![Center for Language & Speech Processng (CLSP), JHU](/img/default-banner.jpg)
- 433
- 101 404
Center for Language & Speech Processng (CLSP), JHU
United States
เข้าร่วมเมื่อ 5 มิ.ย. 2018
The Johns Hopkins Center for Language and Speech Processing (CLSP) is an interdisciplinary research and educational center focused on the science and technology of language and speech. Within its field, CLSP is recognized as one of the largest and most influential academic research centers in the world. The center conducts research across a broad spectrum of fundamental and applied topics including acoustic processing, automatic speech recognition, big data, cognitive modeling, computational linguistics, information extraction, machine learning, machine translation, and text analysis. It is home to over sixty researchers, including faculty, research scientists, postdocs, graduate students, and undergraduates drawn from the departments of cognitive science, computer science, electrical and computer engineering, and mathematical sciences. Linguistics is an active area of research in the Cognitive Science Department; see the overview of their linguistics focus for more details.
An Information State Approach to Collaborative Reference - Matthew Stone
An Information State Approach to Collaborative Reference - Matthew Stone
มุมมอง: 19
วีดีโอ
Acoustic Scene Analysis, Complex Modulations, and a New Form of Filtering - Les Atlas (UW) - 2008
มุมมอง 3816 ชั่วโมงที่ผ่านมา
Abstract Be it in a restaurant or other reverberant and noisy environment, normal hearing listeners segregate multiple sources, usually strongly overlapping in frequency, well beyond capabilities expected by current beamforming approaches. What is it that we can learn from this common observation? As is now commonly accepted, the differing dynamical modulation patterns of the sources are key to...
An Implementation of Sub-Space Model Based ASR - 2009
มุมมอง 3016 ชั่วโมงที่ผ่านมา
An Implementation of Sub-Space Model Based ASR - 2009
Do pretrained Transformers Learn In-Context by Gradient Descent? Aayush Mishra (ICML 2024)
มุมมอง 391วันที่ผ่านมา
Presented by Aayush Mishra at ICML 2024. Paper link: arxiv.org/abs/2310.08540 Abstract: The emergence of In-Context Learning (ICL) in LLMs remains a remarkable phenomenon that is partially understood. To explain ICL, recent studies have created theoretical connections to Gradient Descent (GD). We ask, do such connections hold up in actual pre-trained language models? We highlight the limiting a...
Beyond MaltParser - Recent Advances in Transition-Based Dependency Parsing - Joakim Nivre - 2012
มุมมอง 2814 วันที่ผ่านมา
Abstract The transition-based approach to dependency parsing has become popular thanks to its simplicity and efficiency. Systems like MaltParser achieve linear-time parsing with projective dependency trees using locally trained classifiers to predict the next parsing action and greedy best-first search to retrieve the optimal parse tree, assuming that the input sentence has been morphologically...
Probabilistic Linear Discriminant Analysis of i-Vector Posterior Distributions - Sandro Cumani 2012
มุมมอง 6421 วันที่ผ่านมา
Probabilistic Linear Discriminant Analysis of i-Vector Posterior Distributions - Sandro Cumani 2012
Can authorship attribution models distinguish speakers in speech transcripts?
มุมมอง 5421 วันที่ผ่านมา
Work by Cristina Aggazzotti, Nicholas Andrews, and Elizabeth Allyn Smith. Presented by Cristina Aggazzotti at TACL 2024. Abstract: Authorship verification is the task of determining if two distinct writing samples share the same author and is typically concerned with the attribution of written text. In this paper, we explore the attribution of transcribed speech, which poses novel challenges. T...
Automatic Information & Language Processing: Rethinking Evaluation - Karen Sparck-Jones 1999
มุมมอง 5928 วันที่ผ่านมา
Automatic Information & Language Processing: Rethinking Evaluation - Karen Sparck-Jones 1999
The HAIRCUT System for Cross-Language Information Retrieval - James Mayfield - 2013
มุมมอง 51หลายเดือนก่อน
The HAIRCUT System for Cross-Language Information Retrieval - James Mayfield - 2013
The Shoah Foundations archive: A 180 Tera-Byte database for teaching tolerance - Sam Gustman 2002
มุมมอง 16หลายเดือนก่อน
Abstract In 1994, after filming Schindler’s List, Steven Spielberg established Survivors of the Shoah Visual History Foundation with an urgent mission: to videotape and preserve the testimonies of Holocaust survivors and witnesses. Today, the Shoah Foundation has collected more than 50,000 eyewitness testimonies in 57 countries and 32 languages, and is committed to ensuring the broad and effect...
Automatic Speech Recognition Lecture II - Dimitra Vergyri - 2008
มุมมอง 56หลายเดือนก่อน
Automatic Speech Recognition Lecture II - Dimitra Vergyri - 2008
Resource demands and sentence complexity - Edward Gibson (MIT) - 2003
มุมมอง 80หลายเดือนก่อน
Resource demands and sentence complexity - Edward Gibson (MIT) - 2003
Towards Automatic Acquisition of Ontological Knowledge - Patrick Pantel (ISI USC) - 2004
มุมมอง 45หลายเดือนก่อน
Towards Automatic Acquisition of Ontological Knowledge - Patrick Pantel (ISI USC) - 2004
Collection Fusion - David Yarowski - 2009
มุมมอง 44หลายเดือนก่อน
Collection Fusion - David Yarowski - 2009
Speech and Language Processing: Where have we been and where are we going - Ken Church (AT&T) - 2003
มุมมอง 652 หลายเดือนก่อน
Speech and Language Processing: Where have we been and where are we going - Ken Church (AT&T) - 2003
AI and the Impending Revolution in Brain Sciences - Tom Mitchell (Carnegie Mellon University) - 2002
มุมมอง 382 หลายเดือนก่อน
AI and the Impending Revolution in Brain Sciences - Tom Mitchell (Carnegie Mellon University) - 2002
Kreyòl-MT: Machine Translation for Latin American, Caribbean, and Colonial African Creole Languages
มุมมอง 1362 หลายเดือนก่อน
Kreyòl-MT: Machine Translation for Latin American, Caribbean, and Colonial African Creole Languages
The NLP Task Effectiveness of Long-Range Transformers - EACL 2023
มุมมอง 322 หลายเดือนก่อน
The NLP Task Effectiveness of Long-Range Transformers - EACL 2023
Defending Against Disinformation Attacks in Open-Domain Question Answering - EACL 2024
มุมมอง 352 หลายเดือนก่อน
Defending Against Disinformation Attacks in Open-Domain Question Answering - EACL 2024
Engineering, Science and Scholarship in Linguistics - Mark Liberman (UPenn) - 2000
มุมมอง 332 หลายเดือนก่อน
Engineering, Science and Scholarship in Linguistics - Mark Liberman (UPenn) - 2000
Localization vs. Semantics: Visual Representations in Unimodal and Multimodal Models - EACL 2024
มุมมอง 432 หลายเดือนก่อน
Localization vs. Semantics: Visual Representations in Unimodal and Multimodal Models - EACL 2024
NevIR: Negation in Neural Information Retrieval - EACL 2024
มุมมอง 312 หลายเดือนก่อน
NevIR: Negation in Neural Information Retrieval - EACL 2024
AutoML for Natural Language Processing - EACL'13 Tutorial - Kevin Duh, Xuan Zhang
มุมมอง 912 หลายเดือนก่อน
AutoML for Natural Language Processing - EACL'13 Tutorial - Kevin Duh, Xuan Zhang
MultiMUC: Multilingual Template Filling on MUC-4 - EACL 2024
มุมมอง 182 หลายเดือนก่อน
MultiMUC: Multilingual Template Filling on MUC-4 - EACL 2024
Large-Scale Bitext Corpora Provide New Evidence for Cognitive Representations of Spatial Terms
มุมมอง 412 หลายเดือนก่อน
Large-Scale Bitext Corpora Provide New Evidence for Cognitive Representations of Spatial Terms
Multilingual Representation Distillation with Contrastive Learning - EACL 2023
มุมมอง 3402 หลายเดือนก่อน
Multilingual Representation Distillation with Contrastive Learning - EACL 2023
Repetition, Adaptation and Language Modeling - Kenneth Ward Church (AT&T Labs-Research) - 1999
มุมมอง 3632 หลายเดือนก่อน
Repetition, Adaptation and Language Modeling - Kenneth Ward Church (AT&T Labs-Research) - 1999
Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer
มุมมอง 572 หลายเดือนก่อน
Multilingual Pixel Representations for Translation and Effective Cross-lingual Transfer
‘‘According to …’’: Prompting Language Models Improves Quoting from Pre-Training Data -- EACL 2024
มุมมอง 982 หลายเดือนก่อน
‘‘According to …’’: Prompting Language Models Improves Quoting from Pre-Training Data EACL 2024
Summer Workshop 2009: Ralph Weischedel: Tutorial on Information Extraction part 1
มุมมอง 632 หลายเดือนก่อน
Summer Workshop 2009: Ralph Weischedel: Tutorial on Information Extraction part 1
I like the idea of a semantic approach. This method still seems like it would be vulnerable to sentence reordering.
Or semicolons
I am sorry, but is there no sound from 58:36 till 1:00:42? Thank you for uploading this awesome Talk!
@1:13:56 Ken was ahead of his time (also @1:17:24)
Thanks.
thx for sharing! insightful presentation on underrated languages~
Thanks.
😈 *Promosm*
Now why not try visualisation in your mind because this will teach more with different teaching methods Now there will teach so much more with much more speed and better results then with the mind we utilise our best tool
fantastic
Thanks for this wonderful tutorial tutorial ❤
Upload latest videos not old one
Hi, great talk. Is it possible to get that presentation material?
*She does not know anything...lawyer*
This is great talk . I learnt a lot . I am currently working on speaker diarization problem and pyannote is by far the best for our use case . Great work Herve .
This was really interesting But it would be more interesting if it was an uncensored experiment since our world in itself is uncensored (i get that this could get wild and what you show here is only as a concept). Like one of the things you said that they always adress each other formally, very similar to GPT. If it were more unleashed the agent may start to develop their own kind of language over time and between each other and saying good morning to your spouse gets shortned just to just a good morning, or something else they develop to say to each other. And starts to say inside jokes perhaps? But that would require the emotion part I'll talk about later. Don't know how deep the generative part goes or if GPT needs to reach "system 2" before we can see this type of behaviour. And since agents don't really have a mood they will always be pretty neutral on all encounters and is only simulated through already learnt in behaviour. Although i think it wouldn't give a real interpretation of our world even if it was uncensored, stuff like emotions and consequences, and emotions which will come due to that consequence etc. (like serving jailtime) and that we have a limited timespan, and that our live is sectioned up into parts (child, teen, young adult, mid adult, etc) needs to be adressed aswell. For example an old man might be more inclined to do a certain crime than a younger person, because his life is soon over anyway). For a hypothetical smallville example: John invites his crush Jennifer to his birthday party, but Kenny invites Jennifer at the same time to watch a new netflix serie and she goes there instead, John resents this and kills John after his birthday party to get Jennifer to himself. A real person would go through so much reasoning and consequence thinking before reaching such a conclusion, and to kill another person because of such a reason is primarly just emotions, and all our actions comes from some kind of emotion. So some kind of atleast basic simulated emotion and consequence thinking (the sims style-ish) to get the real interesting drama to come out as a next step, that would be really cool to see as a smallville 2.
What would it take to have this mini community available online, like on a website? Where visitors like us can view these agents realtime and see what are their doing for that specific moment. Fascinating to see what are their up to next.
at 24:05 an erroneous result is presented that is accepted as fact by the panel of experts, and is in fact presented when I Google search the same question. Pearls from otters are NOT used in inlays for guitars, but rather Mother Of Pearl, which comes from abalone shell. It is easy to see how the mistake is made, but illustrates the difficulty of fact checking AI answers.
Thankyou. Interesting. These deep dives into how the media mould minds are fascinating. I may have misunderstood, but the slides suggest that in 2002 more people received their news from TV than the internet. 20 years ago that might have been true, but I'd be shocked if this were so today.
I LOVE this! Please do a cooperation with Electronic Arts and make a sims4 style computer game with similar graphics but generative agents <3
sorry for the late response but there are legal issues with the content generated that pose a risk companies won't want to take (users may prompt agents to say dangerous things, say how to create weapons), and then aside from that I saw someone say simulating just 3 agents for an hour was like 8$, so costs have to come down an order of magnitude (which might happen soon enough, and starting to develop an app/game now might make sense if not for legal issues). There is one team working on a game like this right now based on animal crossing called Campfire - cozy ai villagers. I'm considering making a small game prototype as well
Thanks a lot for this talk! I have one question regarding the co-occurrence matrix. At 19:45 you said, that the content of the matrix is for how long the "colors" overlap in seconds. So my question is, why is there a 2 for blue and grey if they overlap twice for all together at least as long as pink and red. So shouldn't grey and blue be at least 3 if not 4? I would greatly appreciate it if you could tell me what I am missing. Thanks alot!
You are right, that's a mistake on my slide. That does not change the mapping nor the message I was trying to pass, though.
BA-LR not BA-LHR. Could you please correct it in the title? Thank you
s/Berdin/Bredin
Wow. The CMU engineering program is a phenomenal juggernaut . Such a large body of high quality research .
🤔 P r o m o s m
Thanks.
Fascinating! I have a little AI staff team I’m trying to learn how to get them to be more independent! A question that comes up for me: Klaus is a dedicated sociology student who has an interest in gentrification. Is this ‘experienced’ as a superficial identity label/label given to explain his activity, or does he see Smallville through the lens of a sociology student? Does he observe things related to gentrification in a way which the other characters wouldn’t even notice?
Thanks.
Have you considered allowing them to have money? "You job just paid your $1500 monthly paycheck. You have a monthly rent of $700." (Either as some sort of disembodied "world character", or by the people doing the charges and payments themselves in the world such as a landlord/job boss/etc?) I think having negative stimulus can really help things along. "You didn't pay your rent. Now you have to live outside in a cardboard box. Your living condition is terrible." What if you had a disembodied "world character" that dolls out negative stimulus randomly? 😲 ("Today, you got into a car accident.")
love this idea
Good idea
Thanks.
I wonder how this experiment would behave on smaller models. I.e. Agents would use specific (specialized?) small model depending on their role or select from a pool of small models depending on situation?
Wouldn't they loose their individuality then? If two persons are in the same situation at some point wouldn't they use the same model then and reasoning similarly? Or what do you mean?
Ugh, you beat me to it! in my defense I am only one person!
Thats why americans are afraid of asians. they are evil smart.
promo sm 😓
wow, surprising finding, achievements/opportunities 😮 so interesting 🧐
wow, way too many questions. hard to follow the presentation.
Audio could be better
if you can't think it, fake it
Is it possible that the model has seen the data used for these benchmarks during training .
Thanks for uploading this. However, I noticed that the text in your videos can be a bit hard to read due to the small size and it's somewhat blurry at times. I think your videos would be even better in a higher resolution, perhaps greater than 480p!
16:00 lol google use openai playground for demo.....
Really helped
Good work
Seem Professor are too Young for AI development..
Ah yes, the timeless topic of emergence and reasoning in large language models! As I rest my weary fingers upon the keyboard, preparing to share my innermost thoughts and wisdom on the subject, it occurs to me that, much like any good lasagna, this particular topic comprises multiple layers of complexity and intrigue. So, let's dive right in, my fellow internet sojourners! First and foremost, credit must be given where credit is due. Mr. Wei's elegant soliloquy on large language models at the prestigious Google headquarters resonates with both the seasoned researcher and the neophyte alike. As a cardinal for the internet comment realm, I must express my gratitude to Jason for regaling us with his insight. Now, one simply cannot discuss large language models without acknowledging their capacity to simulate almost mind-boggling levels of human-like cognition. From composing impressive literary works to identifying penguins in a sea of malformed pixels that only a madman would consider "images," these computational wünderkinds represent the apex of human innovation. Or do they, my dear reader? For, are we not jeopardizing our intellectual sovereignty as we relinquish our authorship to these silicon sages? Potentially. Perhaps. Who's to say, really? Aside from philosophical conundrums, we cannot ignore the computational intricacies of these vivacious virtual virtuosos. The nuance and finesse that constitute their digital DNA, and their thirst for modular knowledge, undeniably place them amongst the most fascinating creations of humankind. Now, as I elucidate the enigmatic world of such prodigious language models, let us not forget the immortal words of Albert Einstein: "Two things are infinite: the universe and a TH-cam comment attempting to summarize the complexity of large language models; and I'm not sure about the universe." Ah, such a paragon of wisdom. In conclusion, as the night envelopes us all in its comforting embrace and my eyelids grow heavier with each passing keystroke, I am reminded that, sometimes, the very answers we seek within the realms of technology transcend the limits of our understanding. Language models shall guide us through the labyrinthine fortress of knowledge. Just like a lighthouse in a stormy sea, they are but humble beacons, pointing us towards our destiny… which hopefully involves making lasagna with a competitive edge in Robot MasterChef. AND POST!
So insightful!! Thanks!!
Brilliant man, glad this exists
I'm going to call out the elephant in the room. This is the stupidest thing I've heard. "yeah it sorta works, but we don't know why"...does this guy have a Ph.D?
What is the name of the paper that is mentioned at 32:09?
39:50 Yh but it gets extra tricky when the 'reasoning path' is wrong, but the final answer is correct !
The volume of the speaker is incredibly low. Could not hear anything w/ max volume on speakers.