Computer Science and Artificial Intelligence Laboratory (CSAIL)

Subterms

    More stories

    • 88 Shares99 Views

      in Data Management & Statistics

      MIT researchers make language models scalable self-learners

      by

      Socrates once said: “It is not the size of a thing, but the quality that truly matters. For it is in the nature of substance, not its volume, that true value is found.”

      Does size always matter for large language models (LLMs)? In a technological landscape bedazzled by LLMs taking center stage, a team of MIT Computer Science and Artificial Intelligence Laboratory (CSAIL) researchers think smaller models shouldn’t be overlooked, especially for natural language understanding products widely deployed in the industry.

      To that end, the researchers cooked up an approach to long-standing problems of inefficiency and privacy associated with big, text-based AI models — a logic-aware model that outperforms 500-times-bigger counterparts on some language understanding tasks without human-generated annotations, while preserving privacy and robustness with high performance.

      LLMs, which have shown some promising skills in generating language, art, and code, are computationally expensive, and their data requirements can risk privacy leaks when using application programming interfaces for data upload. Smaller models have been historically less capable, particularly in multitasking and weakly supervised tasks, compared to their larger counterparts.

      So what’s helping these smaller models act so mighty, then? Something called “textual entailment,” a way to help these models understand a variety of language tasks, where if one sentence (the premise) is true, then the other sentence (the hypothesis) is likely to be true as well. For example, if the premise is, “all cats have tails” then the hypothesis “a tabby cat has a tail” would be entailed by the premise. This concept is used to train an “entailment model” that proved to be less biased than other language models, from the team’s previous research. They then created “prompts” that the models can use to figure out if certain information is entailed by a given sentence or phrase according to different tasks. This method improved the model’s ability to adapt to different tasks without any additional training, known as zero-shot adaptation.

      In the realm of “natural language understanding,” there are various applications that hinge on determining the relationship between two pieces of text. For example, in sentiment classification, a statement like “I think the movie is good” can be inferred or entailed from a movie review that says, “I like the story and the acting is great,” indicating a positive sentiment. Another is news classification, where the topic of a news article can be inferred from its content. For example, a statement like “the news article is about sports” can be entailed if the main content of the article reports on an NBA game. The key insight was that many existing natural language understanding tasks could be recast as an entailment (i.e., logical inference in natural language) task. 

      “Our research is about improving the ability of computer programs to understand and process natural language — the way humans speak and write. Our self-trained, 350-million-parameter entailment models, without human-generated labels, outperform supervised language models with 137 to 175 billion parameters,” says MIT CSAIL postdoc Hongyin Luo, lead author on a new paper about the study. “This has potential to reshape the landscape of AI and machine learning, providing a more scalable, trustworthy, and cost-effective solution to language modeling,” says Luo. “By proving that smaller models can perform at the same level as larger ones for language understanding, this work paves the way for more sustainable and privacy-preserving AI technologies.” 

      The team discovered that they could improve the model’s performance even more by using a technique called “self-training,” where the model uses its own predictions to teach itself, effectively learning without human supervision and additional annotated training data.The self-training method significantly improved performance on a bunch of downstream tasks, including sentiment analysis, question-answering, and news classification. It outperformed both Google’s LaMDA and FLAN in zero-shot capabilities, GPT models, and other supervised algorithms. 

      However, one challenge with self-training is that the model can sometimes generate incorrect or noisy labels that harm performance. To overcome this, they developed a new algorithm called ‘SimPLE’ (Simple Pseudo-Label Editing), a process to review and modify the pseudo-labels made in initial rounds of learning. By correcting any mislabeled instances, it improved the overall quality of the self-generated labels. This not only made the models more effective at understanding language, but more robust when faced with adversarial data. 

      As with most research, there are some limitations. The self-training on multi-class classification tasks didn’t perform as well as on binary natural language understanding tasks, indicating the challenge of applying entailment models to multi-choice tasks.“This research presents an efficient and effective way to train large language models (LLMs) by formulating natural language understanding tasks as contextual entailment problems and employing a pseudo-labeling self-training mechanism to incorporate large quantities of unlabelled text data in the training process,” adds CSAIL Senior Research Scientist James Glass, who is also an author on the paper. “While the field of LLMs is undergoing rapid and dramatic changes, this research shows that it is possible to produce relatively compact language models that perform very well on benchmark understanding tasks compared to their peers of roughly the same size, or even much larger language models.”

      “Entailment task is a popular proxy to evaluate “understanding” of a given context by an AI model,” says Leonid Karlinsky, research staff member at the MIT-IBM Watson AI Lab. “It is used in many areas analyzing models with unimodal, like LLMs, and and multi-modal, like VLMs [visual language models] inputs, simplifying the task of question-answering about a given input context to a binary classification problem — does this context entail a certain (e.g., text) conclusion or not? This paper makes two contributions in this space. First, it proposes a way to improve the zero-shot (without additional tuning) NLU performance and robustness to adversarial attacks via tuning with synthesized (specialized) entailment tasks generated for the primal NLU task. Second, it offers a self-supervised SimPLE method including pseudo-labeling and confidence-based filtering to further improve large LLMs’ NLU performance.”

      Luo and Glass wrote the paper with Yoon Kim, a CSAIL member and assistant professor in MIT’s Department of Electrical Engineering and Computer Science, and Jiaxin Ge of Peking University. Their work will be presented at the meeting of the Association for Computational Linguistics in Toronto, Ontario this July. This research was supported by a grant from the Hong Kong Innovation AI program. More

    • 63 Shares119 Views

      in Data Management & Statistics

      Scaling audio-visual learning without labels

      by

      Researchers from MIT, the MIT-IBM Watson AI Lab, IBM Research, and elsewhere have developed a new technique for analyzing unlabeled audio and visual data that could improve the performance of machine-learning models used in applications like speech recognition and object detection. The work, for the first time, combines two architectures of self-supervised learning, contrastive learning and masked data modeling, in an effort to scale machine-learning tasks like event classification in single- and multimodal data without the need for annotation, thereby replicating how humans understand and perceive our world.

      “A larger portion of human knowledge is learned in a self-supervised way, because we don’t always get supervision signals, and we want to enable the machine-learning model to have the same ability,” says Yuan Gong, an MIT postdoc in the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      “So, another way to put it is that self-supervised learning often forms the foundation of an initial model, because it can learn on vast amounts of unlabeled data. And then you can use classical, supervised learning or reinforcement learning to fine tune the model to something particular if you want to,” says Jim Glass, an MIT senior research scientist and member of the MIT-IBM Watson AI Lab.

      The technique, called the contrastive audio-visual masked autoencoder (CAV-MAE), is a type of neural network that can learn to extract and map meaningful latent representations into high-dimensional space from acoustic and visual data by training on large YouTube datasets of audio and video 10-second clips. The researchers say the technique is more effective than previous approaches because it explicitly models the relationships between audio and visual data in a way that other methods do not.

      Joining Gong and Glass on the study are graduate students Andrew Rouditchenko and Alexander H. Liu of MIT, David Harwath PhD ’18 of the University of Texas at Austin, and MIT-IBM Watson AI Lab members Leonid Karlinsky and Hilde Kuehne. Kuehne is also affiliated with Goethe University Frankfurt. The method was recently presented at the International Conference on Learning Representations.

      A joint and coordinated approach

      The CAV-MAE works by “learning by prediction” and “learning by comparison,” says Gong. The masked data modeling, or the prediction method, takes a video along with its coordinated audio waveform, converts the audio to a spectrogram, and masks 75 percent of both. The unmasked data is tokenized, then fed into separate audio and visual encoders before entering a joint encoder/decoder, where the model is asked to recover the missing data. The difference (reconstruction loss) between the resulting reconstructed prediction and the original audio-visual combination is then used to train the model for better performance. An example of this would be covering part of a video of a piano and part of a spectrogram of piano music, and then asking the model to try to determine the masked inputs. Unfortunately, this method may not capture the association between the video and audio pair, whereas contrastive learning leverages this, but may discard some modality-unique information, like the background in a video.

      Contrastive learning aims to map representations that are similar close to each other. For example, the model will attempt to place different video and audio data of different parrots close to each other and further away from pairs of video and audio of guitars playing. In a similar fashion to masked autoencoding, audio-visual pairs are passed into separate modality encoders; however, the audio and visual components are kept separately within the joint encoder before the model performs pooling and contrastive loss. In this way, contrastive learning tries to identify the parts of each audio or video that are most relevant to the other. For example, if a video shows someone speaking and the corresponding audio clip contains speech, the autoencoder will learn to associate the mouth movements of the speaker with the words being spoken. It will then adjust the model’s parameters so that those inputs are represented close to each other. Ultimately, the CAV-MAE method combines both techniques with multiple forward data streams with masking as a first step, modality-specific encoders, and layer normalization so that the representation strengths are similar.

      “We [then] wanted to compare the proposed CAV-MAE with a model trained only with a masked autoencoder and a model trained only with contrastive learning, because we want to show that by combining masked autoencoder and contrastive learning, we can get some performance improvement,” says Gong, “and the results support our hypothesis that there’s obvious improvement.”

      The researchers tested CAV-MAE — as well as their method without contrastive loss or a masked autoencoder — against other state-of-the-art methods on audio-visual retrieval and audio-visual event classification tasks using standard AudioSet (20K and 2M) and VGGSound datasets — labeled, realistic short clips, which could include multiple sounds. Audio-visual retrieval means that the model sees either the audio or visual component of a query pair and searches for the missing one; event classification includes identifying actions or sounds within data, like a person singing or a car driving.

      Overall, they found that contrastive learning and masked data modeling are complementary methods. CAV-MAE was able to outperform previous techniques (with fully self-supervised pre-training) by about 2 percent for event classification performance verses models with comparable computation and, more impressively, kept pace with or outperformed models with industry-level computational resources. The team’s model ranked similarly to models trained with only the contrastive loss. And surprisingly, the team says, the incorporation of multi-modal data into CAV-MAE pre-training greatly improves the fine-tuning of single-modality representation via supervised learning (with some labeled data) and performance on audio-only event classification tasks. This demonstrates that, like humans, multi-modal information provides an additional “soft label” boost even for audio or visual only tasks; for instance, it helps the model to understand if it’s looking for an electric or acoustic guitar — a richer supervision signal.

      “I think people like the elegance of this model for combining information in the different audio and visual streams. It has the contrastive and the reconstruction loss, and compared to models that have been evaluated with similar data, it clearly does very well across a range of these tasks,” says Glass.

      Building on this, “one special thing is, our model can do both classification and the retrieval, which is not common,” Gong adds. “Before this work, these methods are used separately, but after this work, I see that most of the audio-visual learning frameworks use contracting loss and the masked autoencoder together, implicitly or explicitly.”

      Bringing self-supervised audio-visual learning into our world

      The researchers see their contribution of the contrastive audio-visual masked autoencoder (CAV-MAE) as an important milestone and a step forward for applications, which are increasingly moving from single modality to multi-modality and which require or leverage audio-visual fusion. They hypothesize that one day it could be used for action recognition in realms like sports, education, entertainment, motor vehicles, and public safety. It could also, one day, extend to other modalities. At this time, the fact that, “this only applies to audio-visual data may be a limitation, but we are targeting multi-modal learning, which is trend of machine learning,” says Gong. “As humans, we have multi-modalities — we have smell, touch — many more things that just audio-visual. So, when we try to build AI, we try to mimic humans somehow, not necessarily from the biological perspective, and this method could [potentially be] generalized to other unexplored modalities.”

      As machine-learning models continue to play an increasingly important role in our lives, techniques like this one will become increasingly valuable.

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

    • 88 Shares99 Views

      in Data Management & Statistics

      Using data to write songs for progress

      by

      A three-year recipient of MIT’s Emerson Classical Vocal Scholarships, senior Ananya Gurumurthy recalls getting ready to step onto the Carnegie Hall stage to sing a Mozart opera that she once sang with the New York All-State Choir. The choir conductor reminded her to articulate her words and to engage her diaphragm.

      “If you don’t project your voice, how are people going to hear you when you perform?” Gurumurthy recalls her conductor telling her. “This is your moment, your chance to connect with such a tremendous audience.”

      Gurumurthy reflects on the universal truth of those words as she adds her musical talents to her math and computer science studies to campaign for social and economic justice.

      The daughter of immigrants

      Growing up in Edgemont, New York, she was inspired to fight on behalf of others by her South Asian immigrant parents, who came to the United States in the 1980s. Her father is a management consultant and her mother has experience as an investment banker.

      “They came barely 15 years after the passage of the 1965 Immigration and Nationality Act, which removed national origin quotas from the American immigration system,” she says. “I would not be here if it had not been for the Civil Rights Movement, which preceded both me and my parents.”

      Her parents told her about their new home’s anti-immigrant sentiments; for example, her father was a graduate student in Dallas exiting a store when he was pelted with glass bottles and racial slurs.

      “I often consider the amount of bravery that it must have taken them to abandon everything they knew to immigrate to a new, but still imperfect, country in search of something better,” she says. “As a result, I have always felt so grounded in my identity both as a South Asian American and a woman of color. These identities have allowed me to think critically about how I can most effectively reform the institutions surrounding me.”

      Gurumurthy has been singing since she was 11, but in high school, she decided to also build her political voice by working for New York Senator Andrea Stewart-Cousins. At one point, Gurumurthy noted a log was kept for the subjects of constituent calls, such as “affordable housing” and  “infrastructure,” and it was then that she became aware that Stewart-Cousins would address the most pressing of these callers’ issues before the Senate.

      “This experience was my first time witnessing how powerful the mobilization of constituents in vast numbers was for influencing meaningful legislative change,” says Gurumurthy.

      After she began applying her math skills to political campaigns, Gurumurthy was soon tapped to run analytics for the Democratic National Committee’s (DNC) midterm election initiative. As a lead analyst for the New York DNC, she adapted an interactive activation-competition (IAC) model to understand voting patterns in the 2018 and 2020 elections. She collected data from public voting records to predict how constituents would cast their ballots and used an IAC algorithm to strategize alongside grassroots organizations and allocate resources to empower historically disenfranchised groups in municipal, state, and federal elections to encourage them to vote.

      Research and student organizing at MIT

      When she arrived at MIT in 2019 to study mathematics with computer science, along with minors in music and economics, she admits she was saddled with the naïve notion that she would “build digital tools that could single-handedly alleviate all of the collective pressures of systemic injustice in this country.” 

      Since then, she has learned to create what she calls “a more nuanced view.” She picked up data analytics skills to build mobilization platforms for organizations that pursued social and economic justice, including working in Fulton County, Georgia, with Fair Fight Action (through the Kelly-Douglas Fund Scholarship) to analyze patterns of voter suppression, and MIT’s ethics laboratories in the Computer Science and Artificial Intelligence Laboratory to build symbolic artificial intelligence protocols to better understand bias in artificial intelligence algorithms. For her work on the International Monetary Fund (through the MIT Washington Summer Internship Program), Gurumurthy was awarded second place for the 2022 S. Klein Prize in Technical Writing for her paper “The Rapid Rise of Cryptocurrency.”

      “The outcomes of each project gave me more hope to begin the next because I could see the impact of these digital tools,” she says. “I saw people feel empowered to use their voices whether it was voting for the first time, protesting exploitative global monetary policy, or fighting gender discrimination. I’ve been really fortunate to see the power of mathematical analysis firsthand.”

      “I have come to realize that the constructive use of technology could be a powerful voice of resistance against injustice,” she says. “Because numbers matter, and when people bear witness to them, they are pushed to take action in meaningful ways.”

      Hoping to make a difference in her own community, she joined several Institute committees. As co-chair of the Undergraduate Association’s education committee, she propelled MIT’s first-ever digital petition for grade transparency and worked with faculty members on Institute committees to ensure that all students were being provided adequate resources to participate in online education in the wake of the Covid-19 pandemic. The digital petition inspired her to begin a project, called Insite, to develop a more centralized digital means of data collection on student life at MIT to better inform policies made by its governing bodies. As Ring Committee chair, she ensured that the special traditions of the “Brass Rat” were made economically accessible to all class members by helping the committee nearly triple its financial aid budget. For her efforts at MIT, last May she received the William L. Stewart, Jr. Award for “[her] contributions [as] an individual student at MIT to extracurricular activities and student life.”

      Ananya plans on going to law school after graduation, to study constitutional law so that she can use her technical background to build quantitative evidence in cases pertaining to voting rights, social welfare, and ethical technology, and set legal standards ”for the humane use of data,” she says.

      “In building digital tools for a variety of social and economic justice organizations, I hope that we can challenge our existing systems of power and realize the progress we so dearly need to witness. There is strength in numbers, both algorithmically and organizationally. I believe it is our responsibility to simultaneously use these strengths to change the world.”

      Her ambitions, however, began when she began singing lessons when she was 11; without her background as a vocalist, she says she would be voiceless.

      “Operatic performance has given me the ability to truly step into my character and convey powerful emotions in my performance. In the process, I have realized that my voice is most powerful when it reflects my true convictions, whether I am performing or publicly speaking. I truly believe that this honesty has allowed me to become an effective community organizer. I’d like to believe that this voice is what compels those around me to act.”

      Private musical study is available for students through the Emerson/Harris Program, which offers merit-based financial awards to students of outstanding achievement on their instruments or voice in classical, jazz, or world music. The Emerson/Harris Program is funded by the late Cherry L. Emerson Jr. SM ’41, in response to an appeal from Associate Provost Ellen T. Harris (Class of 1949 professor emeritus of music). More

    • 138 Shares129 Views

      in Data Management & Statistics

      Study: AI models fail to reproduce human judgements about rule violations

      by

      In an effort to improve fairness or reduce backlogs, machine-learning models are sometimes designed to mimic human decision making, such as deciding whether social media posts violate toxic content policies.

      But researchers from MIT and elsewhere have found that these models often do not replicate human decisions about rule violations. If models are not trained with the right data, they are likely to make different, often harsher judgements than humans would.

      In this case, the “right” data are those that have been labeled by humans who were explicitly asked whether items defy a certain rule. Training involves showing a machine-learning model millions of examples of this “normative data” so it can learn a task.

      But data used to train machine-learning models are typically labeled descriptively — meaning humans are asked to identify factual features, such as, say, the presence of fried food in a photo. If “descriptive data” are used to train models that judge rule violations, such as whether a meal violates a school policy that prohibits fried food, the models tend to over-predict rule violations.

      This drop in accuracy could have serious implications in the real world. For instance, if a descriptive model is used to make decisions about whether an individual is likely to reoffend, the researchers’ findings suggest it may cast stricter judgements than a human would, which could lead to higher bail amounts or longer criminal sentences.

      “I think most artificial intelligence/machine-learning researchers assume that the human judgements in data and labels are biased, but this result is saying something worse. These models are not even reproducing already-biased human judgments because the data they’re being trained on has a flaw: Humans would label the features of images and text differently if they knew those features would be used for a judgment. This has huge ramifications for machine learning systems in human processes,” says Marzyeh Ghassemi, an assistant professor and head of the Healthy ML Group in the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      Ghassemi is senior author of a new paper detailing these findings, which was published today in Science Advances. Joining her on the paper are lead author Aparna Balagopalan, an electrical engineering and computer science graduate student; David Madras, a graduate student at the University of Toronto; David H. Yang, a former graduate student who is now co-founder of ML Estimation; Dylan Hadfield-Menell, an MIT assistant professor; and Gillian K. Hadfield, Schwartz Reisman Chair in Technology and Society and professor of law at the University of Toronto.

      Labeling discrepancy

      This study grew out of a different project that explored how a machine-learning model can justify its predictions. As they gathered data for that study, the researchers noticed that humans sometimes give different answers if they are asked to provide descriptive or normative labels about the same data.

      To gather descriptive labels, researchers ask labelers to identify factual features — does this text contain obscene language? To gather normative labels, researchers give labelers a rule and ask if the data violates that rule — does this text violate the platform’s explicit language policy?

      Surprised by this finding, the researchers launched a user study to dig deeper. They gathered four datasets to mimic different policies, such as a dataset of dog images that could be in violation of an apartment’s rule against aggressive breeds. Then they asked groups of participants to provide descriptive or normative labels.

      In each case, the descriptive labelers were asked to indicate whether three factual features were present in the image or text, such as whether the dog appears aggressive. Their responses were then used to craft judgements. (If a user said a photo contained an aggressive dog, then the policy was violated.) The labelers did not know the pet policy. On the other hand, normative labelers were given the policy prohibiting aggressive dogs, and then asked whether it had been violated by each image, and why.

      The researchers found that humans were significantly more likely to label an object as a violation in the descriptive setting. The disparity, which they computed using the absolute difference in labels on average, ranged from 8 percent on a dataset of images used to judge dress code violations to 20 percent for the dog images.

      “While we didn’t explicitly test why this happens, one hypothesis is that maybe how people think about rule violations is different from how they think about descriptive data. Generally, normative decisions are more lenient,” Balagopalan says.

      Yet data are usually gathered with descriptive labels to train a model for a particular machine-learning task. These data are often repurposed later to train different models that perform normative judgements, like rule violations.

      Training troubles

      To study the potential impacts of repurposing descriptive data, the researchers trained two models to judge rule violations using one of their four data settings. They trained one model using descriptive data and the other using normative data, and then compared their performance.

      They found that if descriptive data are used to train a model, it will underperform a model trained to perform the same judgements using normative data. Specifically, the descriptive model is more likely to misclassify inputs by falsely predicting a rule violation. And the descriptive model’s accuracy was even lower when classifying objects that human labelers disagreed about.

      “This shows that the data do really matter. It is important to match the training context to the deployment context if you are training models to detect if a rule has been violated,” Balagopalan says.

      It can be very difficult for users to determine how data have been gathered; this information can be buried in the appendix of a research paper or not revealed by a private company, Ghassemi says.

      Improving dataset transparency is one way this problem could be mitigated. If researchers know how data were gathered, then they know how those data should be used. Another possible strategy is to fine-tune a descriptively trained model on a small amount of normative data. This idea, known as transfer learning, is something the researchers want to explore in future work.

      They also want to conduct a similar study with expert labelers, like doctors or lawyers, to see if it leads to the same label disparity.

      “The way to fix this is to transparently acknowledge that if we want to reproduce human judgment, we must only use data that were collected in that setting. Otherwise, we are going to end up with systems that are going to have extremely harsh moderations, much harsher than what humans would do. Humans would see nuance or make another distinction, whereas these models don’t,” Ghassemi says.

      This research was funded, in part, by the Schwartz Reisman Institute for Technology and Society, Microsoft Research, the Vector Institute, and a Canada Research Council Chain. More

    • 113 Shares119 Views

      in Data Management & Statistics

      Researchers create a tool for accurately simulating complex systems

      by

      Researchers often use simulations when designing new algorithms, since testing ideas in the real world can be both costly and risky. But since it’s impossible to capture every detail of a complex system in a simulation, they typically collect a small amount of real data that they replay while simulating the components they want to study.

      Known as trace-driven simulation (the small pieces of real data are called traces), this method sometimes results in biased outcomes. This means researchers might unknowingly choose an algorithm that is not the best one they evaluated, and which will perform worse on real data than the simulation predicted that it should.

      MIT researchers have developed a new method that eliminates this source of bias in trace-driven simulation. By enabling unbiased trace-driven simulations, the new technique could help researchers design better algorithms for a variety of applications, including improving video quality on the internet and increasing the performance of data processing systems.

      The researchers’ machine-learning algorithm draws on the principles of causality to learn how the data traces were affected by the behavior of the system. In this way, they can replay the correct, unbiased version of the trace during the simulation.

      When compared to a previously developed trace-driven simulator, the researchers’ simulation method correctly predicted which newly designed algorithm would be best for video streaming — meaning the one that led to less rebuffering and higher visual quality. Existing simulators that do not account for bias would have pointed researchers to a worse-performing algorithm.

      “Data are not the only thing that matter. The story behind how the data are generated and collected is also important. If you want to answer a counterfactual question, you need to know the underlying data generation story so you only intervene on those things that you really want to simulate,” says Arash Nasr-Esfahany, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper on this new technique.

      He is joined on the paper by co-lead authors and fellow EECS graduate students Abdullah Alomar and Pouya Hamadanian; recent graduate student Anish Agarwal PhD ’21; and senior authors Mohammad Alizadeh, an associate professor of electrical engineering and computer science; and Devavrat Shah, the Andrew and Erna Viterbi Professor in EECS and a member of the Institute for Data, Systems, and Society and of the Laboratory for Information and Decision Systems. The research was recently presented at the USENIX Symposium on Networked Systems Design and Implementation.

      Specious simulations

      The MIT researchers studied trace-driven simulation in the context of video streaming applications.

      In video streaming, an adaptive bitrate algorithm continually decides the video quality, or bitrate, to transfer to a device based on real-time data on the user’s bandwidth. To test how different adaptive bitrate algorithms impact network performance, researchers can collect real data from users during a video stream for a trace-driven simulation.

      They use these traces to simulate what would have happened to network performance had the platform used a different adaptive bitrate algorithm in the same underlying conditions.

      Researchers have traditionally assumed that trace data are exogenous, meaning they aren’t affected by factors that are changed during the simulation. They would assume that, during the period when they collected the network performance data, the choices the bitrate adaptation algorithm made did not affect those data.

      But this is often a false assumption that results in biases about the behavior of new algorithms, making the simulation invalid, Alizadeh explains.

      “We recognized, and others have recognized, that this way of doing simulation can induce errors. But I don’t think people necessarily knew how significant those errors could be,” he says.

      To develop a solution, Alizadeh and his collaborators framed the issue as a causal inference problem. To collect an unbiased trace, one must understand the different causes that affect the observed data. Some causes are intrinsic to a system, while others are affected by the actions being taken.

      In the video streaming example, network performance is affected by the choices the bitrate adaptation algorithm made — but it’s also affected by intrinsic elements, like network capacity.

      “Our task is to disentangle these two effects, to try to understand what aspects of the behavior we are seeing are intrinsic to the system and how much of what we are observing is based on the actions that were taken. If we can disentangle these two effects, then we can do unbiased simulations,” he says.

      Learning from data

      But researchers often cannot directly observe intrinsic properties. This is where the new tool, called CausalSim, comes in. The algorithm can learn the underlying characteristics of a system using only the trace data.

      CausalSim takes trace data that were collected through a randomized control trial, and estimates the underlying functions that produced those data. The model tells the researchers, under the exact same underlying conditions that a user experienced, how a new algorithm would change the outcome.

      Using a typical trace-driven simulator, bias might lead a researcher to select a worse-performing algorithm, even though the simulation indicates it should be better. CausalSim helps researchers select the best algorithm that was tested.

      The MIT researchers observed this in practice. When they used CausalSim to design an improved bitrate adaptation algorithm, it led them to select a new variant that had a stall rate that was nearly 1.4 times lower than a well-accepted competing algorithm, while achieving the same video quality. The stall rate is the amount of time a user spent rebuffering the video.

      By contrast, an expert-designed trace-driven simulator predicted the opposite. It indicated that this new variant should cause a stall rate that was nearly 1.3 times higher. The researchers tested the algorithm on real-world video streaming and confirmed that CausalSim was correct.

      “The gains we were getting in the new variant were very close to CausalSim’s prediction, while the expert simulator was way off. This is really exciting because this expert-designed simulator has been used in research for the past decade. If CausalSim can so clearly be better than this, who knows what we can do with it?” says Hamadanian.

      During a 10-month experiment, CausalSim consistently improved simulation accuracy, resulting in algorithms that made about half as many errors as those designed using baseline methods.

      In the future, the researchers want to apply CausalSim to situations where randomized control trial data are not available or where it is especially difficult to recover the causal dynamics of the system. They also want to explore how to design and monitor systems to make them more amenable to causal analysis. More

    • 100 Shares199 Views

      in Data Management & Statistics

      Drones navigate unseen environments with liquid neural networks

      by

      In the vast, expansive skies where birds once ruled supreme, a new crop of aviators is taking flight. These pioneers of the air are not living creatures, but rather a product of deliberate innovation: drones. But these aren’t your typical flying bots, humming around like mechanical bees. Rather, they’re avian-inspired marvels that soar through the sky, guided by liquid neural networks to navigate ever-changing and unseen environments with precision and ease.

      Inspired by the adaptable nature of organic brains, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) have introduced a method for robust flight navigation agents to master vision-based fly-to-target tasks in intricate, unfamiliar environments. The liquid neural networks, which can continuously adapt to new data inputs, showed prowess in making reliable decisions in unknown domains like forests, urban landscapes, and environments with added noise, rotation, and occlusion. These adaptable models, which outperformed many state-of-the-art counterparts in navigation tasks, could enable potential real-world drone applications like search and rescue, delivery, and wildlife monitoring.

      The researchers’ recent study, published today in Science Robotics, details how this new breed of agents can adapt to significant distribution shifts, a long-standing challenge in the field. The team’s new class of machine-learning algorithms, however, captures the causal structure of tasks from high-dimensional, unstructured data, such as pixel inputs from a drone-mounted camera. These networks can then extract crucial aspects of a task (i.e., understand the task at hand) and ignore irrelevant features, allowing acquired navigation skills to transfer targets seamlessly to new environments.

      Play video

      Drones navigate unseen environments with liquid neural networks.

      “We are thrilled by the immense potential of our learning-based control approach for robots, as it lays the groundwork for solving problems that arise when training in one environment and deploying in a completely distinct environment without additional training,” says Daniela Rus, CSAIL director and the Andrew (1956) and Erna Viterbi Professor of Electrical Engineering and Computer Science at MIT. “Our experiments demonstrate that we can effectively teach a drone to locate an object in a forest during summer, and then deploy the model in winter, with vastly different surroundings, or even in urban settings, with varied tasks such as seeking and following. This adaptability is made possible by the causal underpinnings of our solutions. These flexible algorithms could one day aid in decision-making based on data streams that change over time, such as medical diagnosis and autonomous driving applications.”

      A daunting challenge was at the forefront: Do machine-learning systems understand the task they are given from data when flying drones to an unlabeled object? And, would they be able to transfer their learned skill and task to new environments with drastic changes in scenery, such as flying from a forest to an urban landscape? What’s more, unlike the remarkable abilities of our biological brains, deep learning systems struggle with capturing causality, frequently over-fitting their training data and failing to adapt to new environments or changing conditions. This is especially troubling for resource-limited embedded systems, like aerial drones, that need to traverse varied environments and respond to obstacles instantaneously. 

      The liquid networks, in contrast, offer promising preliminary indications of their capacity to address this crucial weakness in deep learning systems. The team’s system was first trained on data collected by a human pilot, to see how they transferred learned navigation skills to new environments under drastic changes in scenery and conditions. Unlike traditional neural networks that only learn during the training phase, the liquid neural net’s parameters can change over time, making them not only interpretable, but more resilient to unexpected or noisy data. 

      In a series of quadrotor closed-loop control experiments, the drones underwent range tests, stress tests, target rotation and occlusion, hiking with adversaries, triangular loops between objects, and dynamic target tracking. They tracked moving targets, and executed multi-step loops between objects in never-before-seen environments, surpassing performance of other cutting-edge counterparts. 

      The team believes that the ability to learn from limited expert data and understand a given task while generalizing to new environments could make autonomous drone deployment more efficient, cost-effective, and reliable. Liquid neural networks, they noted, could enable autonomous air mobility drones to be used for environmental monitoring, package delivery, autonomous vehicles, and robotic assistants. 

      “The experimental setup presented in our work tests the reasoning capabilities of various deep learning systems in controlled and straightforward scenarios,” says MIT CSAIL Research Affiliate Ramin Hasani. “There is still so much room left for future research and development on more complex reasoning challenges for AI systems in autonomous navigation applications, which has to be tested before we can safely deploy them in our society.”

      “Robust learning and performance in out-of-distribution tasks and scenarios are some of the key problems that machine learning and autonomous robotic systems have to conquer to make further inroads in society-critical applications,” says Alessio Lomuscio, professor of AI safety in the Department of Computing at Imperial College London. “In this context, the performance of liquid neural networks, a novel brain-inspired paradigm developed by the authors at MIT, reported in this study is remarkable. If these results are confirmed in other experiments, the paradigm here developed will contribute to making AI and robotic systems more reliable, robust, and efficient.”

      Clearly, the sky is no longer the limit, but rather a vast playground for the boundless possibilities of these airborne marvels. 

      Hasani and PhD student Makram Chahine; Patrick Kao ’22, MEng ’22; and PhD student Aaron Ray SM ’21 wrote the paper with Ryan Shubert ’20, MEng ’22; MIT postdocs Mathias Lechner and Alexander Amini; and Rus.

      This research was supported, in part, by Schmidt Futures, the U.S. Air Force Research Laboratory, the U.S. Air Force Artificial Intelligence Accelerator, and the Boeing Co. More

    • 63 Shares189 Views

      in Data Management & Statistics

      Learning to grow machine-learning models

      by

      It’s no secret that OpenAI’s ChatGPT has some incredible capabilities — for instance, the chatbot can write poetry that resembles Shakespearean sonnets or debug code for a computer program. These abilities are made possible by the massive machine-learning model that ChatGPT is built upon. Researchers have found that when these types of models become large enough, extraordinary capabilities emerge.

      But bigger models also require more time and money to train. The training process involves showing hundreds of billions of examples to a model. Gathering so much data is an involved process in itself. Then come the monetary and environmental costs of running many powerful computers for days or weeks to train a model that may have billions of parameters. 

      “It’s been estimated that training models at the scale of what ChatGPT is hypothesized to run on could take millions of dollars, just for a single training run. Can we improve the efficiency of these training methods, so we can still get good models in less time and for less money? We propose to do this by leveraging smaller language models that have previously been trained,” says Yoon Kim, an assistant professor in MIT’s Department of Electrical Engineering and Computer Science and a member of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      Rather than discarding a previous version of a model, Kim and his collaborators use it as the building blocks for a new model. Using machine learning, their method learns to “grow” a larger model from a smaller model in a way that encodes knowledge the smaller model has already gained. This enables faster training of the larger model.

      Their technique saves about 50 percent of the computational cost required to train a large model, compared to methods that train a new model from scratch. Plus, the models trained using the MIT method performed as well as, or better than, models trained with other techniques that also use smaller models to enable faster training of larger models.

      Reducing the time it takes to train huge models could help researchers make advancements faster with less expense, while also reducing the carbon emissions generated during the training process. It could also enable smaller research groups to work with these massive models, potentially opening the door to many new advances.

      “As we look to democratize these types of technologies, making training faster and less expensive will become more important,” says Kim, senior author of a paper on this technique.

      Kim and his graduate student Lucas Torroba Hennigen wrote the paper with lead author Peihao Wang, a graduate student at the University of Texas at Austin, as well as others at the MIT-IBM Watson AI Lab and Columbia University. The research will be presented at the International Conference on Learning Representations.

      The bigger the better

      Large language models like GPT-3, which is at the core of ChatGPT, are built using a neural network architecture called a transformer. A neural network, loosely based on the human brain, is composed of layers of interconnected nodes, or “neurons.” Each neuron contains parameters, which are variables learned during the training process that the neuron uses to process data.

      Transformer architectures are unique because, as these types of neural network models get bigger, they achieve much better results.

      “This has led to an arms race of companies trying to train larger and larger transformers on larger and larger datasets. More so than other architectures, it seems that transformer networks get much better with scaling. We’re just not exactly sure why this is the case,” Kim says.

      These models often have hundreds of millions or billions of learnable parameters. Training all these parameters from scratch is expensive, so researchers seek to accelerate the process.

      One effective technique is known as model growth. Using the model growth method, researchers can increase the size of a transformer by copying neurons, or even entire layers of a previous version of the network, then stacking them on top. They can make a network wider by adding new neurons to a layer or make it deeper by adding additional layers of neurons.

      In contrast to previous approaches for model growth, parameters associated with the new neurons in the expanded transformer are not just copies of the smaller network’s parameters, Kim explains. Rather, they are learned combinations of the parameters of the smaller model.

      Learning to grow

      Kim and his collaborators use machine learning to learn a linear mapping of the parameters of the smaller model. This linear map is a mathematical operation that transforms a set of input values, in this case the smaller model’s parameters, to a set of output values, in this case the parameters of the larger model.

      Their method, which they call a learned Linear Growth Operator (LiGO), learns to expand the width and depth of larger network from the parameters of a smaller network in a data-driven way.

      But the smaller model may actually be quite large — perhaps it has a hundred million parameters — and researchers might want to make a model with a billion parameters. So the LiGO technique breaks the linear map into smaller pieces that a machine-learning algorithm can handle.

      LiGO also expands width and depth simultaneously, which makes it more efficient than other methods. A user can tune how wide and deep they want the larger model to be when they input the smaller model and its parameters, Kim explains.

      When they compared their technique to the process of training a new model from scratch, as well as to model-growth methods, it was faster than all the baselines. Their method saves about 50 percent of the computational costs required to train both vision and language models, while often improving performance.

      The researchers also found they could use LiGO to accelerate transformer training even when they didn’t have access to a smaller, pretrained model.

      “I was surprised by how much better all the methods, including ours, did compared to the random initialization, train-from-scratch baselines.” Kim says.

      In the future, Kim and his collaborators are looking forward to applying LiGO to even larger models.

      The work was funded, in part, by the MIT-IBM Watson AI Lab, Amazon, the IBM Research AI Hardware Center, Center for Computational Innovation at Rensselaer Polytechnic Institute, and the U.S. Army Research Office. More

    • 100 Shares109 Views

      in Data Management & Statistics

      New method accelerates data retrieval in huge databases

      by

      Hashing is a core operation in most online databases, like a library catalogue or an e-commerce website. A hash function generates codes that directly determine the location where data would be stored. So, using these codes, it is easier to find and retrieve the data.

      However, because traditional hash functions generate codes randomly, sometimes two pieces of data can be hashed with the same value. This causes collisions — when searching for one item points a user to many pieces of data with the same hash value. It takes much longer to find the right one, resulting in slower searches and reduced performance.

      Certain types of hash functions, known as perfect hash functions, are designed to place the data in a way that prevents collisions. But they are time-consuming to construct for each dataset and take more time to compute than traditional hash functions.

      Since hashing is used in so many applications, from database indexing to data compression to cryptography, fast and efficient hash functions are critical. So, researchers from MIT and elsewhere set out to see if they could use machine learning to build better hash functions.

      They found that, in certain situations, using learned models instead of traditional hash functions could result in half as many collisions. These learned models are created by running a machine-learning algorithm on a dataset to capture specific characteristics. The team’s experiments also showed that learned models were often more computationally efficient than perfect hash functions.

      “What we found in this work is that in some situations we can come up with a better tradeoff between the computation of the hash function and the collisions we will face. In these situations, the computation time for the hash function can be increased a bit, but at the same time its collisions can be reduced very significantly,” says Ibrahim Sabek, a postdoc in the MIT Data Systems Group of the Computer Science and Artificial Intelligence Laboratory (CSAIL).

      Their research, which will be presented at the 2023 International Conference on Very Large Databases, demonstrates how a hash function can be designed to significantly speed up searches in a huge database. For instance, their technique could accelerate computational systems that scientists use to store and analyze DNA, amino acid sequences, or other biological information.

      Sabek is the co-lead author of the paper with Department of Electrical Engineering and Computer Science (EECS) graduate student Kapil Vaidya. They are joined by co-authors Dominick Horn, a graduate student at the Technical University of Munich; Andreas Kipf, an MIT postdoc; Michael Mitzenmacher, professor of computer science at the Harvard John A. Paulson School of Engineering and Applied Sciences; and senior author Tim Kraska, associate professor of EECS at MIT and co-director of the Data, Systems, and AI Lab.

      Hashing it out

      Given a data input, or key, a traditional hash function generates a random number, or code, that corresponds to the slot where that key will be stored. To use a simple example, if there are 10 keys to be put into 10 slots, the function would generate a random integer between 1 and 10 for each input. It is highly probable that two keys will end up in the same slot, causing collisions.

      Perfect hash functions provide a collision-free alternative. Researchers give the function some extra knowledge, such as the number of slots the data are to be placed into. Then it can perform additional computations to figure out where to put each key to avoid collisions. However, these added computations make the function harder to create and less efficient.

      “We were wondering, if we know more about the data — that it will come from a particular distribution — can we use learned models to build a hash function that can actually reduce collisions?” Vaidya says.

      A data distribution shows all possible values in a dataset, and how often each value occurs. The distribution can be used to calculate the probability that a particular value is in a data sample.

      The researchers took a small sample from a dataset and used machine learning to approximate the shape of the data’s distribution, or how the data are spread out. The learned model then uses the approximation to predict the location of a key in the dataset.

      They found that learned models were easier to build and faster to run than perfect hash functions and that they led to fewer collisions than traditional hash functions if data are distributed in a predictable way. But if the data are not predictably distributed because gaps between data points vary too widely, using learned models might cause more collisions.

      “We may have a huge number of data inputs, and the gaps between consecutive inputs are very different, so learning a model to capture the data distribution of these inputs is quite difficult,” Sabek explains.

      Fewer collisions, faster results

      When data were predictably distributed, learned models could reduce the ratio of colliding keys in a dataset from 30 percent to 15 percent, compared with traditional hash functions. They were also able to achieve better throughput than perfect hash functions. In the best cases, learned models reduced the runtime by nearly 30 percent.

      As they explored the use of learned models for hashing, the researchers also found that throughput was impacted most by the number of sub-models. Each learned model is composed of smaller linear models that approximate the data distribution for different parts of the data. With more sub-models, the learned model produces a more accurate approximation, but it takes more time.

      “At a certain threshold of sub-models, you get enough information to build the approximation that you need for the hash function. But after that, it won’t lead to more improvement in collision reduction,” Sabek says.

      Building off this analysis, the researchers want to use learned models to design hash functions for other types of data. They also plan to explore learned hashing for databases in which data can be inserted or deleted. When data are updated in this way, the model needs to change accordingly, but changing the model while maintaining accuracy is a difficult problem.

      “We want to encourage the community to use machine learning inside more fundamental data structures and algorithms. Any kind of core data structure presents us with an opportunity to use machine learning to capture data properties and get better performance. There is still a lot we can explore,” Sabek says.

      “Hashing and indexing functions are core to a lot of database functionality. Given the variety of users and use cases, there is no one size fits all hashing, and learned models help adapt the database to a specific user. This paper is a great balanced analysis of the feasibility of these new techniques and does a good job of talking rigorously about the pros and cons, and helps us build our understanding of when such methods can be expected to work well,” says Murali Narayanaswamy, a principal machine learning scientist at Amazon, who was not involved with this work. “Exploring these kinds of enhancements is an exciting area of research both in academia and industry, and the kind of rigor shown in this work is critical for these methods to have large impact.”

      This work was supported, in part, by Google, Intel, Microsoft, the U.S. National Science Foundation, the U.S. Air Force Research Laboratory, and the U.S. Air Force Artificial Intelligence Accelerator. More