More stories

  • in

    Nonsense can make sense to machine-learning models

    For all that neural networks can accomplish, we still don’t really understand how they operate. Sure, we can program them to learn, but making sense of a machine’s decision-making process remains much like a fancy puzzle with a dizzying, complex pattern where plenty of integral pieces have yet to be fitted. 

    If a model was trying to classify an image of said puzzle, for example, it could encounter well-known, but annoying adversarial attacks, or even more run-of-the-mill data or processing issues. But a new, more subtle type of failure recently identified by MIT scientists is another cause for concern: “overinterpretation,” where algorithms make confident predictions based on details that don’t make sense to humans, like random patterns or image borders. 

    This could be particularly worrisome for high-stakes environments, like split-second decisions for self-driving cars, and medical diagnostics for diseases that need more immediate attention. Autonomous vehicles in particular rely heavily on systems that can accurately understand surroundings and then make quick, safe decisions. The network used specific backgrounds, edges, or particular patterns of the sky to classify traffic lights and street signs — irrespective of what else was in the image. 

    The team found that neural networks trained on popular datasets like CIFAR-10 and ImageNet suffered from overinterpretation. Models trained on CIFAR-10, for example, made confident predictions even when 95 percent of input images were missing, and the remainder is senseless to humans. 

    “Overinterpretation is a dataset problem that’s caused by these nonsensical signals in datasets. Not only are these high-confidence images unrecognizable, but they contain less than 10 percent of the original image in unimportant areas, such as borders. We found that these images were meaningless to humans, yet models can still classify them with high confidence,” says Brandon Carter, MIT Computer Science and Artificial Intelligence Laboratory PhD student and lead author on a paper about the research. 

    Deep-image classifiers are widely used. In addition to medical diagnosis and boosting autonomous vehicle technology, there are use cases in security, gaming, and even an app that tells you if something is or isn’t a hot dog, because sometimes we need reassurance. The tech in discussion works by processing individual pixels from tons of pre-labeled images for the network to “learn.” 

    Image classification is hard, because machine-learning models have the ability to latch onto these nonsensical subtle signals. Then, when image classifiers are trained on datasets such as ImageNet, they can make seemingly reliable predictions based on those signals. 

    Although these nonsensical signals can lead to model fragility in the real world, the signals are actually valid in the datasets, meaning overinterpretation can’t be diagnosed using typical evaluation methods based on that accuracy. 

    To find the rationale for the model’s prediction on a particular input, the methods in the present study start with the full image and repeatedly ask, what can I remove from this image? Essentially, it keeps covering up the image, until you’re left with the smallest piece that still makes a confident decision. 

    To that end, it could also be possible to use these methods as a type of validation criteria. For example, if you have an autonomously driving car that uses a trained machine-learning method for recognizing stop signs, you could test that method by identifying the smallest input subset that constitutes a stop sign. If that consists of a tree branch, a particular time of day, or something that’s not a stop sign, you could be concerned that the car might come to a stop at a place it’s not supposed to.

    While it may seem that the model is the likely culprit here, the datasets are more likely to blame. “There’s the question of how we can modify the datasets in a way that would enable models to be trained to more closely mimic how a human would think about classifying images and therefore, hopefully, generalize better in these real-world scenarios, like autonomous driving and medical diagnosis, so that the models don’t have this nonsensical behavior,” says Carter. 

    This may mean creating datasets in more controlled environments. Currently, it’s just pictures that are extracted from public domains that are then classified. But if you want to do object identification, for example, it might be necessary to train models with objects with an uninformative background. 

    This work was supported by Schmidt Futures and the National Institutes of Health. Carter wrote the paper alongside Siddhartha Jain and Jonas Mueller, scientists at Amazon, and MIT Professor David Gifford. They are presenting the work at the 2021 Conference on Neural Information Processing Systems. More

  • in

    Differences in T cells’ functional states determine resistance to cancer therapy

    Non-small cell lung cancer (NSCLC) is the most common type of lung cancer in humans. Some patients with NSCLC receive a therapy called immune checkpoint blockade (ICB) that helps kill cancer cells by reinvigorating a subset of immune cells called T cells, which are “exhausted” and have stopped working. However, only about 35 percent of NSCLC patients respond to ICB therapy. Stefani Spranger’s lab at the MIT Department of Biology explores the mechanisms behind this resistance, with the goal of inspiring new therapies to better treat NSCLC patients. In a new study published on Oct. 29 in Science Immunology, a team led by Spranger lab postdoc Brendan Horton revealed what causes T cells to be non-responsive to ICB — and suggests a possible solution.

    Scientists have long thought that the conditions within a tumor were responsible for determining when T cells stop working and become exhausted after being overstimulated or working for too long to fight a tumor. That’s why physicians prescribe ICB to treat cancer — ICB can invigorate the exhausted T cells within a tumor. However, Horton’s new experiments show that some ICB-resistant T cells stop working before they even enter the tumor. These T cells are not actually exhausted, but rather they become dysfunctional due to changes in gene expression that arise early during the activation of a T cell, which occurs in lymph nodes. Once activated, T cells differentiate into certain functional states, which are distinguishable by their unique gene expression patterns.

    The notion that the dysfunctional state that leads to ICB resistance arises before T cells enter the tumor is quite novel, says Spranger, the Howard S. and Linda B. Stern Career Development Professor, a member of the Koch Institute for Integrative Cancer Research, and the study’s senior author.

    “We show that this state is actually a preset condition, and that the T cells are already non-responsive to therapy before they enter the tumor,” she says. As a result, she explains, ICB therapies that work by reinvigorating exhausted T cells within the tumor are less likely to be effective. This suggests that combining ICB with other forms of immunotherapy that target T cells differently might be a more effective approach to help the immune system combat this subset of lung cancer.

    In order to determine why some tumors are resistant to ICB, Horton and the research team studied T cells in murine models of NSCLC. The researchers sequenced messenger RNA from the responsive and non-responsive T cells in order to identify any differences between the T cells. Supported in part by the Koch Institute Frontier Research Program, they used a technique called Seq-Well, developed in the lab of fellow Koch Institute member J. Christopher Love, the Raymond A. (1921) and Helen E. St. Laurent Professor of Chemical Engineering and a co-author of the study. The technique allows for the rapid gene expression profiling of single cells, which permitted Spranger and Horton to get a very granular look at the gene expression patterns of the T cells they were studying.

    Seq-Well revealed distinct patterns of gene expression between the responsive and non-responsive T cells. These differences, which are determined when the T cells assume their specialized functional states, may be the underlying cause of ICB resistance.

    Now that Horton and his colleagues had a possible explanation for why some T cells did not respond to ICB, they decided to see if they could help the ICB-resistant T cells kill the tumor cells. When analyzing the gene expression patterns of the non-responsive T cells, the researchers had noticed that these T cells had a lower expression of receptors for certain cytokines, small proteins that control immune system activity. To counteract this, the researchers treated lung tumors in murine models with extra cytokines. As a result, the previously non-responsive T cells were then able to fight the tumors — meaning that the cytokine therapy prevented, and potentially even reversed, the dysfunctionality.

    Administering cytokine therapy to human patients is not currently safe, because cytokines can cause serious side effects as well as a reaction called a “cytokine storm,” which can produce severe fevers, inflammation, fatigue, and nausea. However, there are ongoing efforts to figure out how to safely administer cytokines to specific tumors. In the future, Spranger and Horton suspect that cytokine therapy could be used in combination with ICB.

    “This is potentially something that could be translated into a therapeutic that could increase the therapy response rate in non-small cell lung cancer,” Horton says.

    Spranger agrees that this work will help researchers develop more innovative cancer therapies, especially because researchers have historically focused on T cell exhaustion rather than the earlier role that T cell functional states might play in cancer.

    “If T cells are rendered dysfunctional early on, ICB is not going to be effective, and we need to think outside the box,” she says. “There’s more evidence, and other labs are now showing this as well, that the functional state of the T cell actually matters quite substantially in cancer therapies.” To Spranger, this means that cytokine therapy “might be a therapeutic avenue” for NSCLC patients beyond ICB.

    Jeffrey Bluestone, the A.W. and Mary Margaret Clausen Distinguished Professor of Metabolism and Endocrinology at the University of California-San Francisco, who was not involved with the paper, agrees. “The study provides a potential opportunity to ‘rescue’ immunity in the NSCLC non-responder patients with appropriate combination therapies,” he says.

    This research was funded by the Pew-Stewart Scholars for Cancer Research, the Ludwig Center for Molecular Oncology, the Koch Institute Frontier Research Program through the Kathy and Curt Mable Cancer Research Fund, and the National Cancer Institute. More

  • in

    Exploring the human stories behind the data

    Shaking in the back of a police cruiser, handcuffs digging into his wrists, Brian Williams was overwhelmed with fear. He had been pulled over, but before he was asked for his name, license, or registration, a police officer ordered him out of his car and into back of the police cruiser, saying into his radio, “Black male detained.” The officer’s explanation for these actions was: “for your safety and mine.”

    Williams walked away from the experience with two tickets, a pair of bruised wrists, and a desire to do everything in his power to prevent others from experiencing the utter powerlessness he had felt.

    Now an MIT senior majoring in biological engineering and minoring in Black studies, Williams has continued working to empower his community. Through experiences in and out of the classroom, he has leveraged his background in bioengineering to explore interests in public health and social justice, specifically looking at how the medical sector can uplift and support communities of color.

    Williams grew up in a close-knit family and community in Broward County, Florida, where he found comfort in the routine of Sunday church services, playing outside with friends, and cookouts on the weekends. Broward County was home to him — a home he felt deeply invested in and indebted to.

    “It takes a village. The Black community has invested a lot in me, and I have a lot to invest back in it,” he says.

    Williams initially focused on STEM subjects at MIT, but in his sophomore year, his interests in exploring data science and humanities research led him to an Undergraduate Research Opportunities Program (UROP) project in the Department of Political Science. Working with Professor Ariel White, he analyzed information on incarceration and voting rights, studied the behavior patterns of police officers, and screened 911 calls to identify correlations between how people described events to how the police responded to them.

    In the summer before his junior year, Williams also joined MIT’s Civic Data Design Lab, where he worked as a researcher for the Missing Data Project, which uses both journalism and data science to visualize statistics and humanize the people behind the numbers. As the project’s name suggests, there is often much to be learned from seeking out data that aren’t easily available. Using datasets and interviews describing how the pandemic affected Black communities, Williams and a team of researchers created a series called the Color of Covid, which told the stories behind the grim statistics on race and the pandemic.

    The following year, Williams undertook a research-and-development internship with the biopharmaceutical company Amgen in San Francisco, working on protein engineering to combat autoimmune diseases. Because this work was primarily in the lab, focusing on science-based applications, he saw it as an opportunity to ask himself: “Do I want to dedicate my life to this area of bioengineering?” He found the issue of social justice to be more compelling.

    At the same time, Williams was drawn toward tackling problems the local Black community was experiencing related to the pandemic. He found himself thinking deeply about how to educate the public, address disparities in case rates, and, above all, help people.

    Working through Amgen’s Black Employee Resource Group and its Diversity, Inclusion, and Belonging Team, Williams crafted a proposal, which the company adopted, for addressing Covid-19 vaccination misinformation in Black and Brown communities in San Mateo and San Francisco County. He paid special attention to how to frame vaccine hesitancy among members of these communities, understanding that a longstanding history of racism in scientific discovery and medicine led many Black and Brown people to distrust the entire medical industry.

    “Trying to meet people where they are is important,” Williams says.

    This experience reinforced the idea for Williams that he wanted to do everything in his power to uplift the Black community.

    “I think it’s only right that I go out and I shine bright because it’s not easy being Black. You know, you have to work twice as hard to get half as much,” he says.

    As the current political action co-chair of the MIT Black Students’ Union (BSU), Williams also works to inspire change on campus, promoting and participating in events that uplift the BSU. During his Amgen internship, he also organized the MIT Black History Month Takeover Series, which involved organizing eight events from February through the beginning of spring semester. These included promotions through social media and virtual meetings for students and faculty. For his leadership, he received the “We Are Family” award from the BSU executive board.

    Williams prioritizes community in everything he does, whether in the classroom, at a campus event, or spending time outside in local communities of color around Boston.

    “The things that really keep me going are the stories of other people,” says Williams, who is currently applying to a variety of postgraduate programs. After receiving MIT endorsement, he applied to the Rhodes and Marshall Fellowships; he also plans to apply to law school with a joint master’s degree in public health and policy.

    Ultimately, Williams hopes to bring his fight for racial justice to the policy level, looking at how a long, ongoing history of medical racism has led marginalized communities to mistrust current scientific endeavors. He wants to help bring about new legislation to fix old systems which disproportionately harm communities of color. He says he aims to be “an engineer of social solutions, one who reaches deep into their toolbox of social justice, pulling the levers of activism, advocacy, democracy, and legislation to radically change our world — to improve our social institutions at the root and liberate our communities.” While he understands this is a big feat, he sees his ambition as an asset.

    “I’m just another person with huge aspirations, and an understanding that you have to go get it if you want it,” he says. “You feel me? At the end of the day, this is just the beginning of my story. And I’m grateful to everyone in my life that’s helping me write it. Tap in.” More

  • in

    Enabling AI-driven health advances without sacrificing patient privacy

    There’s a lot of excitement at the intersection of artificial intelligence and health care. AI has already been used to improve disease treatment and detection, discover promising new drugs, identify links between genes and diseases, and more.

    By analyzing large datasets and finding patterns, virtually any new algorithm has the potential to help patients — AI researchers just need access to the right data to train and test those algorithms. Hospitals, understandably, are hesitant to share sensitive patient information with research teams. When they do share data, it’s difficult to verify that researchers are only using the data they need and deleting it after they’re done.

    Secure AI Labs (SAIL) is addressing those problems with a technology that lets AI algorithms run on encrypted datasets that never leave the data owner’s system. Health care organizations can control how their datasets are used, while researchers can protect the confidentiality of their models and search queries. Neither party needs to see the data or the model to collaborate.

    SAIL’s platform can also combine data from multiple sources, creating rich insights that fuel more effective algorithms.

    “You shouldn’t have to schmooze with hospital executives for five years before you can run your machine learning algorithm,” says SAIL co-founder and MIT Professor Manolis Kellis, who co-founded the company with CEO Anne Kim ’16, SM ’17. “Our goal is to help patients, to help machine learning scientists, and to create new therapeutics. We want new algorithms — the best algorithms — to be applied to the biggest possible data set.”

    SAIL has already partnered with hospitals and life science companies to unlock anonymized data for researchers. In the next year, the company hopes to be working with about half of the top 50 academic medical centers in the country.

    Unleashing AI’s full potential

    As an undergraduate at MIT studying computer science and molecular biology, Kim worked with researchers in the Computer Science and Artificial Intelligence Laboratory (CSAIL) to analyze data from clinical trials, gene association studies, hospital intensive care units, and more.

    “I realized there is something severely broken in data sharing, whether it was hospitals using hard drives, ancient file transfer protocol, or even sending stuff in the mail,” Kim says. “It was all just not well-tracked.”

    Kellis, who is also a member of the Broad Institute of MIT and Harvard, has spent years establishing partnerships with hospitals and consortia across a range of diseases including cancers, heart disease, schizophrenia, and obesity. He knew that smaller research teams would struggle to get access to the same data his lab was working with.

    In 2017, Kellis and Kim decided to commercialize technology they were developing to allow AI algorithms to run on encrypted data.

    In the summer of 2018, Kim participated in the delta v startup accelerator run by the Martin Trust Center for MIT Entrepreneurship. The founders also received support from the Sandbox Innovation Fund and the Venture Mentoring Service, and made various early connections through their MIT network.

    To participate in SAIL’s program, hospitals and other health care organizations make parts of their data available to researchers by setting up a node behind their firewall. SAIL then sends encrypted algorithms to the servers where the datasets reside in a process called federated learning. The algorithms crunch the data locally in each server and transmit the results back to a central model, which updates itself. No one — not the researchers, the data owners, or even SAIL —has access to the models or the datasets.

    The approach allows a much broader set of researchers to apply their models to large datasets. To further engage the research community, Kellis’ lab at MIT has begun holding competitions in which it gives access to datasets in areas like protein function and gene expression, and challenges researchers to predict results.

    “We invite machine learning researchers to come and train on last year’s data and predict this year’s data,” says Kellis. “If we see there’s a new type of algorithm that is performing best in these community-level assessments, people can adopt it locally at many different institutions and level the playing field. So, the only thing that matters is the quality of your algorithm rather than the power of your connections.”

    By enabling a large number of datasets to be anonymized into aggregate insights, SAIL’s technology also allows researchers to study rare diseases, in which small pools of relevant patient data are often spread out among many institutions. That has historically made the data difficult to apply AI models to.

    “We’re hoping that all of these datasets will eventually be open,” Kellis says. “We can cut across all the silos and enable a new era where every patient with every rare disorder across the entire world can come together in a single keystroke to analyze data.”

    Enabling the medicine of the future

    To work with large amounts of data around specific diseases, SAIL has increasingly sought to partner with patient associations and consortia of health care groups, including an international health care consulting company and the Kidney Cancer Association. The partnerships also align SAIL with patients, the group they’re most trying to help.

    Overall, the founders are happy to see SAIL solving problems they faced in their labs for researchers around the world.

    “The right place to solve this is not an academic project. The right place to solve this is in industry, where we can provide a platform not just for my lab but for any researcher,” Kellis says. “It’s about creating an ecosystem of academia, researchers, pharma, biotech, and hospital partners. I think it’s the blending all of these different areas that will make that vision of medicine of the future become a reality.” More