Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 63 Shares129 Views

      in Data Management & Statistics

      Fighting discrimination in mortgage lending

      by

      Although the U.S. Equal Credit Opportunity Act prohibits discrimination in mortgage lending, biases still impact many borrowers. One 2021 Journal of Financial Economics study found that borrowers from minority groups were charged interest rates that were nearly 8 percent higher and were rejected for loans 14 percent more often than those from privileged groups.

      When these biases bleed into machine-learning models that lenders use to streamline decision-making, they can have far-reaching consequences for housing fairness and even contribute to widening the racial wealth gap.

      If a model is trained on an unfair dataset, such as one in which a higher proportion of Black borrowers were denied loans versus white borrowers with the same income, credit score, etc., those biases will affect the model’s predictions when it is applied to real situations. To stem the spread of mortgage lending discrimination, MIT researchers created a process that removes bias in data that are used to train these machine-learning models.

      While other methods try to tackle this bias, the researchers’ technique is new in the mortgage lending domain because it can remove bias from a dataset that has multiple sensitive attributes, such as race and ethnicity, as well as several “sensitive” options for each attribute, such as Black or white, and Hispanic or Latino or non-Hispanic or Latino. Sensitive attributes and options are features that distinguish a privileged group from an underprivileged group.

      The researchers used their technique, which they call DualFair, to train a machine-learning classifier that makes fair predictions of whether borrowers will receive a mortgage loan. When they applied it to mortgage lending data from several U.S. states, their method significantly reduced the discrimination in the predictions while maintaining high accuracy.

      “As Sikh Americans, we deal with bias on a frequent basis and we think it is unacceptable to see that transform to algorithms in real-world applications. For things like mortgage lending and financial systems, it is very important that bias not infiltrate these systems because it can emphasize the gaps that are already in place against certain groups,” says Jashandeep Singh, a senior at Floyd Buchanan High School and co-lead author of the paper with his twin brother, Arashdeep. The Singh brothers were recently accepted into MIT.

      Joining Arashdeep and Jashandeep Singh on the paper are MIT sophomore Ariba Khan and senior author Amar Gupta, a researcher in the Computer Science and Artificial Intelligence Laboratory at MIT, who studies the use of evolving technology to address inequity and other societal issues. The research was recently published online and will appear in a special issue of Machine Learning and Knowledge Extraction.

      Double take

      DualFair tackles two types of bias in a mortgage lending dataset — label bias and selection bias. Label bias occurs when the balance of favorable or unfavorable outcomes for a particular group is unfair. (Black applicants are denied loans more frequently than they should be.) Selection bias is created when data are not representative of the larger population. (The dataset only includes individuals from one neighborhood where incomes are historically low.)

      The DualFair process eliminates label bias by subdividing a dataset into the largest number of subgroups based on combinations of sensitive attributes and options, such as white men who are not Hispanic or Latino, Black women who are Hispanic or Latino, etc.

      By breaking down the dataset into as many subgroups as possible, DualFair can simultaneously address discrimination based on multiple attributes.

      “Researchers have mostly tried to classify biased cases as binary so far. There are multiple parameters to bias, and these multiple parameters have their own impact in different cases. They are not equally weighed. Our method is able to calibrate it much better,” says Gupta.

      After the subgroups have been generated, DualFair evens out the number of borrowers in each subgroup by duplicating individuals from minority groups and deleting individuals from the majority group. DualFair then balances the proportion of loan acceptances and rejections in each subgroup so they match the median in the original dataset before recombining the subgroups.

      DualFair then eliminates selection bias by iterating on each data point to see if discrimination is present. For instance, if an individual is a non-Hispanic or Latino Black woman who was rejected for a loan, the system will adjust her race, ethnicity, and gender one at a time to see if the outcome changes. If this borrower is granted a loan when her race is changed to white, DualFair considers that data point biased and removes it from the dataset.

      Fairness vs. accuracy

      To test DualFair, the researchers used the publicly available Home Mortgage Disclosure Act dataset, which spans 88 percent of all mortgage loans in the U.S. in 2019, and includes 21 features, including race, sex, and ethnicity. They used DualFair to “de-bias” the entire dataset and smaller datasets for six states, and then trained a machine-learning model to predict loan acceptances and rejections.

      After applying DualFair, the fairness of predictions increased while the accuracy level remained high across all states. They used an existing fairness metric known as average odds difference, but it can only measure fairness in one sensitive attribute at a time.

      So, they created their own fairness metric, called alternate world index, that considers bias from multiple sensitive attributes and options as a whole. Using this metric, they found that DualFair increased fairness in predictions for four of the six states while maintaining high accuracy.

      “It is the common belief that if you want to be accurate, you have to give up on fairness, or if you want to be fair, you have to give up on accuracy. We show that we can make strides toward lessening that gap,” Khan says.

      The researchers now want to apply their method to de-bias different types of datasets, such as those that capture health care outcomes, car insurance rates, or job applications. They also plan to address limitations of DualFair, including its instability when there are small amounts of data with multiple sensitive attributes and options.

      While this is only a first step, the researchers are hopeful their work can someday have an impact on mitigating bias in lending and beyond.

      “Technology, very bluntly, works only for a certain group of people. In the mortgage loan domain in particular, African American women have been historically discriminated against. We feel passionate about making sure that systemic racism does not extend to algorithmic models. There is no point in making an algorithm that can automate a process if it doesn’t work for everyone equally,” says Khan.

      This research is supported, in part, by the FinTech@CSAIL initiative. More

    • 88 Shares119 Views

      in Data Management & Statistics

      Security tool guarantees privacy in surveillance footage

      by

      Surveillance cameras have an identity problem, fueled by an inherent tension between utility and privacy. As these powerful little devices have cropped up seemingly everywhere, the use of machine learning tools has automated video content analysis at a massive scale — but with increasing mass surveillance, there are currently no legally enforceable rules to limit privacy invasions. 

      Security cameras can do a lot — they’ve become smarter and supremely more competent than their ghosts of grainy pictures past, the ofttimes “hero tool” in crime media. (“See that little blurry blue blob in the right hand corner of that densely populated corner — we got him!”) Now, video surveillance can help health officials measure the fraction of people wearing masks, enable transportation departments to monitor the density and flow of vehicles, bikes, and pedestrians, and provide businesses with a better understanding of shopping behaviors. But why has privacy remained a weak afterthought? 

      The status quo is to retrofit video with blurred faces or black boxes. Not only does this prevent analysts from asking some genuine queries (e.g., Are people wearing masks?), it also doesn’t always work; the system may miss some faces and leave them unblurred for the world to see. Dissatisfied with this status quo, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), in collaboration with other institutions, came up with a system to better guarantee privacy in video footage from surveillance cameras. Called “Privid,” the system lets analysts submit video data queries, and adds a little bit of noise (extra data) to the end result to ensure that an individual can’t be identified. The system builds on a formal definition of privacy — “differential privacy” — which allows access to aggregate statistics about private data without revealing personally identifiable information.

      Typically, analysts would just have access to the entire video to do whatever they want with it, but Privid makes sure the video isn’t a free buffet. Honest analysts can get access to the information they need, but that access is restrictive enough that malicious analysts can’t do too much with it. To enable this, rather than running the code over the entire video in one shot, Privid breaks the video into small pieces and runs processing code over each chunk. Instead of getting results back from each piece, the segments are aggregated, and that additional noise is added. (There’s also information on the error bound you’re going to get on your result — maybe a 2 percent error margin, given the extra noisy data added). 

      For example, the code might output the number of people observed in each video chunk, and the aggregation might be the “sum,” to count the total number of people wearing face coverings, or the “average” to estimate the density of crowds. 

      Privid allows analysts to use their own deep neural networks that are commonplace for video analytics today. This gives analysts the flexibility to ask questions that the designers of Privid did not anticipate. Across a variety of videos and queries, Privid was accurate within 79 to 99 percent of a non-private system.

      “We’re at a stage right now where cameras are practically ubiquitous. If there’s a camera on every street corner, every place you go, and if someone could actually process all of those videos in aggregate, you can imagine that entity building a very precise timeline of when and where a person has gone,” says MIT CSAIL PhD student ​​Frank Cangialosi, the lead author on a paper about Privid. “People are already worried about location privacy with GPS — video data in aggregate could capture not only your location history, but also moods, behaviors, and more at each location.” 

      Privid introduces a new notion of “duration-based privacy,” which decouples the definition of privacy from its enforcement — with obfuscation, if your privacy goal is to protect all people, the enforcement mechanism needs to do some work to find the people to protect, which it may or may not do perfectly. With this mechanism, you don’t need to fully specify everything, and you’re not hiding more information than you need to. 

      Let’s say we have a video overlooking a street. Two analysts, Alice and Bob, both claim they want to count the number of people that pass by each hour, so they submit a video processing module and ask for a sum aggregation.

      The first analyst is the city planning department, which hopes to use this information to understand footfall patterns and plan sidewalks for the city. Their model counts people and outputs this count for each video chunk.

      The other analyst is malicious. They hope to identify every time “Charlie” passes by the camera. Their model only looks for Charlie’s face, and outputs a large number if Charlie is present (i.e., the “signal” they’re trying to extract), or zero otherwise. Their hope is that the sum will be non-zero if Charlie was present. 

      From Privid’s perspective, these two queries look identical. It’s hard to reliably determine what their models might be doing internally, or what the analyst hopes to use the data for. This is where the noise comes in. Privid executes both of the queries, and adds the same amount of noise for each. In the first case, because Alice was counting all people, this noise will only have a small impact on the result, but likely won’t impact the usefulness. 

      In the second case, since Bob was looking for a specific signal (Charlie was only visible for a few chunks), the noise is enough to prevent them from knowing if Charlie was there or not. If they see a non-zero result, it might be because Charlie was actually there, or because the model outputs “zero,” but the noise made it non-zero. Privid didn’t need to know anything about when or where Charlie appeared, the system just needed to know a rough upper bound on how long Charlie might appear for, which is easier to specify than figuring out the exact locations, which prior methods rely on. 

      The challenge is determining how much noise to add — Privid wants to add just enough to hide everyone, but not so much that it would be useless for analysts. Adding noise to the data and insisting on queries over time windows means that your result isn’t going to be as accurate as it could be, but the results are still useful while providing better privacy. 

      Cangialosi wrote the paper with Princeton PhD student Neil Agarwal, MIT CSAIL PhD student Venkat Arun, assistant professor at the University of Chicago Junchen Jiang, assistant professor at Rutgers University and former MIT CSAIL postdoc Srinivas Narayana, associate professor at Rutgers University Anand Sarwate, and assistant professor at Princeton University and Ravi Netravali SM ’15, PhD ’18. Cangialosi will present the paper at the USENIX Symposium on Networked Systems Design and Implementation Conference in April in Renton, Washington. 

      This work was partially supported by a Sloan Research Fellowship and National Science Foundation grants. More

    • 125 Shares149 Views

      in Data Management & Statistics

      When it comes to AI, can we ditch the datasets?

      by

      Huge amounts of data are needed to train machine-learning models to perform image classification tasks, such as identifying damage in satellite photos following a natural disaster. However, these data are not always easy to come by. Datasets may cost millions of dollars to generate, if usable data exist in the first place, and even the best datasets often contain biases that negatively impact a model’s performance.

      To circumvent some of the problems presented by datasets, MIT researchers developed a method for training a machine learning model that, rather than using a dataset, uses a special type of machine-learning model to generate extremely realistic synthetic data that can train another model for downstream vision tasks.

      Their results show that a contrastive representation learning model trained using only these synthetic data is able to learn visual representations that rival or even outperform those learned from real data.

      This special machine-learning model, known as a generative model, requires far less memory to store or share than a dataset. Using synthetic data also has the potential to sidestep some concerns around privacy and usage rights that limit how some real data can be distributed. A generative model could also be edited to remove certain attributes, like race or gender, which could address some biases that exist in traditional datasets.

      “We knew that this method should eventually work; we just needed to wait for these generative models to get better and better. But we were especially pleased when we showed that this method sometimes does even better than the real thing,” says Ali Jahanian, a research scientist in the Computer Science and Artificial Intelligence Laboratory (CSAIL) and lead author of the paper.

      Jahanian wrote the paper with CSAIL grad students Xavier Puig and Yonglong Tian, and senior author Phillip Isola, an assistant professor in the Department of Electrical Engineering and Computer Science. The research will be presented at the International Conference on Learning Representations.

      Generating synthetic data

      Once a generative model has been trained on real data, it can generate synthetic data that are so realistic they are nearly indistinguishable from the real thing. The training process involves showing the generative model millions of images that contain objects in a particular class (like cars or cats), and then it learns what a car or cat looks like so it can generate similar objects.

      Essentially by flipping a switch, researchers can use a pretrained generative model to output a steady stream of unique, realistic images that are based on those in the model’s training dataset, Jahanian says.

      But generative models are even more useful because they learn how to transform the underlying data on which they are trained, he says. If the model is trained on images of cars, it can “imagine” how a car would look in different situations — situations it did not see during training — and then output images that show the car in unique poses, colors, or sizes.

      Having multiple views of the same image is important for a technique called contrastive learning, where a machine-learning model is shown many unlabeled images to learn which pairs are similar or different.

      The researchers connected a pretrained generative model to a contrastive learning model in a way that allowed the two models to work together automatically. The contrastive learner could tell the generative model to produce different views of an object, and then learn to identify that object from multiple angles, Jahanian explains.

      “This was like connecting two building blocks. Because the generative model can give us different views of the same thing, it can help the contrastive method to learn better representations,” he says.

      Even better than the real thing

      The researchers compared their method to several other image classification models that were trained using real data and found that their method performed as well, and sometimes better, than the other models.

      One advantage of using a generative model is that it can, in theory, create an infinite number of samples. So, the researchers also studied how the number of samples influenced the model’s performance. They found that, in some instances, generating larger numbers of unique samples led to additional improvements.

      “The cool thing about these generative models is that someone else trained them for you. You can find them in online repositories, so everyone can use them. And you don’t need to intervene in the model to get good representations,” Jahanian says.

      But he cautions that there are some limitations to using generative models. In some cases, these models can reveal source data, which can pose privacy risks, and they could amplify biases in the datasets they are trained on if they aren’t properly audited.

      He and his collaborators plan to address those limitations in future work. Another area they want to explore is using this technique to generate corner cases that could improve machine learning models. Corner cases often can’t be learned from real data. For instance, if researchers are training a computer vision model for a self-driving car, real data wouldn’t contain examples of a dog and his owner running down a highway, so the model would never learn what to do in this situation. Generating that corner case data synthetically could improve the performance of machine learning models in some high-stakes situations.

      The researchers also want to continue improving generative models so they can compose images that are even more sophisticated, he says.

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

    • 63 Shares149 Views

      in Data Management & Statistics

      Injecting fairness into machine-learning models

      by

      If a machine-learning model is trained using an unbalanced dataset, such as one that contains far more images of people with lighter skin than people with darker skin, there is serious risk the model’s predictions will be unfair when it is deployed in the real world.

      But this is only one part of the problem. MIT researchers have found that machine-learning models that are popular for image recognition tasks actually encode bias when trained on unbalanced data. This bias within the model is impossible to fix later on, even with state-of-the-art fairness-boosting techniques, and even when retraining the model with a balanced dataset.      

      So, the researchers came up with a technique to introduce fairness directly into the model’s internal representation itself. This enables the model to produce fair outputs even if it is trained on unfair data, which is especially important because there are very few well-balanced datasets for machine learning.

      The solution they developed not only leads to models that make more balanced predictions, but also improves their performance on downstream tasks like facial recognition and animal species classification.

      “In machine learning, it is common to blame the data for bias in models. But we don’t always have balanced data. So, we need to come up with methods that actually fix the problem with imbalanced data,” says lead author Natalie Dullerud, a graduate student in the Healthy ML Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT.

      Dullerud’s co-authors include Kimia Hamidieh, a graduate student in the Healthy ML Group; Karsten Roth, a former visiting researcher who is now a graduate student at the University of Tubingen; Nicolas Papernot, an assistant professor in the University of Toronto’s Department of Electrical Engineering and Computer Science; and senior author Marzyeh Ghassemi, an assistant professor and head of the Healthy ML Group. The research will be presented at the International Conference on Learning Representations.

      Defining fairness

      The machine-learning technique the researchers studied is known as deep metric learning, which is a broad form of representation learning. In deep metric learning, a neural network learns the similarity between objects by mapping similar photos close together and dissimilar photos far apart. During training, this neural network maps images in an “embedding space” where a similarity metric between photos corresponds to the distance between them.

      For example, if a deep metric learning model is being used to classify bird species, it will map photos of golden finches together in one part of the embedding space and cardinals together in another part of the embedding space. Once trained, the model can effectively measure the similarity of new images it hasn’t seen before. It would learn to cluster images of an unseen bird species close together, but farther from cardinals or golden finches within the embedding space.

      The similarity metrics the model learns are very robust, which is why deep metric learning is so often employed for facial recognition, Dullerud says. But she and her colleagues wondered how to determine if a similarity metric is biased.

      “We know that data reflect the biases of processes in society. This means we have to shift our focus to designing methods that are better suited to reality,” says Ghassemi.

      The researchers defined two ways that a similarity metric can be unfair. Using the example of facial recognition, the metric will be unfair if it is more likely to embed individuals with darker-skinned faces closer to each other, even if they are not the same person, than it would if those images were people with lighter-skinned faces. Second, it will be unfair if the features it learns for measuring similarity are better for the majority group than for the minority group.

      The researchers ran a number of experiments on models with unfair similarity metrics and were unable to overcome the bias the model had learned in its embedding space.

      “This is quite scary because it is a very common practice for companies to release these embedding models and then people finetune them for some downstream classification task. But no matter what you do downstream, you simply can’t fix the fairness problems that were induced in the embedding space,” Dullerud says.

      Even if a user retrains the model on a balanced dataset for the downstream task, which is the best-case scenario for fixing the fairness problem, there are still performance gaps of at least 20 percent, she says.

      The only way to solve this problem is to ensure the embedding space is fair to begin with.

      Learning separate metrics

      The researchers’ solution, called Partial Attribute Decorrelation (PARADE), involves training the model to learn a separate similarity metric for a sensitive attribute, like skin tone, and then decorrelating the skin tone similarity metric from the targeted similarity metric. If the model is learning the similarity metrics of different human faces, it will learn to map similar faces close together and dissimilar faces far apart using features other than skin tone.

      Any number of sensitive attributes can be decorrelated from the targeted similarity metric in this way. And because the similarity metric for the sensitive attribute is learned in a separate embedding space, it is discarded after training so only the targeted similarity metric remains in the model.

      Their method is applicable to many situations because the user can control the amount of decorrelation between similarity metrics. For instance, if the model will be diagnosing breast cancer from mammogram images, a clinician likely wants some information about biological sex to remain in the final embedding space because it is much more likely that women will have breast cancer than men, Dullerud explains.

      They tested their method on two tasks, facial recognition and classifying bird species, and found that it reduced performance gaps caused by bias, both in the embedding space and in the downstream task, regardless of the dataset they used.

      Moving forward, Dullerud is interested in studying how to force a deep metric learning model to learn good features in the first place.

      “How do you properly audit fairness? That is an open question right now. How can you tell that a model is going to be fair, or that it is only going to be fair in certain situations, and what are those situations? Those are questions I am really interested in moving forward,” she says. More

    • 75 Shares179 Views

      in Data Management & Statistics

      Deep-learning technique predicts clinical treatment outcomes

      by

      When it comes to treatment strategies for critically ill patients, clinicians want to be able to consider all their options and timing of administration, and make the optimal decision for their patients. While clinician experience and study has helped them to be successful in this effort, not all patients are the same, and treatment decisions at this crucial time could mean the difference between patient improvement and quick deterioration. Therefore, it would be helpful for doctors to be able to take a patient’s previous known health status and received treatments and use that to predict that patient’s health outcome under different treatment scenarios, in order to pick the best path.

      Now, a deep-learning technique, called G-Net, from researchers at MIT and IBM provides a window into causal counterfactual prediction, affording physicians the opportunity to explore how a patient might fare under different treatment plans. The foundation of G-Net is the g-computation algorithm, a causal inference method that estimates the effect of dynamic exposures in the presence of measured confounding variables — ones that may influence both treatments and outcomes. Unlike previous implementations of the g-computation framework, which have used linear modeling approaches, G-Net uses recurrent neural networks (RNN), which have node connections that allow them to better model temporal sequences with complex and nonlinear dynamics, like those found in the physiological and clinical time series data. In this way, physicians can develop alternative plans based on patient history and test them before making a decision.

      “Our ultimate goal is to develop a machine learning technique that would allow doctors to explore various ‘What if’ scenarios and treatment options,” says Li-wei Lehman, MIT research scientist in the MIT Institute for Medical Engineering and Science and an MIT-IBM Watson AI Lab project lead. “A lot of work has been done in terms of deep learning for counterfactual prediction but [it’s] been focusing on a point exposure setting,” or a static, time-varying treatment strategy, which doesn’t allow for adjustment of treatments as patient history changes. However, her team’s new prediction approach provides for treatment plan flexibility and chances for treatment alteration over time as patient covariate history and past treatments change. “G-Net is the first deep-learning approach based on g-computation that can predict both the population-level and individual-level treatment effects under dynamic and time varying treatment strategies.”

      The research, which was recently published in the Proceedings of Machine Learning Research, was co-authored by Rui Li MEng ’20, Stephanie Hu MEng ’21, former MIT postdoc Mingyu Lu MD, graduate student Yuria Utsumi, IBM research staff member Prithwish Chakraborty, IBM Research director of Hybrid Cloud Services Daby Sow, IBM data scientist Piyush Madan, IBM research scientist Mohamed Ghalwash, and IBM research scientist Zach Shahn.

      Tracking disease progression

      To build, validate, and test G-Net’s predictive abilities, the researchers considered the circulatory system in septic patients in the ICU. During critical care, doctors need to make trade-offs and judgement calls, such as ensuring the organs are receiving adequate blood supply without overworking the heart. For this, they could give intravenous fluids to patients to increase blood pressure; however, too much can cause edema. Alternatively, physicians can administer vasopressors, which act to contract blood vessels and raise blood pressure.

      In order to mimic this and demonstrate G-Net’s proof-of-concept, the team used CVSim, a mechanistic model of a human cardiovascular system that’s governed by 28 input variables characterizing the system’s current state, such as arterial pressure, central venous pressure, total blood volume, and total peripheral resistance, and modified it to simulate various disease processes (e.g., sepsis or blood loss) and effects of interventions (e.g., fluids and vasopressors). The researchers used CVSim to generate observational patient data for training and for “ground truth” comparison against counterfactual prediction. In their G-Net architecture, the researchers ran two RNNs to handle and predict variables that are continuous, meaning they can take on a range of values, like blood pressure, and categorical variables, which have discrete values, like the presence or absence of pulmonary edema. The researchers simulated the health trajectories of thousands of “patients” exhibiting symptoms under one treatment regime, let’s say A, for 66 timesteps, and used them to train and validate their model.

      Testing G-Net’s prediction capability, the team generated two counterfactual datasets. Each contained roughly 1,000 known patient health trajectories, which were created from CVSim using the same “patient” condition as the starting point under treatment A. Then at timestep 33, treatment changed to plan B or C, depending on the dataset. The team then performed 100 prediction trajectories for each of these 1,000 patients, whose treatment and medical history was known up until timestep 33 when a new treatment was administered. In these cases, the prediction agreed well with the “ground-truth” observations for individual patients and averaged population-level trajectories.

      A cut above the rest

      Since the g-computation framework is flexible, the researchers wanted to examine G-Net’s prediction using different nonlinear models — in this case, long short-term memory (LSTM) models, which are a type of RNN that can learn from previous data patterns or sequences — against the more classical linear models and a multilayer perception model (MLP), a type of neural network that can make predictions using a nonlinear approach. Following a similar setup as before, the team found that the error between the known and predicted cases was smallest in the LSTM models compared to the others. Since G-Net is able to model the temporal patterns of the patient’s ICU history and past treatment, whereas a linear model and MLP cannot, it was better able to predict the patient’s outcome.

      The team also compared G-Net’s prediction in a static, time-varying treatment setting against two state-of-the-art deep-learning based counterfactual prediction approaches, a recurrent marginal structural network (rMSN) and a counterfactual recurrent neural network (CRN), as well as a linear model and an MLP. For this, they investigated a model for tumor growth under no treatment, radiation, chemotherapy, and both radiation and chemotherapy scenarios. “Imagine a scenario where there’s a patient with cancer, and an example of a static regime would be if you only give a fixed dosage of chemotherapy, radiation, or any kind of drug, and wait until the end of your trajectory,” comments Lu. For these investigations, the researchers generated simulated observational data using tumor volume as the primary influence dictating treatment plans and demonstrated that G-Net outperformed the other models. One potential reason could be because g-computation is known to be more statistically efficient than rMSN and CRN, when models are correctly specified.

      While G-Net has done well with simulated data, more needs to be done before it can be applied to real patients. Since neural networks can be thought of as “black boxes” for prediction results, the researchers are beginning to investigate the uncertainty in the model to help ensure safety. In contrast to these approaches that recommend an “optimal” treatment plan without any clinician involvement, “as a decision support tool, I believe that G-Net would be more interpretable, since the clinicians would input treatment strategies themselves,” says Lehman, and “G-Net will allow them to be able to explore different hypotheses.” Further, the team has moved on to using real data from ICU patients with sepsis, bringing it one step closer to implementation in hospitals.

      “I think it is pretty important and exciting for real-world applications,” says Hu. “It’d be helpful to have some way to predict whether or not a treatment might work or what the effects might be — a quicker iteration process for developing these hypotheses for what to try, before actually trying to implement them in in a years-long, potentially very involved and very invasive type of clinical trial.”

      This research was funded by the MIT-IBM Watson AI Lab. More

    • 75 Shares149 Views

      in Data Management & Statistics

      Can machine-learning models overcome biased datasets?

      by

      Artificial intelligence systems may be able to complete tasks quickly, but that doesn’t mean they always do so fairly. If the datasets used to train machine-learning models contain biased data, it is likely the system could exhibit that same bias when it makes decisions in practice.

      For instance, if a dataset contains mostly images of white men, then a facial-recognition model trained with these data may be less accurate for women or people with different skin tones.

      A group of researchers at MIT, in collaboration with researchers at Harvard University and Fujitsu Ltd., sought to understand when and how a machine-learning model is capable of overcoming this kind of dataset bias. They used an approach from neuroscience to study how training data affects whether an artificial neural network can learn to recognize objects it has not seen before. A neural network is a machine-learning model that mimics the human brain in the way it contains layers of interconnected nodes, or “neurons,” that process data.

      The new results show that diversity in training data has a major influence on whether a neural network is able to overcome bias, but at the same time dataset diversity can degrade the network’s performance. They also show that how a neural network is trained, and the specific types of neurons that emerge during the training process, can play a major role in whether it is able to overcome a biased dataset.

      “A neural network can overcome dataset bias, which is encouraging. But the main takeaway here is that we need to take into account data diversity. We need to stop thinking that if you just collect a ton of raw data, that is going to get you somewhere. We need to be very careful about how we design datasets in the first place,” says Xavier Boix, a research scientist in the Department of Brain and Cognitive Sciences (BCS) and the Center for Brains, Minds, and Machines (CBMM), and senior author of the paper.  

      Co-authors include former MIT graduate students Timothy Henry, Jamell Dozier, Helen Ho, Nishchal Bhandari, and Spandan Madan, a corresponding author who is currently pursuing a PhD at Harvard; Tomotake Sasaki, a former visiting scientist now a senior researcher at Fujitsu Research; Frédo Durand, a professor of electrical engineering and computer science at MIT and a member of the Computer Science and Artificial Intelligence Laboratory; and Hanspeter Pfister, the An Wang Professor of Computer Science at the Harvard School of Enginering and Applied Sciences. The research appears today in Nature Machine Intelligence.

      Thinking like a neuroscientist

      Boix and his colleagues approached the problem of dataset bias by thinking like neuroscientists. In neuroscience, Boix explains, it is common to use controlled datasets in experiments, meaning a dataset in which the researchers know as much as possible about the information it contains.

      The team built datasets that contained images of different objects in varied poses, and carefully controlled the combinations so some datasets had more diversity than others. In this case, a dataset had less diversity if it contains more images that show objects from only one viewpoint. A more diverse dataset had more images showing objects from multiple viewpoints. Each dataset contained the same number of images.

      The researchers used these carefully constructed datasets to train a neural network for image classification, and then studied how well it was able to identify objects from viewpoints the network did not see during training (known as an out-of-distribution combination). 

      For example, if researchers are training a model to classify cars in images, they want the model to learn what different cars look like. But if every Ford Thunderbird in the training dataset is shown from the front, when the trained model is given an image of a Ford Thunderbird shot from the side, it may misclassify it, even if it was trained on millions of car photos.

      The researchers found that if the dataset is more diverse — if more images show objects from different viewpoints — the network is better able to generalize to new images or viewpoints. Data diversity is key to overcoming bias, Boix says.

      “But it is not like more data diversity is always better; there is a tension here. When the neural network gets better at recognizing new things it hasn’t seen, then it will become harder for it to recognize things it has already seen,” he says.

      Testing training methods

      The researchers also studied methods for training the neural network.

      In machine learning, it is common to train a network to perform multiple tasks at the same time. The idea is that if a relationship exists between the tasks, the network will learn to perform each one better if it learns them together.

      But the researchers found the opposite to be true — a model trained separately for each task was able to overcome bias far better than a model trained for both tasks together.

      “The results were really striking. In fact, the first time we did this experiment, we thought it was a bug. It took us several weeks to realize it was a real result because it was so unexpected,” he says.

      They dove deeper inside the neural networks to understand why this occurs.

      They found that neuron specialization seems to play a major role. When the neural network is trained to recognize objects in images, it appears that two types of neurons emerge — one that specializes in recognizing the object category and another that specializes in recognizing the viewpoint.

      When the network is trained to perform tasks separately, those specialized neurons are more prominent, Boix explains. But if a network is trained to do both tasks simultaneously, some neurons become diluted and don’t specialize for one task. These unspecialized neurons are more likely to get confused, he says.

      “But the next question now is, how did these neurons get there? You train the neural network and they emerge from the learning process. No one told the network to include these types of neurons in its architecture. That is the fascinating thing,” he says.

      That is one area the researchers hope to explore with future work. They want to see if they can force a neural network to develop neurons with this specialization. They also want to apply their approach to more complex tasks, such as objects with complicated textures or varied illuminations.

      Boix is encouraged that a neural network can learn to overcome bias, and he is hopeful their work can inspire others to be more thoughtful about the datasets they are using in AI applications.

      This work was supported, in part, by the National Science Foundation, a Google Faculty Research Award, the Toyota Research Institute, the Center for Brains, Minds, and Machines, Fujitsu Research, and the MIT-Sensetime Alliance on Artificial Intelligence. More

    • 100 Shares149 Views

      in Data Management & Statistics

      The downside of machine learning in health care

      by

      While working toward her dissertation in computer science at MIT, Marzyeh Ghassemi wrote several papers on how machine-learning techniques from artificial intelligence could be applied to clinical data in order to predict patient outcomes. “It wasn’t until the end of my PhD work that one of my committee members asked: ‘Did you ever check to see how well your model worked across different groups of people?’”

      That question was eye-opening for Ghassemi, who had previously assessed the performance of models in aggregate, across all patients. Upon a closer look, she saw that models often worked differently — specifically worse — for populations including Black women, a revelation that took her by surprise. “I hadn’t made the connection beforehand that health disparities would translate directly to model disparities,” she says. “And given that I am a visible minority woman-identifying computer scientist at MIT, I am reasonably certain that many others weren’t aware of this either.”

      In a paper published Jan. 14 in the journal Patterns, Ghassemi — who earned her doctorate in 2017 and is now an assistant professor in the Department of Electrical Engineering and Computer Science and the MIT Institute for Medical Engineering and Science (IMES) — and her coauthor, Elaine Okanyene Nsoesie of Boston University, offer a cautionary note about the prospects for AI in medicine. “If used carefully, this technology could improve performance in health care and potentially reduce inequities,” Ghassemi says. “But if we’re not actually careful, technology could worsen care.”

      It all comes down to data, given that the AI tools in question train themselves by processing and analyzing vast quantities of data. But the data they are given are produced by humans, who are fallible and whose judgments may be clouded by the fact that they interact differently with patients depending on their age, gender, and race, without even knowing it.

      Furthermore, there is still great uncertainty about medical conditions themselves. “Doctors trained at the same medical school for 10 years can, and often do, disagree about a patient’s diagnosis,” Ghassemi says. That’s different from the applications where existing machine-learning algorithms excel — like object-recognition tasks — because practically everyone in the world will agree that a dog is, in fact, a dog.

      Machine-learning algorithms have also fared well in mastering games like chess and Go, where both the rules and the “win conditions” are clearly defined. Physicians, however, don’t always concur on the rules for treating patients, and even the win condition of being “healthy” is not widely agreed upon. “Doctors know what it means to be sick,” Ghassemi explains, “and we have the most data for people when they are sickest. But we don’t get much data from people when they are healthy because they’re less likely to see doctors then.”

      Even mechanical devices can contribute to flawed data and disparities in treatment. Pulse oximeters, for example, which have been calibrated predominately on light-skinned individuals, do not accurately measure blood oxygen levels for people with darker skin. And these deficiencies are most acute when oxygen levels are low — precisely when accurate readings are most urgent. Similarly, women face increased risks during “metal-on-metal” hip replacements, Ghassemi and Nsoesie write, “due in part to anatomic differences that aren’t taken into account in implant design.” Facts like these could be buried within the data fed to computer models whose output will be undermined as a result.

      Coming from computers, the product of machine-learning algorithms offers “the sheen of objectivity,” according to Ghassemi. But that can be deceptive and dangerous, because it’s harder to ferret out the faulty data supplied en masse to a computer than it is to discount the recommendations of a single possibly inept (and maybe even racist) doctor. “The problem is not machine learning itself,” she insists. “It’s people. Human caregivers generate bad data sometimes because they are not perfect.”

      Nevertheless, she still believes that machine learning can offer benefits in health care in terms of more efficient and fairer recommendations and practices. One key to realizing the promise of machine learning in health care is to improve the quality of data, which is no easy task. “Imagine if we could take data from doctors that have the best performance and share that with other doctors that have less training and experience,” Ghassemi says. “We really need to collect this data and audit it.”

      The challenge here is that the collection of data is not incentivized or rewarded, she notes. “It’s not easy to get a grant for that, or ask students to spend time on it. And data providers might say, ‘Why should I give my data out for free when I can sell it to a company for millions?’ But researchers should be able to access data without having to deal with questions like: ‘What paper will I get my name on in exchange for giving you access to data that sits at my institution?’

      “The only way to get better health care is to get better data,” Ghassemi says, “and the only way to get better data is to incentivize its release.”

      It’s not only a question of collecting data. There’s also the matter of who will collect it and vet it. Ghassemi recommends assembling diverse groups of researchers — clinicians, statisticians, medical ethicists, and computer scientists — to first gather diverse patient data and then “focus on developing fair and equitable improvements in health care that can be deployed in not just one advanced medical setting, but in a wide range of medical settings.”

      The objective of the Patterns paper is not to discourage technologists from bringing their expertise in machine learning to the medical world, she says. “They just need to be cognizant of the gaps that appear in treatment and other complexities that ought to be considered before giving their stamp of approval to a particular computer model.” More

    • 100 Shares189 Views

      in Data Management & Statistics

      When should someone trust an AI assistant’s predictions?

      by

      In a busy hospital, a radiologist is using an artificial intelligence system to help her diagnose medical conditions based on patients’ X-ray images. Using the AI system can help her make faster diagnoses, but how does she know when to trust the AI’s predictions?

      She doesn’t. Instead, she may rely on her expertise, a confidence level provided by the system itself, or an explanation of how the algorithm made its prediction — which may look convincing but still be wrong — to make an estimation.

      To help people better understand when to trust an AI “teammate,” MIT researchers created an onboarding technique that guides humans to develop a more accurate understanding of those situations in which a machine makes correct predictions and those in which it makes incorrect predictions.

      By showing people how the AI complements their abilities, the training technique could help humans make better decisions or come to conclusions faster when working with AI agents.

      “We propose a teaching phase where we gradually introduce the human to this AI model so they can, for themselves, see its weaknesses and strengths,” says Hussein Mozannar, a graduate student in the Social and Engineering Systems doctoral program within the Institute for Data, Systems, and Society (IDSS) who is also a researcher with the Clinical Machine Learning Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL) and the Institute for Medical Engineering and Science. “We do this by mimicking the way the human will interact with the AI in practice, but we intervene to give them feedback to help them understand each interaction they are making with the AI.”

      Mozannar wrote the paper with Arvind Satyanarayan, an assistant professor of computer science who leads the Visualization Group in CSAIL; and senior author David Sontag, an associate professor of electrical engineering and computer science at MIT and leader of the Clinical Machine Learning Group. The research will be presented at the Association for the Advancement of Artificial Intelligence in February.

      Mental models

      This work focuses on the mental models humans build about others. If the radiologist is not sure about a case, she may ask a colleague who is an expert in a certain area. From past experience and her knowledge of this colleague, she has a mental model of his strengths and weaknesses that she uses to assess his advice.

      Humans build the same kinds of mental models when they interact with AI agents, so it is important those models are accurate, Mozannar says. Cognitive science suggests that humans make decisions for complex tasks by remembering past interactions and experiences. So, the researchers designed an onboarding process that provides representative examples of the human and AI working together, which serve as reference points the human can draw on in the future. They began by creating an algorithm that can identify examples that will best teach the human about the AI.

      “We first learn a human expert’s biases and strengths, using observations of their past decisions unguided by AI,” Mozannar says. “We combine our knowledge about the human with what we know about the AI to see where it will be helpful for the human to rely on the AI. Then we obtain cases where we know the human should rely on the AI and similar cases where the human should not rely on the AI.”

      The researchers tested their onboarding technique on a passage-based question answering task: The user receives a written passage and a question whose answer is contained in the passage. The user then has to answer the question and can click a button to “let the AI answer.” The user can’t see the AI answer in advance, however, requiring them to rely on their mental model of the AI. The onboarding process they developed begins by showing these examples to the user, who tries to make a prediction with the help of the AI system. The human may be right or wrong, and the AI may be right or wrong, but in either case, after solving the example, the user sees the correct answer and an explanation for why the AI chose its prediction. To help the user generalize from the example, two contrasting examples are shown that explain why the AI got it right or wrong.

      For instance, perhaps the training question asks which of two plants is native to more continents, based on a convoluted paragraph from a botany textbook. The human can answer on her own or let the AI system answer. Then, she sees two follow-up examples that help her get a better sense of the AI’s abilities. Perhaps the AI is wrong on a follow-up question about fruits but right on a question about geology. In each example, the words the system used to make its prediction are highlighted. Seeing the highlighted words helps the human understand the limits of the AI agent, explains Mozannar.

      To help the user retain what they have learned, the user then writes down the rule she infers from this teaching example, such as “This AI is not good at predicting flowers.” She can then refer to these rules later when working with the agent in practice. These rules also constitute a formalization of the user’s mental model of the AI.

      The impact of teaching

      The researchers tested this teaching technique with three groups of participants. One group went through the entire onboarding technique, another group did not receive the follow-up comparison examples, and the baseline group didn’t receive any teaching but could see the AI’s answer in advance.

      “The participants who received teaching did just as well as the participants who didn’t receive teaching but could see the AI’s answer. So, the conclusion there is they are able to simulate the AI’s answer as well as if they had seen it,” Mozannar says.

      The researchers dug deeper into the data to see the rules individual participants wrote. They found that almost 50 percent of the people who received training wrote accurate lessons of the AI’s abilities. Those who had accurate lessons were right on 63 percent of the examples, whereas those who didn’t have accurate lessons were right on 54 percent. And those who didn’t receive teaching but could see the AI answers were right on 57 percent of the questions.

      “When teaching is successful, it has a significant impact. That is the takeaway here. When we are able to teach participants effectively, they are able to do better than if you actually gave them the answer,” he says.

      But the results also show there is still a gap. Only 50 percent of those who were trained built accurate mental models of the AI, and even those who did were only right 63 percent of the time. Even though they learned accurate lessons, they didn’t always follow their own rules, Mozannar says.

      That is one question that leaves the researchers scratching their heads — even if people know the AI should be right, why won’t they listen to their own mental model? They want to explore this question in the future, as well as refine the onboarding process to reduce the amount of time it takes. They are also interested in running user studies with more complex AI models, particularly in health care settings.

      “When humans collaborate with other humans, we rely heavily on knowing what our collaborators’ strengths and weaknesses are — it helps us know when (and when not) to lean on the other person for assistance. I’m glad to see this research applying that principle to humans and AI,” says Carrie Cai, a staff research scientist in the People + AI Research and Responsible AI groups at Google, who was not involved with this research. “Teaching users about an AI’s strengths and weaknesses is essential to producing positive human-AI joint outcomes.” 

      This research was supported, in part, by the National Science Foundation. More