SigProcessing
SigProcessing
  • 130
  • 248 409
Research Seminar: "Robust learning: Worst-case, average-case, or in-between?" by Prof. Hamed Hassani
Spring 2022 SIP Seminar Series: April 5, 2022
[www.inspirelab.us/seminars/]
Speaker: Prof. Hamed Hassani
Abstract: Many of the successes of machine learning are based on minimizing an averaged loss function. However, it is well-known that this paradigm suffers from robustness issues that hinder its applicability in safety-critical domains. These issues are often addressed by training against worst-case perturbations of data, a technique known as adversarial training. Although empirically effective, adversarial training can be overly conservative, leading to unfavorable trade-offs between nominal performance and robustness. To this end, in this paper we propose a framework called probabilistic robustness that bridges the gap between the accurate, yet brittle average case and the robust, yet conservative worst case by enforcing robustness to most rather than to all perturbations. From a theoretical point of view, this framework overcomes the trade-offs between the performance and the sample-complexity of worst-case and average-case learning. From a practical point of view, we propose a novel algorithm based on risk-aware optimization that effectively balances average- and worst-case performance at a considerably lower computational cost relative to adversarial training. Our results on MNIST, CIFAR-10, and SVHN illustrate the advantages of this framework on the spectrum from average- to worst-case robustness.
Biography: Hamed Hassani is currently an assistant professor of Electrical and Systems Engineering department as well as the Computer and Information Systems department, and the Statistics department at the University of Pennsylvania. Prior to that, he was a research fellow at Simons Institute for the Theory of Computing (UC Berkeley) affiliated with the program of Foundations of Machine Learning, and a post-doctoral researcher in the Institute of Machine Learning at ETH Zurich. He received a Ph.D. degree in Computer and Communication Sciences from EPFL, Lausanne. He is the recipient of the 2014 IEEE Information Theory Society Thomas M. Cover Dissertation Award, 2015 IEEE International Symposium on Information Theory Student Paper Award, 2017 Simons-Berkeley Fellowship, 2018 NSF-CRII Research Initiative Award, 2020 Air Force Office of Scientific Research (AFOSR) Young Investigator Award, 2020 National Science Foundation (NSF) CAREER Award, and 2020 Intel Rising Star award. He has recently been selected as the distinguished lecturer of the IEEE Information Theory Society in 2022-2023.
Переглядів: 122

Відео

Research Seminar: "Systematic Function-Correcting Codes" by Prof. Antonia Wachter-Zeh
Переглядів 183Рік тому
Spring 2022 SIP Seminar Series: March 22, 2022 [www.inspirelab.us/seminars/] Speaker: Prof. Antonia Wachter-Zeh Abstract: In this talk, we study function-correcting codes, a new class of codes designed to protect the function evaluation of a message against errors. We show that function-correcting codes are equivalent to irregular-distance codes, i.e., codes that obey some given distance requir...
Research Seminar: "Mixture Models and Mixture of Models" by Prof. Arya Mazumdar
Переглядів 181Рік тому
Spring 2022 SIP Seminar Series: March 1, 2022 [www.inspirelab.us/seminars/] Speaker: Prof. Arya Mazumdar Abstract: The problems of heterogeneity pose major challenges in extracting meaningful information from data as well as in the subsequent decision making or prediction tasks. Heterogeneity brings forward some very fundamental theoretical questions of machine learning. For unsupervised learni...
Research Seminar: "A User Guide to Low-Pass Graph Signal Processing" by Prof. Anna Scaglione
Переглядів 423Рік тому
Spring 2022 SIP Seminar Series: February 15, 2022 [www.inspirelab.us/seminars/] Speaker: Prof. Anna Scaglione Abstract: The notion of graph filters can be used to define generative models for graph data. In fact, the data obtained from many examples of network dynamics may be viewed as the output of a graph filter. With this interpretation, classical signal processing tools such as frequency an...
Research Seminar: "Privacy against Pattern Matching Attacks" by Prof. Hossein Pishro-Nik
Переглядів 1572 роки тому
Fall 2021 SIP Seminar Series: November 30, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Hossein Pishro-Nik Abstract: Privacy of users in modern IoT applications has recently attracted attention. In this talk we focus on the following scenario: Suppose we are given a large number of sequences on a given alphabet, and an adversary is interested in identifying (de-anonymizing) a specific targ...
Research Seminar: "Spatio-Temporal Inference and Learning" by Prof. Visa Koivunen
Переглядів 1042 роки тому
Fall 2021 SIP Seminar Series: November 24, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Visa Koivunen Abstract: We address the problems of learning and inference for large-scale sensor networks observing streaming spatio-temporally varying data. The developed methods stem from distributed learning, multiple hypothesis testing (MHT) and multiple change-point detection for multiple physical ...
Research Seminar: "Convertible Codes" by Prof. Rashmi K. Vinayak
Переглядів 1522 роки тому
Fall 2021 SIP Seminar Series: November 16, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Rashmi K. Vinayak Abstract: In large-scale data storage systems, erasure codes are employed to store data in a redundant fashion to protect against data loss. In this setting, a set of k data blocks to be stored is encoded using an [n, k] code to generate n blocks that are then stored on distinct storag...
Research Seminar: "Scalable and Energy Efficient IoT Networks" by Dr. Alexei Ashikhmin
Переглядів 702 роки тому
Fall 2021 SIP Seminar Series: November 2, 2021 [www.inspirelab.us/seminars/] Speaker: Dr. Alexei Ashikhmin Abstract: In recent years Internet of Things (IoT) has received significant momentum as a hot topic for both industry and academia. This technology promises to turn each physical object into an Internet node and to enable a whole new class of exciting applications and services for Industry...
Research Seminar: "Overcoming Data Availability Attacks in Blockchain Systems" by Prof. Lara Dolecek
Переглядів 3042 роки тому
Fall 2021 SIP Seminar Series: October 19, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Lara Dolecek Abstract: Blockchain systems are already gaining popularity in a variety of applications due to their decentralized design that is favorable in many settings. To overcome excessive storage and latency burden, light nodes and side blockchains have been proposed to, respectively, enhance the b...
Research Seminar: "Machine Learning-Aided Channel Coding" by Prof. Hessam Mahdavifar
Переглядів 5712 роки тому
Fall 2021 SIP Seminar Series: October 5, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Hessam Mahdavifar Abstract: Today, channel codes are among the fundamental parts of any communication system, including cellular, WiFi, and deep space, among others, enabling reliable communications in the presence of noise. Decades of research have led to breakthrough inventions of various families of ch...
Research Seminar: "Privacy-preserving Federated Learning" by Dr. Swanand Kadhe
Переглядів 5722 роки тому
Fall 2021 SIP Seminar Series: September 21, 2021 [www.inspirelab.us/seminars/] Speaker: Dr. Swanand Kadhe Abstract: In modern large-scale machine learning, federated learning has emerged as an important paradigm, where the training data remains distributed over a large number of clients (e.g., mobile phones, smart devices). In federated learning, each client trains a neural network model locall...
Research Seminar: "Black-box Optimization" by Prof. Tara Javidi
Переглядів 4962 роки тому
Spring 2021 SIP Seminar Series: April 21, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Tara Javidi Abstract: In this talk, we will consider the problem of maximizing a black-box function via noisy and costly queries from a theoretical perspective (a lot of it) as well as applications (an exciting bit). We first motivate the problem by considering a wide variety of engineering design applic...
Research Seminar: "Graph Ricci Flow and Applications" by Prof. Jie Gao
Переглядів 1,3 тис.2 роки тому
Spring 2021 SIP Seminar Series: April 7, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Jie Gao Abstract: The notion of curvature describes how spaces are bent at each point and Ricci flow deforms the space such that curvature changes in a way analogous to the diffusion of heat. In this talk I will discuss some recent work in my group on discrete Ollivier Ricci curvature defined on graphs. D...
Research Seminar: "Identifying Dynamics From Finite Data" by Prof. Konstantin Mischaikow
Переглядів 902 роки тому
Spring 2021 SIP Seminar Series: March 24, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Konstantin Mischaikow Abstract: The classical theory of nonlinear dynamics exhibits wonderfully rich and exotic structures. I would argue that as we move to an era of data driven dynamics it offers too many riches. Stated differently, if we only have finite data at our disposal we need a simpler theory o...
Research Seminar: "Low-dimensional Structure in Messy Data" by Prof. Laura Balzano
Переглядів 2383 роки тому
Spring 2021 SIP Seminar Series: March 3, 2021 [www.inspirelab.us/seminars/] Speaker: Prof. Laura Balzano Abstract: In order to draw inferences from large, high-dimensional datasets, we often seek simple structure that model the phenomena represented in those data. Low-rank linear structure is one of the most flexible and efficient such models, allowing efficient prediction, inference, and anoma...
Research Seminar: "Computational Imaging" by Prof. Ulugbek Kamilov
Переглядів 4793 роки тому
Research Seminar: "Computational Imaging" by Prof. Ulugbek Kamilov
Research Seminar: "Coded Computing and Its Applications" by Prof. Salman Avestimehr
Переглядів 9113 роки тому
Research Seminar: "Coded Computing and Its Applications" by Prof. Salman Avestimehr
Research Seminar: "Privacy & Adversarial Robustness in Statistical Estimation" by Prof. Linjun Zhang
Переглядів 2533 роки тому
Research Seminar: "Privacy & Adversarial Robustness in Statistical Estimation" by Prof. Linjun Zhang
Research Seminar: "Distributed Machine Learning" by Prof. Usman Khan
Переглядів 2433 роки тому
Research Seminar: "Distributed Machine Learning" by Prof. Usman Khan
Research Seminar: "Accelerated Gradient Methods on Riemannian Manifolds" by Prof. Suvrit Sra
Переглядів 1,2 тис.3 роки тому
Research Seminar: "Accelerated Gradient Methods on Riemannian Manifolds" by Prof. Suvrit Sra
Research Seminar: "Data-efficient Reinforcement Learning" by Prof. Mengdi Wang
Переглядів 5163 роки тому
Research Seminar: "Data-efficient Reinforcement Learning" by Prof. Mengdi Wang
Research Seminar: "Pooled Coronavirus Testing" by Prof. Dror Baron
Переглядів 1583 роки тому
Research Seminar: "Pooled Coronavirus Testing" by Prof. Dror Baron
Research Seminar: "Statistical Modeling and Uncertainty Quantification" by Prof. Ying Hung
Переглядів 1413 роки тому
Research Seminar: "Statistical Modeling and Uncertainty Quantification" by Prof. Ying Hung
Research Seminar: "Robust and Flexible Distributed Optimization Algorithms" by Prof. Ermin Wei
Переглядів 3983 роки тому
Research Seminar: "Robust and Flexible Distributed Optimization Algorithms" by Prof. Ermin Wei
Research Seminar: "Projective Splitting" by Dr. Patrick Johnstone
Переглядів 2114 роки тому
Research Seminar: "Projective Splitting" by Dr. Patrick Johnstone
Basic Filtering, Sampling, and Aliasing: A Solved Problem
Переглядів 9205 років тому
Basic Filtering, Sampling, and Aliasing: A Solved Problem
Continuous-time Fourier Transform Problems Involving Cosine Signals
Переглядів 1,6 тис.5 років тому
Continuous-time Fourier Transform Problems Involving Cosine Signals
Sampling-Rate Conversion: Non-Integer Sampling-Rate Changes
Переглядів 4,2 тис.7 років тому
Sampling-Rate Conversion: Non-Integer Sampling-Rate Changes
Sampling-Rate Conversion: Interpolation in Time-Domain
Переглядів 2,5 тис.7 років тому
Sampling-Rate Conversion: Interpolation in Time-Domain
Sampling-Rate Conversion: Understanding Interpolation
Переглядів 3,9 тис.7 років тому
Sampling-Rate Conversion: Understanding Interpolation

КОМЕНТАРІ

  • @preetamsingh8067
    @preetamsingh8067 2 місяці тому

    Sir please give some real life example also ,

  • @ninhnguyen4307
    @ninhnguyen4307 10 місяців тому

    Thanks for your lecture!

  • @samisiddiqi5411
    @samisiddiqi5411 Рік тому

    Audio could use some Signal Processing

  • @waniubaid7718
    @waniubaid7718 Рік тому

    ❤️

  • @BTCIVBHUSHANKOLPE
    @BTCIVBHUSHANKOLPE 2 роки тому

    can you please share this ppt ?

    • @SigProcessing
      @SigProcessing 2 роки тому

      Please reach out to the speaker for the presentation.

  • @chinmay.prabhakar
    @chinmay.prabhakar 2 роки тому

    Thank you so much for sharing. This is a wonderful lecture, clearly explained and very intuitive.

  • @vi5hnupradeep
    @vi5hnupradeep 2 роки тому

    Thank you so much for sharing this excellent lecture ! This is gold

  • @Jawaid.Alam_IITB
    @Jawaid.Alam_IITB 2 роки тому

    Really good lecture.

  • @piyushsinha3344
    @piyushsinha3344 2 роки тому

    Frequency is scaled by M(2) ,SO it sud be( pie/2 )and not( pie*2)??

  • @fazilhamza1476
    @fazilhamza1476 3 роки тому

    I liked this video. Just wanted basic explanation using physical world. Great work Prof.

  • @mehdis.7404
    @mehdis.7404 3 роки тому

    great talk, by "Von Neuman computing theory" probably it was meant to refer to "Von Neuman computing architecture". I wonder if methods similar to Shamir's secret sharing techniques can be developed based on this work, but for secure computing.

  • @shayakbhattacharyya2160
    @shayakbhattacharyya2160 3 роки тому

    brilliant series. Thank you!

  • @YasirAmirKhanOffcial
    @YasirAmirKhanOffcial 3 роки тому

    course on Image Processing --> ua-cam.com/video/1DAR0zqOfCg/v-deo.html

  • @liupengwu5205
    @liupengwu5205 3 роки тому

    "Scaling of frequnecy axis by a fator of M" is not true for down sampling I think. When the sentence is true, it is not downsampling , it only change some samples as zero. Downsampling will no change bandwidth if no aliasing happening.

    • @SigProcessing
      @SigProcessing 3 роки тому

      I am afraid that's not true. Sampling-rate conversion is always going to have an effect on your discrete-time spectrum.

  • @Bijoyjouti
    @Bijoyjouti 3 роки тому

    Thank you for such clear representation on Sampling-rate conversion

  • @developingstrong8386
    @developingstrong8386 3 роки тому

    How do you convert DTFT signal into DFT signal

    • @SigProcessing
      @SigProcessing 3 роки тому

      The DTFT of a signal can be converted to the DFT of that signal by replacing \omega with 2\pi k / N (i.e., sampling the DTFT every 2\pi/N radians)

  • @abdulwaheedarshad9763
    @abdulwaheedarshad9763 4 роки тому

    very helpful thanks

  • @shaojunma4737
    @shaojunma4737 4 роки тому

    You should have applied your FIR filter to your audio before publish.

  • @HMWaleed
    @HMWaleed 4 роки тому

    Try to talk in your natural accent, don't try to be a native it is ruining your speaking skills

  • @lovecatxo2471
    @lovecatxo2471 4 роки тому

    thank u

  • @TheExecutioner21
    @TheExecutioner21 4 роки тому

    Please rerecord these videos if you use them for the next classes

  • @vaibhavtiwari6992
    @vaibhavtiwari6992 4 роки тому

    Thankyou so much sir.

  • @SigProcessing
    @SigProcessing 4 роки тому

    Correction #1: At 8:50 in the video, I depict the positive-frequency complex sinusoid as rotating clockwise; it should have been anti-clockwise. Negative-frequency complex sinusoids rotate clockwise, not positive-frequency ones. Clarification #1: The fact that the sinusoid at 8:50 is not starting at the x-axis is because I am not being precise with the plotting. Since the sinusoid has no phase shift, it will indeed start from the x-axis at t=0.

  • @sonalijeswani2554
    @sonalijeswani2554 4 роки тому

    Thank you for the videos. They are very helpful

  • @sln7736
    @sln7736 4 роки тому

    that’t just nice

  • @Mustafa-yc9zh
    @Mustafa-yc9zh 4 роки тому

    Video starts at 20:15

  • @gerudobombshell
    @gerudobombshell 4 роки тому

    Awesome video! Thanks for this - I feel like I'm beginning to wrap my head around the Laplace / Z Transform now!

  • @jameshopkins3541
    @jameshopkins3541 4 роки тому

    NO ME GUSTAN LOS GARABATOS

  • @kiranrm1935
    @kiranrm1935 5 років тому

    Acha english bolna seekhiye sir ji

    • @SigProcessing
      @SigProcessing 5 років тому

      I am too old for that, but hopefully you can find other videos on UA-cam that satisfy your "English-accent cravings." :)

  • @ROHANKUMAR-ve5nr
    @ROHANKUMAR-ve5nr 5 років тому

    Brilliant work sir

  • @xiaotianqiang9866
    @xiaotianqiang9866 5 років тому

    one of the examples: θ(ω)=-2ω^2, should Τg(ω) be 4w instead of 2ω?

    • @SigProcessing
      @SigProcessing 5 років тому

      You are right; that looks like a typo.

  • @waseemshabir3733
    @waseemshabir3733 5 років тому

    very well explained sir

  • @ameynaik2743
    @ameynaik2743 5 років тому

    Good video, you can conveniently run at 1.5x speed to save time.

    •  4 роки тому

      2x for me

  • @adhit528
    @adhit528 5 років тому

    very nice work

  • @user-vg5kt9gm7g
    @user-vg5kt9gm7g 5 років тому

    your voice。。。OK

  • @sant2maya
    @sant2maya 5 років тому

    audio is horrible. pls take care of that next time

    • @SigProcessing
      @SigProcessing 5 років тому

      Yes, I need to redo some of these videos. One of these days hopefully.

  • @sunkarasaigoutham
    @sunkarasaigoutham 6 років тому

    I am working as High Speed Analog IP design Engineer with Intel for 3 years and I can say this basic course helped me a lot to get intuition and then I could go on ! Thank you so much

    • @SigProcessing
      @SigProcessing 5 років тому

      Thank you for your kind comment. Glad to hear the videos have been useful for you.

  • @MuddasirJahangir22
    @MuddasirJahangir22 6 років тому

    Very well explained Sir. Thank you very much for the video :)

    • @SigProcessing
      @SigProcessing 5 років тому

      Glad to hear you got something out of it.

  • @TmWGaM3rS
    @TmWGaM3rS 6 років тому

    Hi, before anything, thanks for your videos, they are very good. I have one doubt in this video, my doubt is: independently of the sampling period T of the C/D conversor, it will be aliasing in the frequency domain? I mean, if we sample x(t) with the Nyquist frequency, i.e (2*Omega maximum of X(w)), in that case there will be aliasing too? Thanks in advance!

    • @SigProcessing
      @SigProcessing 6 років тому

      Sorry for the late reply; I have been offline for a couple of weeks. As for your query, I might have misunderstood it, in which case you can ask for a clarification. If your original sampling rate is exactly the Nyquist rate and you carry out downsampling without having an anti-aliasing filter (i.e., you don't do decimation, in the language of the video), then you are bound to encounter aliasing (in the frequency domain). If, however, you do decimation and have a low-pass filter before the downsampling stage then you can avoid aliasing (although there will be lost frequency content).

  • @skater1250
    @skater1250 6 років тому

    thumbs up for the song choice

  • @praveenkandula7727
    @praveenkandula7727 6 років тому

    Good explanation.

  • @martinlorenrd4073
    @martinlorenrd4073 6 років тому

    Here is a practical application of the Sinc Interpolation with a software for Oscilloscopes: ua-cam.com/video/1W8B4rQeumM/v-deo.html

  • @chandanjaiswal7011
    @chandanjaiswal7011 8 років тому

    At the interval 16:30 in the above video , you plotted the DTFT from -pi to +pi , but the value of omega(w) can vary from - infinity to +infinity , why ? Aslo please explain why we are sampling omega in 2*pi intervals only as it is a angular frequency and can vary upto infinity in a given ramdom signal , though I heard that DTFT is periodic and continuous but don't know the reason .

    • @SigProcessing
      @SigProcessing 8 років тому

      The reason we focus only on any 2\pi interval of discrete frequencies is because discrete frequencies are only unique up to addition of integer multiples of 2\pi. That is, a frequency \omega_0 and \omega_0 + 2\pi k for any integer k means the same frequency. The easiest way to see this is by noting that e^{j (\omega_0 + 2\pi k)n} = e^{j \omega_0 n} e^{j 2\pi k n} = e^{j \omega_0 n}, since e^{j 2\pi k n} = 1 for all integers k and n. These topics are discussed in an earlier lecture on discrete-time Fourier transform and you may want to watch them for further explanation.

  • @kevintoth5719
    @kevintoth5719 8 років тому

    Thanks again for another great video. Much easier to understand the concepts and properties of the four types of linear-phase FIR filters.

  • @kevintoth5719
    @kevintoth5719 8 років тому

    Excellent description of FIR filters and linear phase/group delay. Thank you very much!

  • @94D33M
    @94D33M 8 років тому

    Sir...you are great!

  • @benbenny1438
    @benbenny1438 8 років тому

    how the DFT can be obtained from the DTFT?

    • @SigProcessing
      @SigProcessing 8 років тому

      +mahinder singh As we discuss in the video, DFT is simply samples of the DTFT every 2\pi/N radians/sample, starting from 0 and going up to 2\pi (N-1)/N radians/sample. So if you are given DTFT, you compute the DTFT at these discrete set of frequencies and the answer would be the DFT sequence.

    • @uanlsoporte5332
      @uanlsoporte5332 5 років тому

      ​@@SigProcessing Hi sir! as i understand DTFT is just the correlation of x[n] with a CONTINUOUS basis complex exponentials? am i right?

    • @SigProcessing
      @SigProcessing 5 років тому

      @@uanlsoporte5332 There are different ways of interpreting DTFT, depending upon one's background. Signal processing folks might interpret it differently than harmonic analysis people, just because of the technical language one is used to. The interpretation you are thinking about is correct (this correlation can also be thought of as a projection step) from the signal processing engineer's perspective. You are correlating your signal to a continuum of frequencies and recording the answer returned by this correlation for each \omega.

  • @kabascoolr
    @kabascoolr 8 років тому

    At around 22:22, using the definition which you have, of the difference equations having having constant coefficients, how is h[n]=sinc IRR? Or can "n" be considered a constant, in this. Thanks.

    • @SigProcessing
      @SigProcessing 8 років тому

      +kabascoolr An IIR filter is simply one that has an impulse response that goes on forever (either in positive or negative 'n' direction, or both). Since sinc satisfies this definition, it is automatically an IIR filter. The point you are referring to is the fact that SOME IIR filters can be expressed in terms of constant-coefficient difference equations. However, that is not a necessary condition for a filter to be an IIR filter. Also, the idea here is that if an IIR filter CANNOT be represented using a constant-coefficient difference equation then we cannot implement it inside digital systems.

  • @sakeekawsar4003
    @sakeekawsar4003 8 років тому

    Thanks!It was very helpful!

  • @TheFinalRevelation1
    @TheFinalRevelation1 8 років тому

    Great lecture