Events

December 04, 2023

Quest | CBMM Seminar Series: Incomplete Objectives and AI Safety: The Theory and Practice of AI Alignment

Dylan Hadfield-Menell

CSAIL MIT

4:00P

- 5:30P

Location

Singleton Auditorium (46-3002)

Singleton Auditorium (46-3002)

Add to Calendar 2023-12-04 16:00:00 2023-12-04 17:30:00 America/New_York Quest | CBMM Seminar Series: Incomplete Objectives and AI Safety: The Theory and Practice of AI Alignment Abstract: For AI systems to be safe and effective, they need to be aligned with the goals and values of users, designers, and society. In this talk, I will discuss the challenges of AI alignment and go over research directions to develop safe AI systems. I'll begin with theoretical results that motivate the alignment problem broadly. In particular, I will show how optimizing incomplete goal specifications reliably causes systems to select unhelpful or harmful actions. Next, I will discuss mitigation measures that counteract this failure mode. I will focus on approaches for incorporating human feedback into objectives, interpreting and understanding learned policies, and maintaining uncertainty about intended goals. Singleton Auditorium (46-3002)

December 05, 2023

High-Performance GPU Code Generation for Mining Motifs in Temporal Graphs

Nishil Talati

University of Michigan

2:00P

- 3:00P

Location

32-D463

32-D463

Add to Calendar 2023-12-05 14:00:00 2023-12-05 15:00:00 America/New_York High-Performance GPU Code Generation for Mining Motifs in Temporal Graphs This is hybrid meeting. The physical room is 32-D463 (Star). The Zoom registration link is https://mit.zoom.us/meeting/register/tJMtdOqgrTwsGNV_k0Nk6JjLMk-G6GFzTRyk.******************IMPORTANT NOTE ABOUT ONLINE REGISTRATION******************- The registration link for 2023 is the same as the link from 2022.- Please save the Zoom link that you receive after you register. This link will stay the same for all subsequent Fast Code seminars.- Zoom does not recognize a second registration, and will not send out the link a second time. The organizers will not be notified of any second registration.- If you have any problems with registration, please contact lindalynch@csail.mit.edu by 12pm on the day of the seminar, so that we can try to resolve it before the seminar begins.*********************************************************************Abstract:Temporal motif mining is the task of finding the occurrences of subgraph patterns withina large input temporal graph that obey the specified structural and temporal constraints.Despite its utility in several critical application domains that demand high performance(e.g., detecting fraud in financial transaction graphs), the performance of existing softwareis limited on commercial hard- ware platforms, in that it runs for tens of hours. In this talk,I will present Everest - a system that efficiently maps the workload of mining (supportsboth enumeration and counting) temporal motifs to the highly parallel GPU architecture.Using input temporal graph and a more expressive user-defined temporal motif querydefinition, Everest generates an execution plan and runtime primitives that optimize theworkload execution by exploiting the high compute throughput of a GPU. Everestgenerates motif-specific mining code to reduce long-latency memory accesses andfrequent thread divergence operations. Everest incorporates novel low-cost runtimemechanisms to enable load balancing to improve GPU hardware utilization. To supportlarge graphs that do not fit on GPU memory, Everest also supports multi-GPU executionby intelligently partitioning the edge list that prevents inter-GPU communication. Everesthides the implementation complexity of presented optimizations away from the targetedsystem user for better usability. Our evaluation shows that, using proposed optimizations,Everest improves the performance of a baseline GPU implementation by 19x, on average.Speaker Bio:Nishil Talati is an Assistant Research Scientist (Research Faculty) at the CSE departmentof University of Michigan. He earned his PhD from University of Michigan. Nishil’sresearch interests include computer architecture and systems software design forimproving the performance of modern data-intensive workloads. His research is publishedat several top-tier venues including ISCA, MICRO, HPCA, ASPLOS, and others. Nishil’swork has been recognized as the 2021 HPCA best paper award, 2023 DATE best paperhonorable mention, and 2023 IISWC best paper nominee.Relevant Publication:Y. Yuan, H. Ye, S. Vedula, W. Kaza, and N. Talati, "Everest: GPU-Accelerated System for Mining TemporalMotifs," at International Conference on Very Large Databases (VLDB 2024). 32-D463

Thesis Defense: Albert Kim, Title: Optimizing Queries with Disjunctions

Albert Kim

CSAIL MIT

3:00P

- 5:00P

Location

Seminar Room G882 (Hewlett Room)

Add to Calendar 2023-12-05 15:00:00 2023-12-05 17:00:00 America/New_York Thesis Defense: Albert Kim, Title: Optimizing Queries with Disjunctions Abstract: Despite decades of research into query optimization, optimizing queries with disjunctive predicate expressions remains a challenge. Solutions employed by existing systems (if any) are often simplistic and lead to much redundant work being performed by the execution engine. In this thesis, we present a two-pronged approach to address this problem: (1) For single-table queries, we provide a formal analysis of the problem for column-oriented engines, and based on our analysis, we present several algorithms capable of generating optimal predicate evaluation plans. Our algorithms are polynomial-time (unlike existing works), theoretically proven to produce optimal plans, and can be implemented in existing systems with minimal effort. (2) For multi-table queries, we present a novel form of query execution called tagged execution, which groups tuples into subrelations based on the predicates they satisfy and tags them with that information. Query operators can take advantage of the additional context provided by tags during runtime, and this allows them to eliminate much of the redundant work performed by traditional engines and even achieve pushdown optimizations for disjunctive predicates. Both our approaches greatly outperform existing solutions, sometimes by orders of magnitude, and we believe the combination of a theoretical approach (1) and a practical approach (2) helps meet the needs of most users. Seminar Room G882 (Hewlett Room)

Bagging is an Optimal PAC Learner

Kasper Green Larsen, Asst. Professor, Dept. of Computer Science, Aarhus University

Aarhus University

4:00P

- 5:15P

Location

G449 KIVA

G449 KIVA

Add to Calendar 2023-12-05 16:00:00 2023-12-05 17:15:00 America/New_York Bagging is an Optimal PAC Learner Abstract:Determining the optimal sample complexity of PAC learning in the realizable setting was a central open problem in learning theory for decades. Finally, the seminal work by Hanneke (2016) gave an algorithm with a provably optimal sample complexity. His algorithm is based on a careful and structured sub-sampling of the training data and then returning a majority vote among hypotheses trained on each of the sub-samples. While being a very exciting theoretical result, it has not had much impact in practice, in part due to inefficiency, since it constructs a polynomial number of sub-samples of the training data, each of linear size.In this talk, we prove the surprising result that the practical and classic heuristic Bagging (a.k.a. bootstrap aggregation), due to Breiman (1996), is in fact also an optimal PAC learner. Bagging pre-dates Hanneke's algorithm by twenty years and is taught in most undergraduate machine learning courses. Moreover, we show that it only requires a logarithmic number of sub-samples to reach optimality.This work was published at COLT’23 and received the Best Paper Award. G449 KIVA

Quest | CBMM Seminar Series: Statistical learning in human sensorimotor control

Daniel Wolpert

Columbia University

4:00P

- 5:30P

Location

Singleton Auditorium (46-3002)

Add to Calendar 2023-12-05 16:00:00 2023-12-05 17:30:00 America/New_York Quest | CBMM Seminar Series: Statistical learning in human sensorimotor control Abstract: Humans spend a lifetime learning, storing and refining a repertoire of motor memories appropriate for the multitude of tasks we perform. However, it is unknown what principle underlies the way our continuous stream of sensorimotor experience is segmented into separate memories and how we adapt and use this growing repertoire. I will review our recent work on how humans learn to make skilled movements focusing on how statistical learning can lead to multi-modal object representations, how we represent the dynamics of objects, the role of context in the expression, updating and creation of motor memories and how families of objects are learned. Bio: Daniel Wolpert FMedSci FRS. Daniel qualified as a medical doctor in 1989. He worked with John Stein and Chris Miall in the Physiology Department of Oxford University where he received his D.Phil. in 1992. He worked as a postdoctoral fellow in the Department of Brain and Cognitive Sciences at MIT in Mike Jordan's group and in 1995 joined the Sobell Department of Motor Neuroscience, Institute of Neurology as a Lecturer. In 2005 moved to the University of Cambridge where he was Professor of Engineering (1875) and a fellow of Trinity College and from 2013 the Royal Society Noreen Murray Research Professorship in Neurobiology. In 2018 Daniel joined the Zuckerman Mind Brain and Behavior Institute at Columbia University as Professor of Neuroscience and is vice-chair of the Department of Neuroscience. Daniel retains a part-time position as Director of Research at the Department of Engineering, University of Cambridge.He was elected a Fellow of the Academy of Medical Sciences in 2004 and a Fellow of the Royal Society in 2012.He was awarded the Royal Society Francis Crick Prize Lecture (2005), the Minerva Foundation Golden Brain Award (2010), the Royal Society Ferrier Medal (2020) and gave the Fred Kavli Distinguished International Scientist Lecture at the Society for Neuroscience (2009).

HCI Seminar - Nadya Peek - Machine Agency: Toolkits for Creative Automation

Nadya Peek

University of Washington

4:00P

- 5:00P

Location

32-D463 (Star)

Add to Calendar 2023-12-05 16:00:00 2023-12-05 17:00:00 America/New_York HCI Seminar - Nadya Peek - Machine Agency: Toolkits for Creative Automation Abstract:How can we harness the precision of machines for the creativity of individuals? Robotics gives us access to precision and repeatability, but there is a high threshold to automating. Domain experts such as ceramicists, plant biologists, wood workers, or chemical engineers have extensive knowledge of intricate processes and workflows, but are not also experts in control systems or programming. Because of this, highly skilled individuals conduct thousands of hours of manual work to support their goals. In my research, I develop end-user systems to lower the threshold to automation without loss of complexity. In particular, I develop low-cost and modular open-source toolkits for domain experts to build and customize while iterating on and refining their workflows. Cost forms only one barrier—the main problem is that robots for many workflows simply do not exist. Not all niche applications form a viable market segment. I show how using rapid prototyping equipment such as 3D printers is a viable strategy for domain experts to build application-specific machines and robots. In this talk, I will show example machines, workflows, and processes we are developing to support machine agency.Bio:Nadya Peek develops unconventional digital fabrication tools, small scale automation, networked controls, and advanced manufacturing systems. Spanning electronics, firmware, software, and mechanics, her research focuses on harnessing the precision of machines for the creativity of individuals. Nadya directs the Machine Agency at the University of Washington where she is an assistant professor in Human-Centered Design and Engineering. Machines and systems Nadya has built have been shared widely, including at the White House Office of Science and Technology Policy, the World Economic Forum, TED, and many Maker Faires and outreach events. Her research has been supported by the National Science Foundation, the Alfred P. Sloan Foundation, and the Gordon and Betty Moore Foundation, and her teaching has been recognized with the University of Washington's Distinguished Teaching Award for Innovation with Technology. She received the MIT Technology Review's 35 under 35 award in 2020. Nadya is an active member of the global fab lab community, making digital fabrication more accessible with better CAD/CAM tools and developing open source hardware machines and control systems. She is on the board of the Open Source Hardware Association, the editor in chief of the Journal of Open Hardware, half of the design studio James and the Giant Peek, plays drum machines and synths in the band Construction, and got her PhD at MIT in the Center for Bits and Atoms.This talk will also be streamed over Zoom: https://mit.zoom.us/j/97928948160. 32-D463 (Star)

December 06, 2023

Using machine learning to increase equity in healthcare and public health.

Emma Pierson

Cornell University

11:30A

- 1:00P

Location

32 G-575

Add to Calendar 2023-12-06 11:30:00 2023-12-06 13:00:00 America/New_York Using machine learning to increase equity in healthcare and public health. Our society remains profoundly unequal. This talk discusses how data science and machine learning can be used to combat inequality in health care and public health by presenting several vignettes from domains like policing, women's health, and cancer risk prediction.Bio: Emma Pierson is an assistant professor of computer science at the Jacobs Technion-Cornell Institute at Cornell Tech and the Technion, and a computer science field member at Cornell University. She holds a secondary joint appointment as an Assistant Professor of Population Health Sciences at Weill Cornell Medical College. She develops data science and machine learning methods to study inequality and healthcare. Her work has been recognized by best paper, poster, and talk awards, an NSF CAREER award, a Rhodes Scholarship, Hertz Fellowship, Rising Star in EECS, MIT Technology Review 35 Innovators Under 35, and Forbes 30 Under 30 in Science. Her research has been published at venues including ICML, KDD, WWW, Nature, and Nature Medicine, and she has also written for The New York Times, FiveThirtyEight, Wired, and various other publications.Zoom link: https://mit.zoom.us/j/93513735220Location: 32 G-575Refreshments will be available 32 G-575

Cloud-Native Database with Storage Disaggregation

Xiangyao Yu

University of Wisconsin - Madison

1:00P

- 2:00P

Location

32-G575

Add to Calendar 2023-12-06 13:00:00 2023-12-06 14:00:00 America/New_York Cloud-Native Database with Storage Disaggregation itle=====Cloud-Native Database with Storage DisaggregationAbstract========Modern databases are moving to the cloud for high elasticity and flexible pricing model (e.g., pay-as-you-go). Cloud-native databases adopt a unique storage-disaggregation architecture, where the computation and storage are decoupled as two separate services connected through the data center network. The new storage-disaggregation architecture demands a revisit of database system design. In this talk, I will discuss our research in optimized disaggregated database systems in both transactional and analytical workloads. In particular, I will present Cornus, an optimized two-phase commit (2PC) protocol that leverages highly-available storage services to reduce latency and enhance reliability for distributed transactions. I will also discuss our recent work on computation pushdown, where the database leverages storage-layer computation (i.e., a pushdown layer) to accelerate data analytics processing. I will conclude the talk with our on-going research and future plans on cloud-native database systems.Short Bio=========Xiangyao Yu is an Assistant Professor at the University of Wisconsin-Madison. His research interests include (1) cloud-native databases, (2) new hardware for databases, and (3) transactions and HTAP. Before joining UW-Madison, he finished postdoc and PhD at MIT and bachelor at Tsinghua University.Zoom option:Tianyu Li is inviting you to a scheduled Zoom meeting.Join Zoom Meetinghttps://mit.zoom.us/j/92698683855?pwd=NWhsVjdVblRvd3g5U0p3VDRIcDJTQT09Password: 329527One tap mobile+16465588656,,92698683855# US (New York)+16699006833,,92698683855# US (San Jose)Meeting ID: 926 9868 3855US : +1 646 558 8656 or +1 669 900 6833International Numbers: https://mit.zoom.us/u/acWjKZOLZiJoin by SIP92698683855@zoomcrc.comJoin by Skype for Businesshttps://mit.zoom.us/skype/92698683855 32-G575

Coin Sampling: Gradient-Based Bayesian Inference without Learning Rates

Louis Sharrock

Lancaster University, UK

2:00P

- 3:00P

Location

Room 32-D507

Add to Calendar 2023-12-06 14:00:00 2023-12-06 15:00:00 America/New_York Coin Sampling: Gradient-Based Bayesian Inference without Learning Rates Abstract: In recent years, particle-based variational inference (ParVI) methods such as Stein variational gradient descent (SVGD) have grown in popularity as scalable methods for Bayesian inference. Unfortunately, the properties of such methods invariably depend on hyperparameters such as the learning rate, which must be carefully tuned by the practitioner in order to ensure convergence to the target measure at a suitable rate. In this work, we introduce a suite of new particle-based methods for scalable Bayesian inference based on coin betting, which are entirely learning-rate free. We illustrate the performance of our approach on a range of numerical examples, including several high-dimensional models and datasets, demonstrating comparable performance to other ParVI algorithms with no need to tune a learning rate. Bio: Louis Sharrock is a Senior Research Associate in Statistical Machine Learning at Lancaster University, UK. Prior to this, he obtained a PhD in Statistics at Imperial College London, and a MA in Mathematics at the University of Cambridge. His research interests include computational statistics, machine learning, and optimisation, with a particular focus on the development and analysis of scalable methods for inference in complex statistical models. Room 32-D507

CSL Seminar - Yi Wu - Language Model Meets Reinforcement Learning: Building Strong Language Agents for Strategic Gameplay

Yi Wu

Tsinghua University

2:00P

- 3:00P

Location

Seminar Room G575

Add to Calendar 2023-12-06 14:00:00 2023-12-06 15:00:00 America/New_York CSL Seminar - Yi Wu - Language Model Meets Reinforcement Learning: Building Strong Language Agents for Strategic Gameplay Computational Sensorimotor Learning (CSL) Seminar1-on-1 meeting sign-up: https://docs.google.com/spreadsheets/d/1riA5Bg72BAaZNokW9JAdHyM8fnwxbK0fhjgRHOm8lyI/edit?usp=sharingZoom: https://mit.zoom.us/j/96702680180?pwd=cXNJTmh2QmdLc2JENDFOWWU5OTc0dz09Abstract:Thanks to the advances in large language models (LLM), there has been a recent trend to develop intelligent language agents for complex tasks. Most existing applications of language agents are purely LLM-based, i.e., by directly prompting LLMs to output actions. Although interesting emergent behaviors can be observed, their performances in complex multi-agent games can be limited due to the lack of domain-specific training.This talk will cover some recent projects in my group on developing language agents that can both yield strong gameplay performances and cooperate with real human players in challenging multi-agent games. The key idea is to combine language modeling and reinforcement learning. The language model will serve as an interface for reasoning and interpreting high-level commands while reinforcement learning helps substantially improve the gameplay performance of the agent. We demonstrate our agents in three domains, including an agent that can follow high-level commands to play a real-time strategy game, an Overcooked agent that can cooperate with humans via languages to cook dishes, and an agent that outperforms average human players in the Werewolf game. Bio:Yi Wu is now an assistant professor at the Institute for Interdisciplinary Information Sciences (IIIS) at Tsinghua University. He obtained his Ph.D. degree from UC Berkeley in 2019 under the supervision of Prof. Stuart Russell. Before moving back to Tsinghua, Yi was a full-time researcher at OpenAI. His research focuses on improving the generalization capabilities of learning agents. He is broadly interested in a variety of topics in AI, including multi-agent reinforcement learning, human-AI interaction, language grounding, and robot learning. His representative works include the MADDPG/MAPPO algorithm, OpenAI's hide-and-seek project, and the value iteration network, which won the best paper award at NIPS 2016. Yi's Website: https://jxwuyi.weebly.com/ Seminar Room G575

December 08, 2023

Aleks Petrov: When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations

Aleks Petrov

11:00A

- 12:00P

Location

G449 (KIVA)

Add to Calendar 2023-12-08 11:00:00 2023-12-08 12:00:00 America/New_York Aleks Petrov: When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations Abstract:Large Language Models (LLMs), typically based on transformer architecture, have undergone extensive research aimed at enhancing their scalability and performance. Despite the technological advancements, the resource-intensive nature of training state-of-the-art transformer models limits their accessibility to a broader spectrum of researchers and smaller enterprises. As a result, context-based fine-tuning methods, including prompting, in-context learning, soft prompting (also known as prompt tuning), and prefix-tuning, have gained popularity due to their ability to often match the performance of full fine-tuning with a fraction of the parameters. However, despite their empirical successes, there is little theoretical understanding of how these techniques influence the internal computation of the model and their expressiveness limitations. Our research presents the first theoretical framework for context-based fine-tuning methods. We establish that these methods cannot alter the attention patterns over the content; they can only skew the outputs of an attention layer in a predetermined direction. This inherent limitation restricts the scope of what can be achieved through context-based fine-tuning. While these methods can effectively prompt a model to manifest a skill it has already acquired, combining pre-existing skills presents a more complex challenge and is harder to optimise. Crucially, our analysis demonstrates that context-based fine-tuning methods, including prompting, are incapable of learning entirely new behaviours by the model. G449 (KIVA)

Nikola Konstantinov: Incentivizing Collaboration in Federated Learning

Nikola Konstantinov, INSAIT

1:00P

- 2:30P

Location

D463 (STAR)

Add to Calendar 2023-12-08 13:00:00 2023-12-08 14:30:00 America/New_York Nikola Konstantinov: Incentivizing Collaboration in Federated Learning Abstract: Collaborative and federated learning techniques have the potential to enable training powerful machine learning models from distributed data. However, in many cases the potential participants in such collaborative schemes have additional incentives to end-model accuracy. For example, they may be competitors on downstream tasks or may be concerned about the privacy of their data. This creates incentives for individual participants to obfuscate their messages or even damage the training process for the other players, which often undermines the benefits of collaboration. In this talk I will present recent results regarding participation incentives in federated learning (FL). In particular, I will cover several theoretical models for rational data-sharing decision making in the context of market competition and privacy concerns. I will also show how to design FL protocols that provably incentivize honesty during training, even in the presence of conflicting incentives. D463 (STAR)

December 11, 2023

Ross Anderson - Crypto War 3 – The DMA, Chatcontrol and now Patchcontrol?

Ross Anderson

Cambridge University

2:00P

- 3:00P

Location

32-G449 (Kiva)

Add to Calendar 2023-12-11 14:00:00 2023-12-11 15:00:00 America/New_York Ross Anderson - Crypto War 3 – The DMA, Chatcontrol and now Patchcontrol? Abstract:The intelligence and law-enforcement communities have been tussling with industry for over thirty years to limit the use of strong cryptography. In Crypto War 1, the Clinton administration pushed the Clipper chip, export keylengths and 'trusted third parties'. After 9/11, illegal wiretapping programs extended to large-scale intercept at mail servers; Ed Snowden exposed this in 2013. Industry fought back with end-to-end encryption, which the agencies are now working to undermine using a variety of strategies. In this third crypto war, the weapons are mandated interoperability of messaging systems; client-side scanning and other mandated "safety" tech; and even a power to demand official approval for security enhancements and bug fixes. I will describe how these tussles have been working out in the EU and the UK.Population-scale text scanning cannot be effective for the claimed purposes, as the false alarms would swamp the police. It could also not be legal as it would contravene the European courts' ban on bulk surveillance without warrant or suspicion. And crimes of violence against children mostly occur in the family. They are associated with violent crime against women and with misogyny in general. Both require an entirely different and local response involving police, social workers, teachers, and family members. And the most effective means of detecting abuse online is by making it easier for users to report it.Bio:Ross Anderson is Professor of Security Engineering at the universities of Cambridge and Edinburgh. He is a Fellow of the Royal Society and the Royal Academy of Engineering, and won the Lovelace Medal, Britain's top award in computing. 32-G449 (Kiva)

December 13, 2023

Learning complex functional constraints in proteins and non-coding DNA

Yun Song

Berkeley

11:30A

- 1:00P

Location

32 G-575

32 G-575

Add to Calendar 2023-12-13 11:30:00 2023-12-13 13:00:00 America/New_York Learning complex functional constraints in proteins and non-coding DNA Abstract: Predicting the impact of genetic variants is a significant challenge in computational biology, with crucial applications in disease diagnosis, gene regulation modeling, and protein engineering. In this talk, I will describe my lab's recent work on improving variant effect prediction for both coding and non-coding regions by leveraging advances in unsupervised learning, particularly self-supervised learning from natural language processing. For coding variants, I will introduce a robust learning framework to transfer properties between unrelated proteins and discuss how this approach fares in comparison with the recently published PrimateAI-3D and AlphaMissense methods. Regarding non-coding variants, I will present our work on DNA language models, highlighting their efficacy in genome-wide variant effect prediction.Zoom link: https://mit.zoom.us/j/93513735220Location: 32 G-575Refreshments will be available 32 G-575

Deep Image Analysis Based on Deformable Shapes and Its Application in Neuroimaging

Jian Wang

Harvard Medical School

2:00P

- 3:00P

Location

32-D451

Add to Calendar 2023-12-13 14:00:00 2023-12-13 15:00:00 America/New_York Deep Image Analysis Based on Deformable Shapes and Its Application in Neuroimaging Deformable shapes are crucial for various image analysis tasks, e.g., image registration for real-time image-guided navigation systems of tumor removal surgery, image classification for neuro-degenerative diseases, and template-based image segmentation. Although recent advances in deep learning-based image analysis have achieved groundbreaking performance by providing a universal mechanism to extract image features in the context of texture, intensity, or simple geometry features, they fall short in capturing much more complex and detailed geometric information behind the image data. This greatly hinders the power of image analysis models when analyzing and quantifying geometric shapes are important. The modeling of deformable shapes presents significant challenges due to their high-dimensional and non-linear nature of data. Existing deep learning algorithms suffer from high computational cost of network training and inference, as well as severely decreased model performance caused by broken assumptions of image quality (i.e., missing data, corrupted signals, or occurrence of new objects). To address these challenges, I first developed deep neural networks to learn low-dimensional shape representations based on fine-grained deformations derived from image registration algorithms with much lower computational complexity in training. I then investigated a new paradigm of deep learning models that are capable of analyzing such learned shape representations to improve the current performance of image analysis tasks, including but not limited to population-based image studies. In order to enhance the robustness of shape-based deep networks, I developed geometric metamorphic learning algorithms to properly consider image conditions where missing data, or appearance changes occur. The algorithmic foundation of my research work could potentially impact a variety of real-world clinical applications, including but not limited to automated diagnosis for neurodegenerative diseases (i.e., Alzheimer’s) and real-time image-guided navigation systems for neurosurgery (i.e., brain tumor resection). 32-D451

Characterizing and Optimizing End-to-End Systems for Private Inference

Karthik Garimella

NYU

4:00P

- 5:00P

Location

H-882, Hewlett Room

Add to Calendar 2023-12-13 16:00:00 2023-12-13 17:00:00 America/New_York Characterizing and Optimizing End-to-End Systems for Private Inference In two-party machine learning prediction services, the client's goal is to query a remote server's trained machine learning model to perform neural network inference in some application domain. However, sensitive information can be obtained during this process by either the client or the server, leading to potential collection, unauthorized secondary use, and inappropriate access to personal information. These security concerns have given rise to Private Inference (PI), in which both the client's personal data and the server's trained model are kept confidential. State-of-the-art PI protocols consist of a pre-processing or offline phase and an online phase that combine several cryptographic primitives: Homomorphic Encryption (HE), Secret Sharing (SS), Garbled Circuits (GC), and Oblivious Transfer (OT). Despite the need and recent performance improvements, PI remains largely arcane today and is too slow for practical use.This paper addresses PI's shortcomings with a detailed characterization of a standard high-performance protocol to build foundational knowledge and intuition in the systems community. Our characterization pinpoints all sources of inefficiency -- compute, communication, and storage. In contrast to prior work, we consider inference request arrival rates rather than studying individual inferences in isolation and we find that the pre-processing phase cannot be ignored and is often incurred online as there is insufficient downtime to hide pre-compute latency. Finally, we leverage insights from our characterization and propose three optimizations to address the storage (Client-Garbler), computation (layer-parallel HE), and communication (wireless slot allocation) overheads. H-882, Hewlett Room