USC at NeurIPS 2022 – USC Viterbi

USC at NeurIPS 2022 - USC Viterbi

From extra clear AI to raised MRI expertise and safer robotics, USC Analysis is advancing innovative expertise in machine studying at NeurIPS 2022. Photograph/iStock.

The Neural Information Processing Systems Conference (NeurIPS) It is among the best worldwide venues for machine studying analysis and the most important pool of researchers within the area. At this 12 months’s occasion (November 28 – December 9), USC analysis represents among the groundbreaking work in machine studying, from making AI extra explainable to designing safer robotic techniques and sooner, extra detailed MRI scans.

Open the black field

Sparse interactive additive networks via feature interaction detection and sparse selection

James Inwin (College of Southern California); Yan Liu (College of Southern California)

What’s the matter of your paper? The paper focuses on making AI techniques extra explainable and reliable. Presently, we feed these algorithms a number of information and hope they study the fitting factor. Sadly, they will choose up on the improper issues (biases, correlations, and many others.) and we might haven’t any manner of realizing. Due to this fact, this work is one step in direction of making these AI insights extra interpretable and digestible to permit people to confirm the algorithm.

Who can profit from this analysis? individuals who need to perceive the “algorithm” or “black field”; medical doctors who need to perceive AI techniques as they predict their sufferers; Credit score auditors are stakeholders who’re required to clarify their selections.

Botnet check

Create a deep alternative with the help of generating environments

Varun Bhatt (College of Southern California); Brion Tjanaka (College of Southern California); Matthew Christopher Fontaine (College of Southern California); Stefanos Nikolaidis (College of Southern California)

What’s the matter of your paper? In July 2022, a chess-playing robotic grabbed and broke the finger of a seven-year-old little one as a result of it mistook the finger for a chess piece. The bot appears to imagine that the human opponent is at all times ready between strikes and has not been examined in conditions the place the human makes strikes rapidly. This incident may have been averted if such a situation had occurred I used to be discovered throughout testing. Incidents like these are why we wished to create an environment friendly technique for locating nook circumstances involving robots or different clever brokers. On this paper, we use machine studying to foretell how our bot will behave specifically situation and dramatically velocity up the method of discovering failure situations. This permits us to seek out very complicated situations comparatively rapidly that break down the complicated components.

Who can profit from this analysis? MProducers of robots, self-driving automobiles, and many others. to raised check brokers earlier than deploying them in the true world. Higher testing results in safer bots, which additionally advantages finish customers.

Perceive the impression of faux information and rumours

A neural counterfactual time point process to estimate the causal impact of misinformation on social media

Yizhou Zhang (College of Southern California); Defoe Kao (College of Southern California); Yan Liu (College of Southern California)

What’s the matter of your paper? search helps We higher perceive how pretend information and rumors change folks’s minds. We apply our mannequin to an actual dataset of social media posts and interactions about COVID-19 vaccines. Experimental outcomes point out that our mannequin acknowledged a particular causal impact of misinformation that hurts folks’s subjective emotions about vaccines.

Who can profit from this analysis? Journalists can use our fashions to grasp how pretend information and rumors change folks’s minds after which discover higher methods to take action Reveals Deceptive content material.

Acceleration of MRI scans

HUMUS-Net: an unpatched multiscale hybrid network architecture for accelerated MRI reconstruction

Zlan Fabian (College of Southern California); Burke Tenaz (College of Southern California); Mehdi Sultan Alkoutibi (College of Southern California)

What’s the matter of this paper? On this work, the workforce proposes a deep studying algorithm that may reconstruct high-quality MRI photos from fast scans. MRI is among the hottest and highly effective medical imaging modalities. Nonetheless, scans can take for much longer than different diagnostic strategies corresponding to CT scans. Present strategies of dashing up MRI scans take fewer measurements of the physique, which degrades picture high quality. Current data-driven AI methods have been efficiently deployed to reconstruct MR photos from accelerated measurements, however their efficiency has stalled in recent times. The workforce’s new technique combines the effectivity of conventional convolutional neural networks with the ability of lately proposed adapter-based architectures for imaginative and prescient functions, creating a brand new state-of-the-art in accelerated MRI reconstruction.

Who can profit from this analysis? This technique might help radiologists and different physicians in two methods. First, it allows the reconstruction of very nice particulars on medical photos that may in any other case have been missed by different methods, which significantly improves the diagnostic worth of such photos. Second, as a result of this technique can get better high-quality photos from accelerated measurements, the length of MRI scans may be considerably lowered. This will likely result in extra environment friendly use of scanners and may scale back their excessive value. General, the purpose is to make MRI extra dependable and efficient for everybody.

Coaching autonomous autos in altering environments

Goal-directed, semi-optimal reinforcement learning in non-static environments

Liu Chen (College of Southern California); Haiping Lu (College of Southern California)

What’s the matter of your paper? this paper About how an agent learns to behave optimally in a altering surroundings. Waymo lately launched a self-service shuttle to the Phoenix airport. Self-driving automotive coaching may be framed as a goal-oriented reinforcement studying downside, which falls inside the framework of the work studied on this paper. Moreover, this analysis examined a non-static altering surroundings, appropriate for capturing altering visitors circumstances.

Understanding the world by way of language and imaginative and prescient

CLiMB: Continuing Learning Benchmark for Vision and Language Tasks

Tejas Srinivasan (College of Southern California)And the Ting Yun Chang (College of Southern California)And the Leticia Leonor Pinto Alva (College of Southern California)And the Georgios Chochlakis (College of Southern California, ISI)And the Mohammad Rostami (College of Southern California)And the Jesse Thomason (College of Southern California)

What’s the matter of your paper? We set a benchmark to review how language and vision-sensitive paradigms can study duties sequentially, corresponding to answering open-ended questions on photos versus answering sure/no questions on pairs of images. This criterion additionally makes it potential to review what occurs when language or imaginative and prescient disappears, corresponding to duties corresponding to categorizing whether or not film evaluations are optimistic or detrimental or figuring out the salient factor in an image.

All papers

Navigating memory construction by simulating pseudo-global continuous learning tasks

Yijia Liu (College of California, Riverside); Wang Ju (College of Southern California); Shaoli Ren (College CaliforniaRiverside)

Near-perfect learning dynamics with no regrets for general convex games

Gabriel Farina (Faculty of Laptop Science, Carnegie Mellon College); Ioannis Anagnostidis (Carnegie Mellon College); Haiping Lu (College of Southern California); Chung Wei Lee (College of Southern California); Christian Krewer (Columbia College); Thomas Sandholm (Carnegie Mellon College)

A near-perfect apology for MDP opponents with delayed reactions for bandits

Tiancheng Jin (College of Southern California); Tal Lancewicki (Tel Aviv College); Haiping Lu (College of Southern California); Yishai Mansour (School of Laptop Science, Tel Aviv College); Aviv Rosenberg (Amazon)

Non-paired learning dynamics with O(log T) switch regression in multiplayer games

Ioannis Anagnostidis (Carnegie Mellon College); Gabriel Farina (Faculty of Laptop Science, Carnegie Mellon College); Christian Krewer (Columbia College); Chung Wei Lee (College of Southern California); Haiping Lu (College of Southern California); Thomas Sandholm (Carnegie Mellon College)

HUMUS-Net: an unpatched multiscale hybrid network architecture for accelerated MRI reconstruction

Zlan Fabian (College of Southern California); Burke Tenaz (College of Southern California); Mehdi Sultan Alkoutibi (College of Southern California)

Strong sparse estimation out by non-convex optimization

Yu Cheng (Brown College); Elias Diakonikolas (College of Wisconsin, Madison); Rong Jie (Duke College); Shivam Gupta (College of Texas, Austin); Daniel Kane (College of California – San Diego); Mehdi Sultan Alkoutibi (College of Southern California)

Self-customized Federated Learning

Huili Chen (College of California, San Diego); Ji Ding (College of Minnesota, Minneapolis); Eric William Tramell (Amazon); Shuang Wu (Amazon); Anit Kumar Sahu (Amazon Alexa AI); Salman Avstemehr (College of Southern California); Tao Chang

NS3: Searching for neural semantic codes

Shushan Araklian (College of Southern California); anna Hakhverdyan (Nationwide Polytechnic College of Armenia); Miltiadis Allamanis (Google); Luis Antonio Garcia (USC ISI); Christoph Hauser (USC/ISI); Xiang Ren (College of Southern California)

Training uncertainty-aware classifiers with matching deep learning

Bat-Shiva Inbender (Technion – Israeli Institute of Know-how, Technion – Israeli Institute of Know-how); Yaniv Romano (Technion, Technion); Mateo Secia (College of Southern California); Yanfei Zhou (College of Southern California)

Estimate the frequency matching with the plotted data

Mateo Secia (College of Southern California); Stefano Favaro (College of Turin)

Why do we need large groups in comparative learning? Gradient perspective and bias

Changyu Chen (SUNY, Buffalo); Jianni Zhang (Duke College); Yi Shu (Amazon); Likun Chen (Duke College); Jiali Duan (College of Southern California); Yiran Chen (Duke College); Son Dinh Tran (College of Maryland, Faculty Park); Belinda Zeng (Amazon); Trishul Chilimbi (Division of Laptop Science, College of Wisconsin-Madison)

Out-of-policy evaluation with policy-dependent improvement response

Wenshuo Guo (College of California, Berkeley); Michael Jordan (College of California, Berkeley); Angela Zhou (College of Southern California)

ALMA: Hierarchical Learning for Composite Multifactor Tasks

Shariq Iqbal (DeepMind); Robbie Costalis (College of Southern California); at SHA (College of Southern California)

Where2comm: Effective collaborative visualization of communication via spatial confidence maps

Yu He (Shanghai Jiao Tong College); Shaoheng Fang (Shanghai Jiao Tong College); Zixing Lei (Shanghai Jiaotong College); Yiqi Zhong (College of Southern California); Siheng Chen (Shanghai Jiao Tong College)

Gattoh’s Empirical Derivatives for Causal Inference

Michael Jordan (College of California, Berkeley); Yixin Wang (UC Berkeley); Angela Zhou (College of Southern California)

Monodynamic display synthesis: a reality check

Hold Zhao (UC Berkeley); Ruilong Lee (College of Southern California); Shubham Tulciani (Carnegie Mellon College); Brian Russell (Adobe Analysis); Anju Kanazawa (College of California, Berkeley)

Unsupervised generalization across tasks via increased recovery

Invoice Jochen Lin (College of Southern California); Kangmin Tan (College of Southern California); Chris Scott Miller (Dartmouth Faculty); Biwen Tian (Tsinghua College, Tsinghua College); Xiang Ren (College of Southern California)

Follow the troubled leader of Markov’s hostile decision ops with his bandit notes

Yan Dai (Institute of Interdisciplinary Info Sciences, Tsinghua College); Haiping Lu (College of Southern California); Liu Chen (College of Southern California)

Posted on November 22, 2022

Final up to date November 22, 2022

#USC #NeurIPS #USC #Viterbi

ZeroToHero

Learn More →

Leave a Reply

Your email address will not be published.