comes to deciding which people should receive a job interview, a biased model could be more likely to turn down Black applicants. The results could be different, however, if instead of explicitly (and prescriptively) telling an employer to “reject this applicant,” a descriptive flag is attached to the file to indicate the applicant’s “possible lack of experience.”

      The implications of this work are broader than just figuring out how to deal with individuals in the midst of mental health crises, Adam maintains.  “Our ultimate goal is to make sure that machine learning models are used in a fair, safe, and robust way.” More

    • 88 Shares179 Views

      in Data Management & Statistics

      A faster way to preserve privacy online

      by

      Searching the internet can reveal information a user would rather keep private. For instance, when someone looks up medical symptoms online, they could reveal their health conditions to Google, an online medical database like WebMD, and perhaps hundreds of these companies’ advertisers and business partners.

      For decades, researchers have been crafting techniques that enable users to search for and retrieve information from a database privately, but these methods remain too slow to be effectively used in practice.

      MIT researchers have now developed a scheme for private information retrieval that is about 30 times faster than other comparable methods. Their technique enables a user to search an online database without revealing their query to the server. Moreover, it is driven by a simple algorithm that would be easier to implement than the more complicated approaches from previous work.

      Their technique could enable private communication by preventing a messaging app from knowing what users are saying or who they are talking to. It could also be used to fetch relevant online ads without advertising servers learning a users’ interests.

      “This work is really about giving users back some control over their own data. In the long run, we’d like browsing the web to be as private as browsing a library. This work doesn’t achieve that yet, but it starts building the tools to let us do this sort of thing quickly and efficiently in practice,” says Alexandra Henzinger, a computer science graduate student and lead author of a paper introducing the technique.

      Co-authors include Matthew Hong, an MIT computer science graduate student; Henry Corrigan-Gibbs, the Douglas Ross Career Development Professor of Software Technology in the MIT Department of Electrical Engineering and Computer Science (EECS) and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL); Sarah Meiklejohn, a professor in cryptography and security at University College London and a staff research scientist at Google; and senior author Vinod Vaikuntanathan, an EECS professor and principal investigator in CSAIL. The research will be presented at the 2023 USENIX Security Symposium. 

      Preserving privacy

      The first schemes for private information retrieval were developed in the 1990s, partly by researchers at MIT. These techniques enable a user to communicate with a remote server that holds a database, and read records from that database without the server knowing what the user is reading.

      To preserve privacy, these techniques force the server to touch every single item in the database, so it can’t tell which entry a user is searching for. If one area is left untouched, the server would learn that the client is not interested in that item. But touching every item when there may be millions of database entries slows down the query process.

      To speed things up, the MIT researchers developed a protocol, known as Simple PIR, in which the server performs much of the underlying cryptographic work in advance, before a client even sends a query. This preprocessing step produces a data structure that holds compressed information about the database contents, and which the client downloads before sending a query.

      In a sense, this data structure is like a hint for the client about what is in the database.

      “Once the client has this hint, it can make an unbounded number of queries, and these queries are going to be much smaller in both the size of the messages you are sending and the work that you need the server to do. This is what makes Simple PIR so much faster,” Henzinger explains.

      But the hint can be relatively large in size. For example, to query a 1-gigabyte database, the client would need to download a 124-megabyte hint. This drives up communication costs, which could make the technique difficult to implement on real-world devices.

      To reduce the size of the hint, the researchers developed a second technique, known as Double PIR, that basically involves running the Simple PIR scheme twice. This produces a much more compact hint that is fixed in size for any database.

      Using Double PIR, the hint for a 1 gigabyte database would only be 16 megabytes.

      “Our Double PIR scheme runs a little bit slower, but it will have much lower communication costs. For some applications, this is going to be a desirable tradeoff,” Henzinger says.

      Hitting the speed limit

      They tested the Simple PIR and Double PIR schemes by applying them to a task in which a client seeks to audit a specific piece of information about a website to ensure that website is safe to visit. To preserve privacy, the client cannot reveal the website it is auditing.

      The researchers’ fastest technique was able to successfully preserve privacy while running at about 10 gigabytes per second. Previous schemes could only achieve a throughput of about 300 megabytes per second.

      They show that their method approaches the theoretical speed limit for private information retrieval — it is nearly the fastest possible scheme one can build in which the server touches every record in the database, adds Corrigan-Gibbs.

      In addition, their method only requires a single server, making it much simpler than many top-performing techniques that require two separate servers with identical databases. Their method outperformed these more complex protocols.

      “I’ve been thinking about these schemes for some time, and I never thought this could be possible at this speed. The folklore was that any single-server scheme is going to be really slow. This work turns that whole notion on its head,” Corrigan-Gibbs says.

      While the researchers have shown that they can make PIR schemes much faster, there is still work to do before they would be able to deploy their techniques in real-world scenarios, says Henzinger. They would like to cut the communication costs of their schemes while still enabling them to achieve high speeds. In addition, they want to adapt their techniques to handle more complex queries, such as general SQL queries, and more demanding applications, such as a general Wikipedia search. And in the long run, they hope to develop better techniques that can preserve privacy without requiring a server to touch every database item. 

      “I’ve heard people emphatically claiming that PIR will never be practical. But I would never bet against technology. That is an optimistic lesson to learn from this work. There are always ways to innovate,” Vaikuntanathan says.

      “This work makes a major improvement to the practical cost of private information retrieval. While it was known that low-bandwidth PIR schemes imply public-key cryptography, which is typically orders of magnitude slower than private-key cryptography, this work develops an ingenious method to bridge the gap. This is done by making a clever use of special properties of a public-key encryption scheme due to Regev to push the vast majority of the computational work to a precomputation step, in which the server computes a short ‘hint’ about the database,” says Yuval Ishai, a professor of computer science at Technion (the Israel Institute of Technology), who was not involved in the study. “What makes their approach particularly appealing is that the same hint can be used an unlimited number of times, by any number of clients. This renders the (moderate) cost of computing the hint insignificant in a typical scenario where the same database is accessed many times.”

      This work is funded, in part, by the National Science Foundation, Google, Facebook, MIT’s Fintech@CSAIL Initiative, an NSF Graduate Research Fellowship, an EECS Great Educators Fellowship, the National Institutes of Health, the Defense Advanced Research Projects Agency, the MIT-IBM Watson AI Lab, Analog Devices, Microsoft, and a Thornton Family Faculty Research Innovation Fellowship. More

    • 88 Shares129 Views

      in Data Management & Statistics

      Large language models help decipher clinical notes

      by

      Electronic health records (EHRs) need a new public relations manager. Ten years ago, the U.S. government passed a law that required hospitals to digitize their health records with the intent of improving and streamlining care. The enormous amount of information in these now-digital records could be used to answer very specific questions beyond the scope of clinical trials: What’s the right dose of this medication for patients with this height and weight? What about patients with a specific genomic profile?

      Unfortunately, most of the data that could answer these questions is trapped in doctor’s notes, full of jargon and abbreviations. These notes are hard for computers to understand using current techniques — extracting information requires training multiple machine learning models. Models trained for one hospital, also, don’t work well at others, and training each model requires domain experts to label lots of data, a time-consuming and expensive process. 

      An ideal system would use a single model that can extract many types of information, work well at multiple hospitals, and learn from a small amount of labeled data. But how? Researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) believed that to disentangle the data, they needed to call on something bigger: large language models. To pull that important medical information, they used a very big, GPT-3 style model to do tasks like expand overloaded jargon and acronyms and extract medication regimens. 

      For example, the system takes an input, which in this case is a clinical note, “prompts” the model with a question about the note, such as “expand this abbreviation, C-T-A.” The system returns an output such as “clear to auscultation,” as opposed to say, a CT angiography. The objective of extracting this clean data, the team says, is to eventually enable more personalized clinical recommendations. 

      Medical data is, understandably, a pretty tricky resource to navigate freely. There’s plenty of red tape around using public resources for testing the performance of large models because of data use restrictions, so the team decided to scrape together their own. Using a set of short, publicly available clinical snippets, they cobbled together a small dataset to enable evaluation of the extraction performance of large language models. 

      “It’s challenging to develop a single general-purpose clinical natural language processing system that will solve everyone’s needs and be robust to the huge variation seen across health datasets. As a result, until today, most clinical notes are not used in downstream analyses or for live decision support in electronic health records. These large language model approaches could potentially transform clinical natural language processing,” says David Sontag, MIT professor of electrical engineering and computer science, principal investigator in CSAIL and the Institute for Medical Engineering and Science, and supervising author on a paper about the work, which will be presented at the Conference on Empirical Methods in Natural Language Processing. “The research team’s advances in zero-shot clinical information extraction makes scaling possible. Even if you have hundreds of different use cases, no problem — you can build each model with a few minutes of work, versus having to label a ton of data for that particular task.”

      For example, without any labels at all, the researchers found these models could achieve 86 percent accuracy at expanding overloaded acronyms, and the team developed additional methods to boost this further to 90 percent accuracy, with still no labels required.

      Imprisoned in an EHR 

      Experts have been steadily building up large language models (LLMs) for quite some time, but they burst onto the mainstream with GPT-3’s widely covered ability to complete sentences. These LLMs are trained on a huge amount of text from the internet to finish sentences and predict the next most likely word. 

      While previous, smaller models like earlier GPT iterations or BERT have pulled off a good performance for extracting medical data, they still require substantial manual data-labeling effort. 

      For example, a note, “pt will dc vanco due to n/v” means that this patient (pt) was taking the antibiotic vancomycin (vanco) but experienced nausea and vomiting (n/v) severe enough for the care team to discontinue (dc) the medication. The team’s research avoids the status quo of training separate machine learning models for each task (extracting medication, side effects from the record, disambiguating common abbreviations, etc). In addition to expanding abbreviations, they investigated four other tasks, including if the models could parse clinical trials and extract detail-rich medication regimens.  

      “Prior work has shown that these models are sensitive to the prompt’s precise phrasing. Part of our technical contribution is a way to format the prompt so that the model gives you outputs in the correct format,” says Hunter Lang, CSAIL PhD student and author on the paper. “For these extraction problems, there are structured output spaces. The output space is not just a string. It can be a list. It can be a quote from the original input. So there’s more structure than just free text. Part of our research contribution is encouraging the model to give you an output with the correct structure. That significantly cuts down on post-processing time.”

      The approach can’t be applied to out-of-the-box health data at a hospital: that requires sending private patient information across the open internet to an LLM provider like OpenAI. The authors showed that it’s possible to work around this by distilling the model into a smaller one that could be used on-site.

      The model — sometimes just like humans — is not always beholden to the truth. Here’s what a potential problem might look like: Let’s say you’re asking the reason why someone took medication. Without proper guardrails and checks, the model might just output the most common reason for that medication, if nothing is explicitly mentioned in the note. This led to the team’s efforts to force the model to extract more quotes from data and less free text.

      Future work for the team includes extending to languages other than English, creating additional methods for quantifying uncertainty in the model, and pulling off similar results with open-sourced models. 

      “Clinical information buried in unstructured clinical notes has unique challenges compared to general domain text mostly due to large use of acronyms, and inconsistent textual patterns used across different health care facilities,” says Sadid Hasan, AI lead at Microsoft and former executive director of AI at CVS Health, who was not involved in the research. “To this end, this work sets forth an interesting paradigm of leveraging the power of general domain large language models for several important zero-/few-shot clinical NLP tasks. Specifically, the proposed guided prompt design of LLMs to generate more structured outputs could lead to further developing smaller deployable models by iteratively utilizing the model generated pseudo-labels.”

      “AI has accelerated in the last five years to the point at which these large models can predict contextualized recommendations with benefits rippling out across a variety of domains such as suggesting novel drug formulations, understanding unstructured text, code recommendations or create works of art inspired by any number of human artists or styles,” says Parminder Bhatia, who was formerly Head of Machine Learning at AWS Health AI and is currently Head of ML for low-code applications leveraging large language models at AWS AI Labs. “One of the applications of these large models [the team has] recently launched is Amazon CodeWhisperer, which is [an] ML-powered coding companion that helps developers in building applications.”

      As part of the MIT Abdul Latif Jameel Clinic for Machine Learning in Health, Agrawal, Sontag, and Lang wrote the paper alongside Yoon Kim, MIT assistant professor and CSAIL principal investigator, and Stefan Hegselmann, a visiting PhD student from the University of Muenster. First-author Agrawal’s research was supported by a Takeda Fellowship, the MIT Deshpande Center for Technological Innovation, and the MLA@CSAIL Initiatives. More

    • 163 Shares129 Views

      in Data Management & Statistics

      Busy GPUs: Sampling and pipelining method speeds up deep learning on large graphs

      by

      Graphs, a potentially extensive web of nodes connected by edges, can be used to express and interrogate relationships between data, like social connections, financial transactions, traffic, energy grids, and molecular interactions. As researchers collect more data and build out these graphical pictures, researchers will need faster and more efficient methods, as well as more computational power, to conduct deep learning on them, in the way of graph neural networks (GNN).  

      Now, a new method, called SALIENT (SAmpling, sLIcing, and data movemeNT), developed by researchers at MIT and IBM Research, improves the training and inference performance by addressing three key bottlenecks in computation. This dramatically cuts down on the runtime of GNNs on large datasets, which, for example, contain on the scale of 100 million nodes and 1 billion edges. Further, the team found that the technique scales well when computational power is added from one to 16 graphical processing units (GPUs). The work was presented at the Fifth Conference on Machine Learning and Systems.

      “We started to look at the challenges current systems experienced when scaling state-of-the-art machine learning techniques for graphs to really big datasets. It turned out there was a lot of work to be done, because a lot of the existing systems were achieving good performance primarily on smaller datasets that fit into GPU memory,” says Tim Kaler, the lead author and a postdoc in the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL).

      By vast datasets, experts mean scales like the entire Bitcoin network, where certain patterns and data relationships could spell out trends or foul play. “There are nearly a billion Bitcoin transactions on the blockchain, and if we want to identify illicit activities inside such a joint network, then we are facing a graph of such a scale,” says co-author Jie Chen, senior research scientist and manager of IBM Research and the MIT-IBM Watson AI Lab. “We want to build a system that is able to handle that kind of graph and allows processing to be as efficient as possible, because every day we want to keep up with the pace of the new data that are generated.”

      Kaler and Chen’s co-authors include Nickolas Stathas MEng ’21 of Jump Trading, who developed SALIENT as part of his graduate work; former MIT-IBM Watson AI Lab intern and MIT graduate student Anne Ouyang; MIT CSAIL postdoc Alexandros-Stavros Iliopoulos; MIT CSAIL Research Scientist Tao B. Schardl; and Charles E. Leiserson, the Edwin Sibley Webster Professor of Electrical Engineering at MIT and a researcher with the MIT-IBM Watson AI Lab.     

      For this problem, the team took a systems-oriented approach in developing their method: SALIENT, says Kaler. To do this, the researchers implemented what they saw as important, basic optimizations of components that fit into existing machine-learning frameworks, such as PyTorch Geometric and the deep graph library (DGL), which are interfaces for building a machine-learning model. Stathas says the process is like swapping out engines to build a faster car. Their method was designed to fit into existing GNN architectures, so that domain experts could easily apply this work to their specified fields to expedite model training and tease out insights during inference faster. The trick, the team determined, was to keep all of the hardware (CPUs, data links, and GPUs) busy at all times: while the CPU samples the graph and prepares mini-batches of data that will then be transferred through the data link, the more critical GPU is working to train the machine-learning model or conduct inference. 

      The researchers began by analyzing the performance of a commonly used machine-learning library for GNNs (PyTorch Geometric), which showed a startlingly low utilization of available GPU resources. Applying simple optimizations, the researchers improved GPU utilization from 10 to 30 percent, resulting in a 1.4 to two times performance improvement relative to public benchmark codes. This fast baseline code could execute one complete pass over a large training dataset through the algorithm (an epoch) in 50.4 seconds.                          

      Seeking further performance improvements, the researchers set out to examine the bottlenecks that occur at the beginning of the data pipeline: the algorithms for graph sampling and mini-batch preparation. Unlike other neural networks, GNNs perform a neighborhood aggregation operation, which computes information about a node using information present in other nearby nodes in the graph — for example, in a social network graph, information from friends of friends of a user. As the number of layers in the GNN increase, the number of nodes the network has to reach out to for information can explode, exceeding the limits of a computer. Neighborhood sampling algorithms help by selecting a smaller random subset of nodes to gather; however, the researchers found that current implementations of this were too slow to keep up with the processing speed of modern GPUs. In response, they identified a mix of data structures, algorithmic optimizations, and so forth that improved sampling speed, ultimately improving the sampling operation alone by about three times, taking the per-epoch runtime from 50.4 to 34.6 seconds. They also found that sampling, at an appropriate rate, can be done during inference, improving overall energy efficiency and performance, a point that had been overlooked in the literature, the team notes.      

      In previous systems, this sampling step was a multi-process approach, creating extra data and unnecessary data movement between the processes. The researchers made their SALIENT method more nimble by creating a single process with lightweight threads that kept the data on the CPU in shared memory. Further, SALIENT takes advantage of a cache of modern processors, says Stathas, parallelizing feature slicing, which extracts relevant information from nodes of interest and their surrounding neighbors and edges, within the shared memory of the CPU core cache. This again reduced the overall per-epoch runtime from 34.6 to 27.8 seconds.

      The last bottleneck the researchers addressed was to pipeline mini-batch data transfers between the CPU and GPU using a prefetching step, which would prepare data just before it’s needed. The team calculated that this would maximize bandwidth usage in the data link and bring the method up to perfect utilization; however, they only saw around 90 percent. They identified and fixed a performance bug in a popular PyTorch library that caused unnecessary round-trip communications between the CPU and GPU. With this bug fixed, the team achieved a 16.5 second per-epoch runtime with SALIENT.

      “Our work showed, I think, that the devil is in the details,” says Kaler. “When you pay close attention to the details that impact performance when training a graph neural network, you can resolve a huge number of performance issues. With our solutions, we ended up being completely bottlenecked by GPU computation, which is the ideal goal of such a system.”

      SALIENT’s speed was evaluated on three standard datasets ogbn-arxiv, ogbn-products, and ogbn-papers100M, as well as in multi-machine settings, with different levels of fanout (amount of data that the CPU would prepare for the GPU), and across several architectures, including the most recent state-of-the-art one, GraphSAGE-RI. In each setting, SALIENT outperformed PyTorch Geometric, most notably on the large ogbn-papers100M dataset, containing 100 million nodes and over a billion edges Here, it was three times faster, running on one GPU, than the optimized baseline that was originally created for this work; with 16 GPUs, SALIENT was an additional eight times faster. 

      While other systems had slightly different hardware and experimental setups, so it wasn’t always a direct comparison, SALIENT still outperformed them. Among systems that achieved similar accuracy, representative performance numbers include 99 seconds using one GPU and 32 CPUs, and 13 seconds using 1,536 CPUs. In contrast, SALIENT’s runtime using one GPU and 20 CPUs was 16.5 seconds and was just two seconds with 16 GPUs and 320 CPUs. “If you look at the bottom-line numbers that prior work reports, our 16 GPU runtime (two seconds) is an order of magnitude faster than other numbers that have been reported previously on this dataset,” says Kaler. The researchers attributed their performance improvements, in part, to their approach of optimizing their code for a single machine before moving to the distributed setting. Stathas says that the lesson here is that for your money, “it makes more sense to use the hardware you have efficiently, and to its extreme, before you start scaling up to multiple computers,” which can provide significant savings on cost and carbon emissions that can come with model training.

      This new capacity will now allow researchers to tackle and dig deeper into bigger and bigger graphs. For example, the Bitcoin network that was mentioned earlier contained 100,000 nodes; the SALIENT system can capably handle a graph 1,000 times (or three orders of magnitude) larger.

      “In the future, we would be looking at not just running this graph neural network training system on the existing algorithms that we implemented for classifying or predicting the properties of each node, but we also want to do more in-depth tasks, such as identifying common patterns in a graph (subgraph patterns), [which] may be actually interesting for indicating financial crimes,” says Chen. “We also want to identify nodes in a graph that are similar in a sense that they possibly would be corresponding to the same bad actor in a financial crime. These tasks would require developing additional algorithms, and possibly also neural network architectures.”

      This research was supported by the MIT-IBM Watson AI Lab and in part by the U.S. Air Force Research Laboratory and the U.S. Air Force Artificial Intelligence Accelerator. More

    • 113 Shares169 Views

      in Data Management & Statistics

      In-home wireless device tracks disease progression in Parkinson’s patients

      by

      Parkinson’s disease is the fastest-growing neurological disease, now affecting more than 10 million people worldwide, yet clinicians still face huge challenges in tracking its severity and progression.

      Clinicians typically evaluate patients by testing their motor skills and cognitive functions during clinic visits. These semisubjective measurements are often skewed by outside factors — perhaps a patient is tired after a long drive to the hospital. More than 40 percent of individuals with Parkinson’s are never treated by a neurologist or Parkinson’s specialist, often because they live too far from an urban center or have difficulty traveling.

      In an effort to address these problems, researchers from MIT and elsewhere demonstrated an in-home device that can monitor a patient’s movement and gait speed, which can be used to evaluate Parkinson’s severity, the progression of the disease, and the patient’s response to medication.

      The device, which is about the size of a Wi-Fi router, gathers data passively using radio signals that reflect off the patient’s body as they move around their home. The patient does not need to wear a gadget or change their behavior. (A recent study, for example, showed that this type of device could be used to detect Parkinson’s from a person’s breathing patterns while sleeping.)

      The researchers used these devices to conduct a one-year at-home study with 50 participants. They showed that, by using machine-learning algorithms to analyze the troves of data they passively gathered (more than 200,000 gait speed measurements), a clinician could track Parkinson’s progression and medication response more effectively than they would with periodic, in-clinic evaluations.

      “By being able to have a device in the home that can monitor a patient and tell the doctor remotely about the progression of the disease, and the patient’s medication response so they can attend to the patient even if the patient can’t come to the clinic — now they have real, reliable information — that actually goes a long way toward improving equity and access,” says senior author Dina Katabi, the Thuan and Nicole Pham Professor in the Department of Electrical Engineering and Computer Science (EECS), and a principle investigator in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the MIT Jameel Clinic.

      The co-lead authors are EECS graduate students Yingcheng Liu and Guo Zhang. The research is published today in Science Translational Medicine.

      A human radar

      This work utilizes a wireless device previously developed in the Katabi lab that analyzes radio signals that bounce off people’s bodies. It transmits signals that use a tiny fraction of the power of a Wi-Fi router — these super-low-power signals don’t interfere with other wireless devices in the home. While radio signals pass through walls and other solid objects, they are reflected off humans due to the water in our bodies.  

      This creates a “human radar” that can track the movement of a person in a room. Radio waves always travel at the same speed, so the length of time it takes the signals to reflect back to the device indicates how the person is moving.

      The device incorporates a machine-learning classifier that can pick out the precise radio signals reflected off the patient even when there are other people moving around the room. Advanced algorithms use these movement data to compute gait speed — how fast the person is walking.

      Because the device operates in the background and runs all day, every day, it can collect a massive amount of data. The researchers wanted to see if they could apply machine learning to these datasets to gain insights about the disease over time.

      They gathered 50 participants, 34 of whom had Parkinson’s, and conducted a one-year study of in-home gait measurements Through the study, the researchers collected more than 200,000 individual measurements that they averaged to smooth out variability due to the conditions irrelevant to the disease. (For example, a patient may hurry up to answer an alarm or walk slower when talking on the phone.)

      They used statistical methods to analyze the data and found that in-home gait speed can be used to effectively track Parkinson’s progression and severity. For instance, they showed that gait speed declined almost twice as fast for individuals with Parkinson’s, compared to those without. 

      “Monitoring the patient continuously as they move around the room enabled us to get really good measurements of their gait speed. And with so much data, we were able to perform aggregation that allowed us to see very small differences,” Zhang says.

      Better, faster results

      Drilling down on these variabilities offered some key insights. For instance, the researchers showed that daily fluctuations in a patient’s walking speed correspond with how they are responding to their medication — walking speed may improve after a dose and then begin to decline after a few hours, as the medication impact wears off.

      “This enables us to objectively measure how your mobility responds to your medication. Previously, this was very cumbersome to do because this medication effect could only be measured by having the patient keep a journal,” Liu says.

      A clinician could use these data to adjust medication dosage more effectively and accurately. This is especially important since drugs used to treat disease symptoms can cause serious side effects if the patient receives too much.

      The researchers were able to demonstrate statistically significant results regarding Parkinson’s progression after studying 50 people for just one year. By contrast, an often-cited study by the Michael J. Fox Foundation involved more than 500 individuals and monitored them for more than five years, Katabi says.

      “For a pharmaceutical company or a biotech company trying to develop medicines for this disease, this could greatly reduce the burden and cost and speed up the development of new therapies,” she adds.

      Katabi credits much of the study’s success to the dedicated team of scientists and clinicians who worked together to tackle the many difficulties that arose along the way. For one, they began the study before the Covid-19 pandemic, so team members initially visited people’s homes to set up the devices. When that was no longer possible, they developed a user-friendly phone app to remotely help participants as they deployed the device at home.

      Through the course of the study, they learned to automate processes and reduce effort, especially for the participants and clinical team.

      This knowledge will prove useful as they look to deploy devices in at-home studies of other neurological disorders, such as Alzheimer’s, ALS, and Huntington’s. They also want to explore how these methods could be used, in conjunction with other work from the Katabi lab showing that Parkinson’s can be diagnosed by monitoring breathing, to collect a holistic set of markers that could diagnose the disease early and then be used to track and treat it.

      “This radio-wave sensor can enable more care (and research) to migrate from hospitals to the home where it is most desired and needed,” says Ray Dorsey, a professor of neurology at the University of Rochester Medical Center, co-author of Ending Parkinson’s, and a co-author of this research paper. “Its potential is just beginning to be seen. We are moving toward a day where we can diagnose and predict disease at home. In the future, we may even be able to predict and ideally prevent events like falls and heart attacks.”

      This work is supported, in part, by the National Institutes of Health and the Michael J. Fox Foundation. More

    • 113 Shares159 Views

      in Data Management & Statistics

      AI that can learn the patterns of human language

      by

      Human languages are notoriously complex, and linguists have long thought it would be impossible to teach a machine how to analyze speech sounds and word structures in the way human investigators do.

      But researchers at MIT, Cornell University, and McGill University have taken a step in this direction. They have demonstrated an artificial intelligence system that can learn the rules and patterns of human languages on its own.

      When given words and examples of how those words change to express different grammatical functions (like tense, case, or gender) in one language, this machine-learning model comes up with rules that explain why the forms of those words change. For instance, it might learn that the letter “a” must be added to end of a word to make the masculine form feminine in Serbo-Croatian.

      This model can also automatically learn higher-level language patterns that can apply to many languages, enabling it to achieve better results.

      The researchers trained and tested the model using problems from linguistics textbooks that featured 58 different languages. Each problem had a set of words and corresponding word-form changes. The model was able to come up with a correct set of rules to describe those word-form changes for 60 percent of the problems.

      This system could be used to study language hypotheses and investigate subtle similarities in the way diverse languages transform words. It is especially unique because the system discovers models that can be readily understood by humans, and it acquires these models from small amounts of data, such as a few dozen words. And instead of using one massive dataset for a single task, the system utilizes many small datasets, which is closer to how scientists propose hypotheses — they look at multiple related datasets and come up with models to explain phenomena across those datasets.

      “One of the motivations of this work was our desire to study systems that learn models of datasets that is represented in a way that humans can understand. Instead of learning weights, can the model learn expressions or rules? And we wanted to see if we could build this system so it would learn on a whole battery of interrelated datasets, to make the system learn a little bit about how to better model each one,” says Kevin Ellis ’14, PhD ’20, an assistant professor of computer science at Cornell University and lead author of the paper.

      Joining Ellis on the paper are MIT faculty members Adam Albright, a professor of linguistics; Armando Solar-Lezama, a professor and associate director of the Computer Science and Artificial Intelligence Laboratory (CSAIL); and Joshua B. Tenenbaum, the Paul E. Newton Career Development Professor of Cognitive Science and Computation in the Department of Brain and Cognitive Sciences and a member of CSAIL; as well as senior author

      Timothy J. O’Donnell, assistant professor in the Department of Linguistics at McGill University, and Canada CIFAR AI Chair at the Mila – Quebec Artificial Intelligence Institute.

      The research is published today in Nature Communications.

      Looking at language 

      In their quest to develop an AI system that could automatically learn a model from multiple related datasets, the researchers chose to explore the interaction of phonology (the study of sound patterns) and morphology (the study of word structure).

      Data from linguistics textbooks offered an ideal testbed because many languages share core features, and textbook problems showcase specific linguistic phenomena. Textbook problems can also be solved by college students in a fairly straightforward way, but those students typically have prior knowledge about phonology from past lessons they use to reason about new problems.

      Ellis, who earned his PhD at MIT and was jointly advised by Tenenbaum and Solar-Lezama, first learned about morphology and phonology in an MIT class co-taught by O’Donnell, who was a postdoc at the time, and Albright.

      “Linguists have thought that in order to really understand the rules of a human language, to empathize with what it is that makes the system tick, you have to be human. We wanted to see if we can emulate the kinds of knowledge and reasoning that humans (linguists) bring to the task,” says Albright.

      To build a model that could learn a set of rules for assembling words, which is called a grammar, the researchers used a machine-learning technique known as Bayesian Program Learning. With this technique, the model solves a problem by writing a computer program.

      In this case, the program is the grammar the model thinks is the most likely explanation of the words and meanings in a linguistics problem. They built the model using Sketch, a popular program synthesizer which was developed at MIT by Solar-Lezama.

      But Sketch can take a lot of time to reason about the most likely program. To get around this, the researchers had the model work one piece at a time, writing a small program to explain some data, then writing a larger program that modifies that small program to cover more data, and so on.

      They also designed the model so it learns what “good” programs tend to look like. For instance, it might learn some general rules on simple Russian problems that it would apply to a more complex problem in Polish because the languages are similar. This makes it easier for the model to solve the Polish problem.

      Tackling textbook problems

      When they tested the model using 70 textbook problems, it was able to find a grammar that matched the entire set of words in the problem in 60 percent of cases, and correctly matched most of the word-form changes in 79 percent of problems.

      The researchers also tried pre-programming the model with some knowledge it “should” have learned if it was taking a linguistics course, and showed that it could solve all problems better.

      “One challenge of this work was figuring out whether what the model was doing was reasonable. This isn’t a situation where there is one number that is the single right answer. There is a range of possible solutions which you might accept as right, close to right, etc.,” Albright says.

      The model often came up with unexpected solutions. In one instance, it discovered the expected answer to a Polish language problem, but also another correct answer that exploited a mistake in the textbook. This shows that the model could “debug” linguistics analyses, Ellis says.

      The researchers also conducted tests that showed the model was able to learn some general templates of phonological rules that could be applied across all problems.

      “One of the things that was most surprising is that we could learn across languages, but it didn’t seem to make a huge difference,” says Ellis. “That suggests two things. Maybe we need better methods for learning across problems. And maybe, if we can’t come up with those methods, this work can help us probe different ideas we have about what knowledge to share across problems.”

      In the future, the researchers want to use their model to find unexpected solutions to problems in other domains. They could also apply the technique to more situations where higher-level knowledge can be applied across interrelated datasets. For instance, perhaps they could develop a system to infer differential equations from datasets on the motion of different objects, says Ellis.

      “This work shows that we have some methods which can, to some extent, learn inductive biases. But I don’t think we’ve quite figured out, even for these textbook problems, the inductive bias that lets a linguist accept the plausible grammars and reject the ridiculous ones,” he adds.

      “This work opens up many exciting venues for future research. I am particularly intrigued by the possibility that the approach explored by Ellis and colleagues (Bayesian Program Learning, BPL) might speak to how infants acquire language,” says T. Florian Jaeger, a professor of brain and cognitive sciences and computer science at the University of Rochester, who was not an author of this paper. “Future work might ask, for example, under what additional induction biases (assumptions about universal grammar) the BPL approach can successfully achieve human-like learning behavior on the type of data infants observe during language acquisition. I think it would be fascinating to see whether inductive biases that are even more abstract than those considered by Ellis and his team — such as biases originating in the limits of human information processing (e.g., memory constraints on dependency length or capacity limits in the amount of information that can be processed per time) — would be sufficient to induce some of the patterns observed in human languages.”

      This work was funded, in part, by the Air Force Office of Scientific Research, the Center for Brains, Minds, and Machines, the MIT-IBM Watson AI Lab, the Natural Science and Engineering Research Council of Canada, the Fonds de Recherche du Québec – Société et Culture, the Canada CIFAR AI Chairs Program, the National Science Foundation (NSF), and an NSF graduate fellowship. More

    • 100 Shares149 Views

      in Data Management & Statistics

      Taking a magnifying glass to data center operations

      by

      When the MIT Lincoln Laboratory Supercomputing Center (LLSC) unveiled its TX-GAIA supercomputer in 2019, it provided the MIT community a powerful new resource for applying artificial intelligence to their research. Anyone at MIT can submit a job to the system, which churns through trillions of operations per second to train models for diverse applications, such as spotting tumors in medical images, discovering new drugs, or modeling climate effects. But with this great power comes the great responsibility of managing and operating it in a sustainable manner — and the team is looking for ways to improve.

      “We have these powerful computational tools that let researchers build intricate models to solve problems, but they can essentially be used as black boxes. What gets lost in there is whether we are actually using the hardware as effectively as we can,” says Siddharth Samsi, a research scientist in the LLSC. 

      To gain insight into this challenge, the LLSC has been collecting detailed data on TX-GAIA usage over the past year. More than a million user jobs later, the team has released the dataset open source to the computing community.

      Their goal is to empower computer scientists and data center operators to better understand avenues for data center optimization — an important task as processing needs continue to grow. They also see potential for leveraging AI in the data center itself, by using the data to develop models for predicting failure points, optimizing job scheduling, and improving energy efficiency. While cloud providers are actively working on optimizing their data centers, they do not often make their data or models available for the broader high-performance computing (HPC) community to leverage. The release of this dataset and associated code seeks to fill this space.

      “Data centers are changing. We have an explosion of hardware platforms, the types of workloads are evolving, and the types of people who are using data centers is changing,” says Vijay Gadepally, a senior researcher at the LLSC. “Until now, there hasn’t been a great way to analyze the impact to data centers. We see this research and dataset as a big step toward coming up with a principled approach to understanding how these variables interact with each other and then applying AI for insights and improvements.”

      Papers describing the dataset and potential applications have been accepted to a number of venues, including the IEEE International Symposium on High-Performance Computer Architecture, the IEEE International Parallel and Distributed Processing Symposium, the Annual Conference of the North American Chapter of the Association for Computational Linguistics, the IEEE High-Performance and Embedded Computing Conference, and International Conference for High Performance Computing, Networking, Storage and Analysis. 

      Workload classification

      Among the world’s TOP500 supercomputers, TX-GAIA combines traditional computing hardware (central processing units, or CPUs) with nearly 900 graphics processing unit (GPU) accelerators. These NVIDIA GPUs are specialized for deep learning, the class of AI that has given rise to speech recognition and computer vision.

      The dataset covers CPU, GPU, and memory usage by job; scheduling logs; and physical monitoring data. Compared to similar datasets, such as those from Google and Microsoft, the LLSC dataset offers “labeled data, a variety of known AI workloads, and more detailed time series data compared with prior datasets. To our knowledge, it’s one of the most comprehensive and fine-grained datasets available,” Gadepally says. 

      Notably, the team collected time-series data at an unprecedented level of detail: 100-millisecond intervals on every GPU and 10-second intervals on every CPU, as the machines processed more than 3,000 known deep-learning jobs. One of the first goals is to use this labeled dataset to characterize the workloads that different types of deep-learning jobs place on the system. This process would extract features that reveal differences in how the hardware processes natural language models versus image classification or materials design models, for example.   

      The team has now launched the MIT Datacenter Challenge to mobilize this research. The challenge invites researchers to use AI techniques to identify with 95 percent accuracy the type of job that was run, using their labeled time-series data as ground truth.

      Such insights could enable data centers to better match a user’s job request with the hardware best suited for it, potentially conserving energy and improving system performance. Classifying workloads could also allow operators to quickly notice discrepancies resulting from hardware failures, inefficient data access patterns, or unauthorized usage.

      Too many choices

      Today, the LLSC offers tools that let users submit their job and select the processors they want to use, “but it’s a lot of guesswork on the part of users,” Samsi says. “Somebody might want to use the latest GPU, but maybe their computation doesn’t actually need it and they could get just as impressive results on CPUs, or lower-powered machines.”

      Professor Devesh Tiwari at Northeastern University is working with the LLSC team to develop techniques that can help users match their workloads to appropriate hardware. Tiwari explains that the emergence of different types of AI accelerators, GPUs, and CPUs has left users suffering from too many choices. Without the right tools to take advantage of this heterogeneity, they are missing out on the benefits: better performance, lower costs, and greater productivity.

      “We are fixing this very capability gap — making users more productive and helping users do science better and faster without worrying about managing heterogeneous hardware,” says Tiwari. “My PhD student, Baolin Li, is building new capabilities and tools to help HPC users leverage heterogeneity near-optimally without user intervention, using techniques grounded in Bayesian optimization and other learning-based optimization methods. But, this is just the beginning. We are looking into ways to introduce heterogeneity in our data centers in a principled approach to help our users achieve the maximum advantage of heterogeneity autonomously and cost-effectively.”

      Workload classification is the first of many problems to be posed through the Datacenter Challenge. Others include developing AI techniques to predict job failures, conserve energy, or create job scheduling approaches that improve data center cooling efficiencies.

      Energy conservation 

      To mobilize research into greener computing, the team is also planning to release an environmental dataset of TX-GAIA operations, containing rack temperature, power consumption, and other relevant data.

      According to the researchers, huge opportunities exist to improve the power efficiency of HPC systems being used for AI processing. As one example, recent work in the LLSC determined that simple hardware tuning, such as limiting the amount of power an individual GPU can draw, could reduce the energy cost of training an AI model by 20 percent, with only modest increases in computing time. “This reduction translates to approximately an entire week’s worth of household energy for a mere three-hour time increase,” Gadepally says.

      They have also been developing techniques to predict model accuracy, so that users can quickly terminate experiments that are unlikely to yield meaningful results, saving energy. The Datacenter Challenge will share relevant data to enable researchers to explore other opportunities to conserve energy.

      The team expects that lessons learned from this research can be applied to the thousands of data centers operated by the U.S. Department of Defense. The U.S. Air Force is a sponsor of this work, which is being conducted under the USAF-MIT AI Accelerator.

      Other collaborators include researchers at MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). Professor Charles Leiserson’s Supertech Research Group is investigating performance-enhancing techniques for parallel computing, and research scientist Neil Thompson is designing studies on ways to nudge data center users toward climate-friendly behavior.

      Samsi presented this work at the inaugural AI for Datacenter Optimization (ADOPT’22) workshop last spring as part of the IEEE International Parallel and Distributed Processing Symposium. The workshop officially introduced their Datacenter Challenge to the HPC community.

      “We hope this research will allow us and others who run supercomputing centers to be more responsive to user needs while also reducing the energy consumption at the center level,” Samsi says. More

    • 100 Shares99 Views

      in Data Management & Statistics

      Researchers discover major roadblock in alleviating network congestion

      by

      When users want to send data over the internet faster than the network can handle, congestion can occur — the same way traffic congestion snarls the morning commute into a big city.

      Computers and devices that transmit data over the internet break the data down into smaller packets and use a special algorithm to decide how fast to send those packets. These congestion control algorithms seek to fully discover and utilize available network capacity while sharing it fairly with other users who may be sharing the same network. These algorithms try to minimize delay caused by data waiting in queues in the network.

      Over the past decade, researchers in industry and academia have developed several algorithms that attempt to achieve high rates while controlling delays. Some of these, such as the BBR algorithm developed by Google, are now widely used by many websites and applications.

      But a team of MIT researchers has discovered that these algorithms can be deeply unfair. In a new study, they show there will always be a network scenario where at least one sender receives almost no bandwidth compared to other senders; that is, a problem known as “starvation” cannot be avoided.

      “What is really surprising about this paper and the results is that when you take into account the real-world complexity of network paths and all the things they can do to data packets, it is basically impossible for delay-controlling congestion control algorithms to avoid starvation using current methods,” says Mohammad Alizadeh, associate professor of electrical engineering and computer science (EECS).

      While Alizadeh and his co-authors weren’t able to find a traditional congestion control algorithm that could avoid starvation, there may be algorithms in a different class that could prevent this problem. Their analysis also suggests that changing how these algorithms work, so that they allow for larger variations in delay, could help prevent starvation in some network situations.

      Alizadeh wrote the paper with first author and EECS graduate student Venkat Arun and senior author Hari Balakrishnan, the Fujitsu Professor of Computer Science and Artificial Intelligence. The research will be presented at the ACM Special Interest Group on Data Communications (SIGCOMM) conference.

      Controlling congestion

      Congestion control is a fundamental problem in networking that researchers have been trying to tackle since the 1980s.

      A user’s computer does not know how fast to send data packets over the network because it lacks information, such as the quality of the network connection or how many other senders are using the network. Sending packets too slowly makes poor use of the available bandwidth. But sending them too quickly can overwhelm the network, and in doing so, packets will start to get dropped. These packets must be resent, which leads to longer delays. Delays can also be caused by packets waiting in queues for a long time.

      Congestion control algorithms use packet losses and delays as signals to infer congestion and decide how fast to send data. But the internet is complicated, and packets can be delayed and lost for reasons unrelated to network congestion. For instance, data could be held up in a queue along the way and then released with a burst of other packets, or the receiver’s acknowledgement might be delayed. The authors call delays that are not caused by congestion “jitter.”

      Even if a congestion control algorithm measures delay perfectly, it can’t tell the difference between delay caused by congestion and delay caused by jitter. Delay caused by jitter is unpredictable and confuses the sender. Because of this ambiguity, users start estimating delay differently, which causes them to send packets at unequal rates. Eventually, this leads to a situation where starvation occurs and someone gets shut out completely, Arun explains.

      “We started the project because we lacked a theoretical understanding of congestion control behavior in the presence of jitter. To place it on a firmer theoretical footing, we built a mathematical model that was simple enough to think about, yet able to capture some of the complexities of the internet. It has been very rewarding to have math tell us things we didn’t know and that have practical relevance,” he says.

      Studying starvation

      The researchers fed their mathematical model to a computer, gave it a series of commonly used congestion control algorithms, and asked the computer to find an algorithm that could avoid starvation, using their model.

      “We couldn’t do it. We tried every algorithm that we are aware of, and some new ones we made up. Nothing worked. The computer always found a situation where some people get all the bandwidth and at least one person gets basically nothing,” Arun says.

      The researchers were surprised by this result, especially since these algorithms are widely believed to be reasonably fair. They started suspecting that it may not be possible to avoid starvation, an extreme form of unfairness. This motivated them to define a class of algorithms they call “delay-convergent algorithms” that they proved will always suffer from starvation under their network model. All existing congestion control algorithms that control delay (that the researchers are aware of) are delay-convergent.

      The fact that such simple failure modes of these widely used algorithms remained unknown for so long illustrates how difficult it is to understand algorithms through empirical testing alone, Arun adds. It underscores the importance of a solid theoretical foundation.

      But all hope is not lost. While all the algorithms they tested failed, there may be other algorithms which are not delay-convergent that might be able to avoid starvation This suggests that one way to fix the problem might be to design congestion control algorithms that vary the delay range more widely, so the range is larger than any delay that might occur due to jitter in the network.

      “To control delays, algorithms have tried to also bound the variations in delay about a desired equilibrium, but there is nothing wrong in potentially creating greater delay variation to get better measurements of congestive delays. It is just a new design philosophy you would have to adopt,” Balakrishnan adds.

      Now, the researchers want to keep pushing to see if they can find or build an algorithm that will eliminate starvation. They also want to apply this approach of mathematical modeling and computational proofs to other thorny, unsolved problems in networked systems.

      “We are increasingly reliant on computer systems for very critical things, and we need to put their reliability on a firmer conceptual footing. We’ve shown the surprising things you can discover when you put in the time to come up with these formal specifications of what the problem actually is,” says Alizadeh.

      The NASA University Leadership Initiative (grant #80NSSC20M0163) provided funds to assist the authors with their research, but the research paper solely reflects the opinions and conclusions of its authors and not any NASA entity. This work was also partially funded by the National Science Foundation, award number 1751009. More