October 2024 - talks - lists.cs.princeton.edu

Machine Learning Seminar: Prof. Ramya Korlakai Vinayak visiting on Friday, Oct 4th
by Emily C. Lawrence 03 Oct '24

03 Oct '24

Machine Learning Seminar Prof. Ramya Korlakai Vinayak from UW-Madison is visiting this Friday, Oct 4th from 11:00 am -12:00 pm at Friend Center 008. Title: Towards Pluralistic Alignment: Foundations for Learning from Diverse Human Preferences Abstract: Large pre-trained models trained on internet-scale data are often not ready for safe deployment out-of-the-box. They are heavily fine-tuned and aligned using large quantities of human preference data, usually elicited using pairwise comparisons. While aligning an AI/ML model to human preferences or values, it is important to ask whose preference and values we are aligning it to ? The current approaches of alignment are severely limited due to their inherent uniformity assumption. While there is rich literature on learning preferences from human judgements using comparison queries, the models often focus on learning average preference over the population due to the limitations on the amount of data available per individual or on learning an individual's preference using a lot of queries. Furthermore, the knowledge of the metric, i.e., the way humans judge similarity and dissimilarity, is assumed to be known which does not hold in practice. We aim to overcome these limitations by building mathematical foundations for learning from diverse human preferences. In this talk, I will present, PAL, a personalize-able reward modelling framework for pluralistic alignment, which captures diversity in preferences while also capturing commonalities that can be learned by pooling together data from individuals. I will also discuss some recent theoretical results on per user sample complexity for generalization and fundamental limitations when there are limited pairwise comparisons. Based on work with Daiwei Chen, Yi Chen, Aniket Rege, Zhi Wang, Geelon So, Greg Canal, Blake Mason, Gokcan Tatli, and Rob Nowak. References: 1. [ https://arxiv.org/pdf/2406.08469 | PAL: Pluralistic Alignment Framework for learning from heterogeneous preferences ] (preprint, 2024) 2. [ https://arxiv.org/abs/2207.03609 | One-for-all: Simultaneous metric and preference learning ] (appeared in Neurips 2022) 3. [ https://arxiv.org/pdf/2403.19629 | Metric learning via limited pairwise comparisons ] (appeared in UAI 2024), and 4. [ https://proceedings.mlr.press/v238/tatli24a.html | Learning Populations of Preferences via pairwise comparisons ] (appeared in AISTATS 2024). Bio: Ramya Korlakai Vinayak is the Dugald C. Jackson assistant professor in the Dept. of ECE and affiliated faculty in the Dept. of Computer Science and the Dept. of Statistics at the UW-Madison. Her research interests span the areas of machine learning, statistical inference, and crowdsourcing, with a focus on preference learning and alignment under heterogeneity, reliable and efficient dataset creation, and human-in-the-loop systems. Her works aim to address theoretical and practical challenges that arise when learning from heterogeneous societal data. Prior to joining UW-Madison, Ramya was a postdoctoral researcher in the Paul G. Allen School of Computer Science and Engineering at the University of Washington. She received her Ph.D. in Electrical Engineering from Caltech. She obtained her Masters from Caltech and Bachelors from IIT Madras. She is a recipient of the Schlumberger Foundation Faculty of the Future fellowship from 2013-15, and an invited participant at the Rising Stars in EECS workshop in 2019. She is the recipient of NSF CAREER Award in 2023.

1 0

NextG Distinguished Seminar Series: Thursday Oct. 10 at 4:30 pm
by Emily C. Lawrence 03 Oct '24

03 Oct '24

Speaker: Arjuna Madanayake, Florida International University Day: October 10, 2024 Time: 4:30pm Location: B205 Engineering Quadrangle Hosts: Yasaman Ghasempour and Kaushik Sengupta Event Page: [ https://ece.princeton.edu/events/spectre-signal-processing-efficient-comput… | https://ece.princeton.edu/events/spectre-signal-processing-efficient-comput… ] Title: SPECTRE: Signal Processing with Efficient Compute for Transmit and Receive Environments Abstract: Wideband radio frequency (RF) array processors form an integral building block of modern wireless communications, location, sensing, and electronic warfare systems. There is interest in multifunctional and frequency agile antenna arrays with fully digital beam processing and machine learning (ML) algorithms at the backend that operate across a wide span of frequencies starting from VHF/UHF up to mm-waves and even sub-THz ranges. Digital signal processing (DSP) across massive bandwidths – today up to about 30 GHz for some sub-THz applications – requires careful consideration of multiple design constraints. Most array DSP operations consist of typical functions, such as discrete Fourier transforms (DFTs), finite impulse response (FIR) filters, cross-correlators, multiply-accumulators, and linear algebra (LA) functions like matrix inversion. Arithmetic complexity of basic DSP algorithms must be carefully factored into their design. A classic example is the use of sparse matrix factorizations on the DFT matrix leading to the suite of fast algorithms known as Fast Fourier Transforms (FFTs) of which the most famous is the Cooley-Tukey FFT invented at Princeton and IBM. The claim to fame for the FFT is its enormously useful factorizations having low arithmetic complexity at O(N log2 N), which is tiny compared to the case of the DFT. The DFT is computed brute-force at arithmetic complexity O(N2) where N is the number of points used. The theoretical lower bounds of DFT complexity have been rigorously proven decades in the past by Heideman and Burrus. Luckily, we can get around these bounds via a sequence of approximations. We achieve O(N) complexity for FFT-like operations by recognizing how real-world RF systems have multiple sources of error. The theoretical complexity bounds of the DFT could be circumvented by computing judiciously tailored low-complexity DFT matrix approximations capable of closely preserving mathematical properties of the exact DFT. In this talk, we discuss the use of approximate DFTs (ADFT) with corresponding sparse factorizations that are free of multiplications. The author and his collaborator RJ Cintra have developed an approach for superfast wideband RF signal processing for FFT applications, including multi-beam spacetime orthogonal beamforming, orthogonal frequency division multiplexing (OFDM) and maximally-decimated uniform-DFT polyphase FIR filterbanks, that replace O(N log2 N) FFTs of Cooley-Tukey fame with factorized ADFTs that are within 2 dB of performance of exact DFT but show much lower arithmetic complexity at just O(N). This extra log2 N reduction beyond what is possible from FFTs – which of course comes at a small price – has massive benefits in both chip area and power consumption for many wideband RF systems. In this talk, the author discusses our recent DARPA, ONR, and NSF funded research activities surrounding low-complexity low-power DSP for real-world RF multi-beam beamforming, low-SWaP spatio-temporal multidimensional spectrum sensing, and low-complexity adaptive digital null-forming using Howells-Applebaum adaptive arrays that make use of the O(N) ADFTs. The talk covers the basic mathematical theory, algorithms, RF/microwave systems, digital compute and FPGA/RF-SoC and chiplet based CMOS realizations underway at Florida International University. Bio: Arjuna Madanayake is an Associate Professor of Electrical and Computer Engineering at Florida International University (FIU), Miami, Florida. He runs the RAND Lab which has 15 PhD students at present. His area of specialization is multidimensional systems and signal processing. Dr. Madanayake’s wider research interests span one- and multi-dimensional signal processing and filter design (both analog and digital), RF/microwave systems, analog-CMOS and analog computer architecture, digital systems and FPGA/RF-SoC systems, fast algorithms and digital ASIC, digital hardware based ``software defined’’ radios, electromagnetics and sub-THz nearfield processing, wireless communications, 5G/6G and mm-wave systems, spectrum, RF artificial intelligence and machine learning (AI/ML) and semiconductor technologies.

1 0

Liyi Zhang will present his General Exam "Explaining Distributed Representations by Deep Neural Networks with Bayesian Inference" on Friday, October 4, 2024 at 10:00 AM in Friend Center 005 and via zoom.
by CS Grad Department 01 Oct '24

01 Oct '24

Liyi Zhang will present his General Exam " Explaining Distributed Representations by Deep Neural Networks with Bayesian Inference " on Friday, October 4, 2024 at 10:00 AM in Friend Center 005 and via zoom. Zoom link: https://princeton.zoom.us/j/9603487644 Committee Members: Tom Griffiths (advisor), Ryan Adams , Danqi Chen Abstract: Human-like intelligence in machines has been of interest to philosophers and engineers for centuries. Today’s deep learning allows computational models to obtain meaningful representations on high-dimensional inputs such as text and images. However, understanding these representations is challenging and why they work remains opaque. Our aim goes in two directions: 1) using computational foundations of human cognition to understand and to improve deep neural networks; 2) reverse-engineering the mind by using success in deep neural networks to understand human cognitive processes. In both directions, Bayesian inference serves as a bridge between the two fields. In the first goal, we present a Bayesian interpretation of the autoregressive objective under several distributional assumptions to show where the optimal content of language model embeddings can be identified. In the second goal, we show how variants of prior distributions in variational auto-encoders, a probabilistic generative model, explain how forming prototypes contributes to semi-supervised categorization. Reading List: https://docs.google.com/document/d/13_DPdcPTKklMbzzabiUXY7q-BiLmjnBJGoPZvge… Everyone is invited to attend the talk, and those faculty wishing to remain for the oral exam following are welcome to do so.

1 0