Search

Your search keyword '"Sedghi, Hanie"' showing total 14 results
Start Over Author "Sedghi, Hanie" Publication Year Range Last 3 years
14 results on '"Sedghi, Hanie"'

1. Training Language Models on the Knowledge Graph: Insights on Hallucinations and Their Detectability

Author

Hron, Jiri, Culp, Laura, Elsayed, Gamaleldin, Liu, Rosanne, Adlam, Ben, Bileschi, Maxwell, Bohnet, Bernd, Co-Reyes, JD, Fiedel, Noah, Freeman, C. Daniel, Gur, Izzeddin, Kenealy, Kathleen, Lee, Jaehoon, Liu, Peter J., Mishra, Gaurav, Mordatch, Igor, Nova, Azade, Novak, Roman, Parisi, Aaron, Pennington, Jeffrey, Rizkowsky, Alex, Simpson, Isabelle, Sedghi, Hanie, Sohl-dickstein, Jascha, Swersky, Kevin, Vikram, Sharad, Warkentin, Tris, Xiao, Lechao, Xu, Kelvin, Snoek, Jasper, and Kornblith, Simon

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence, Computer Science - Machine Learning
Abstract

While many capabilities of language models (LMs) improve with increased training budget, the influence of scale on hallucinations is not yet fully understood. Hallucinations come in many forms, and there is no universally accepted definition. We thus focus on studying only those hallucinations where a correct answer appears verbatim in the training set. To fully control the training data content, we construct a knowledge graph (KG)-based dataset, and use it to train a set of increasingly large LMs. We find that for a fixed dataset, larger and longer-trained LMs hallucinate less. However, hallucinating on $\leq5$% of the training data requires an order of magnitude larger model, and thus an order of magnitude more compute, than Hoffmann et al. (2022) reported was optimal. Given this costliness, we study how hallucination detectors depend on scale. While we see detector size improves performance on fixed LM's outputs, we find an inverse relationship between the scale of the LM and the detectability of its hallucinations., Comment: Published at COLM 2024. 16 pages, 11 figures

Published
2024

2. Exploring and Benchmarking the Planning Capabilities of Large Language Models

Author

Bohnet, Bernd, Nova, Azade, Parisi, Aaron T, Swersky, Kevin, Goshvadi, Katayoon, Dai, Hanjun, Schuurmans, Dale, Fiedel, Noah, and Sedghi, Hanie

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence, Computer Science - Machine Learning
Abstract

Classical and natural language planning tasks remain a difficult domain for modern large language models (LLMs). In this work, we lay the foundations for improving planning capabilities of LLMs. First, we construct a comprehensive benchmark suite encompassing both classical planning benchmarks and natural language scenarios. This suite includes algorithms to methodically generate instances of tasks with varying levels of difficulty, allowing for rigorous and systematic evaluation of LLM performance. Next, we investigate the use of many-shot in-context learning to enhance LLM planning, exploring the relationship between increased context length and improved planning performance. In addition, we demonstrate the positive impact of fine-tuning LLMs on optimal planning paths. We also probe the efficacy of chain-of-thought reasoning methods to improve LLM planning performance. Moreover, we probe the performance of the proposed methods in out-of-distribution scenarios, assessing the ability to generalize to novel and unseen planning challenges. Finally, we investigate model's failure modes and reveal insights that hold true across different benchmarks.

Published
2024

3. Long-Span Question-Answering: Automatic Question Generation and QA-System Ranking via Side-by-Side Evaluation

Author

Bohnet, Bernd, Swersky, Kevin, Liu, Rosanne, Awasthi, Pranjal, Nova, Azade, Snaider, Javier, Sedghi, Hanie, Parisi, Aaron T, Collins, Michael, Lazaridou, Angeliki, Firat, Orhan, and Fiedel, Noah

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence
Abstract

We explore the use of long-context capabilities in large language models to create synthetic reading comprehension data from entire books. Previous efforts to construct such datasets relied on crowd-sourcing, but the emergence of transformers with a context size of 1 million or more tokens now enables entirely automatic approaches. Our objective is to test the capabilities of LLMs to analyze, understand, and reason over problems that require a detailed comprehension of long spans of text, such as questions involving character arcs, broader themes, or the consequences of early actions later in the story. We propose a holistic pipeline for automatic data generation including question generation, answering, and model scoring using an ``Evaluator''. We find that a relative approach, comparing answers between models in a pairwise fashion and ranking with a Bradley-Terry model, provides a more consistent and differentiating scoring mechanism than an absolute scorer that rates answers individually. We also show that LLMs from different model families produce moderate agreement in their ratings. We ground our approach using the manually curated NarrativeQA dataset, where our evaluator shows excellent agreement with human judgement and even finds errors in the dataset. Using our automatic evaluation approach, we show that using an entire book as context produces superior reading comprehension performance compared to baseline no-context (parametric knowledge only) and retrieval-based approaches.

Published
2024

4. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context

Author

Gemini Team, Georgiev, Petko, Lei, Ving Ian, Burnell, Ryan, Bai, Libin, Gulati, Anmol, Tanzer, Garrett, Vincent, Damien, Pan, Zhufeng, Wang, Shibo, Mariooryad, Soroosh, Ding, Yifan, Geng, Xinyang, Alcober, Fred, Frostig, Roy, Omernick, Mark, Walker, Lexi, Paduraru, Cosmin, Sorokin, Christina, Tacchetti, Andrea, Gaffney, Colin, Daruki, Samira, Sercinoglu, Olcan, Gleicher, Zach, Love, Juliette, Voigtlaender, Paul, Jain, Rohan, Surita, Gabriela, Mohamed, Kareem, Blevins, Rory, Ahn, Junwhan, Zhu, Tao, Kawintiranon, Kornraphop, Firat, Orhan, Gu, Yiming, Zhang, Yujing, Rahtz, Matthew, Faruqui, Manaal, Clay, Natalie, Gilmer, Justin, Co-Reyes, JD, Penchev, Ivo, Zhu, Rui, Morioka, Nobuyuki, Hui, Kevin, Haridasan, Krishna, Campos, Victor, Mahdieh, Mahdis, Guo, Mandy, Hassan, Samer, Kilgour, Kevin, Vezer, Arpi, Cheng, Heng-Tze, de Liedekerke, Raoul, Goyal, Siddharth, Barham, Paul, Strouse, DJ, Noury, Seb, Adler, Jonas, Sundararajan, Mukund, Vikram, Sharad, Lepikhin, Dmitry, Paganini, Michela, Garcia, Xavier, Yang, Fan, Valter, Dasha, Trebacz, Maja, Vodrahalli, Kiran, Asawaroengchai, Chulayuth, Ring, Roman, Kalb, Norbert, Soares, Livio Baldini, Brahma, Siddhartha, Steiner, David, Yu, Tianhe, Mentzer, Fabian, He, Antoine, Gonzalez, Lucas, Xu, Bibo, Kaufman, Raphael Lopez, Shafey, Laurent El, Oh, Junhyuk, Hennigan, Tom, Driessche, George van den, Odoom, Seth, Lucic, Mario, Roelofs, Becca, Lall, Sid, Marathe, Amit, Chan, Betty, Ontanon, Santiago, He, Luheng, Teplyashin, Denis, Lai, Jonathan, Crone, Phil, Damoc, Bogdan, Ho, Lewis, Riedel, Sebastian, Lenc, Karel, Yeh, Chih-Kuan, Chowdhery, Aakanksha, Xu, Yang, Kazemi, Mehran, Amid, Ehsan, Petrushkina, Anastasia, Swersky, Kevin, Khodaei, Ali, Chen, Gowoon, Larkin, Chris, Pinto, Mario, Yan, Geng, Badia, Adria Puigdomenech, Patil, Piyush, Hansen, Steven, Orr, Dave, Arnold, Sebastien M. R., Grimstad, Jordan, Dai, Andrew, Douglas, Sholto, Sinha, Rishika, Yadav, Vikas, Chen, Xi, Gribovskaya, Elena, Austin, Jacob, Zhao, Jeffrey, Patel, Kaushal, Komarek, Paul, Austin, Sophia, Borgeaud, Sebastian, Friso, Linda, Goyal, Abhimanyu, Caine, Ben, Cao, Kris, Chung, Da-Woon, Lamm, Matthew, Barth-Maron, Gabe, Kagohara, Thais, Olszewska, Kate, Chen, Mia, Shivakumar, Kaushik, Agarwal, Rishabh, Godhia, Harshal, Rajwar, Ravi, Snaider, Javier, Dotiwalla, Xerxes, Liu, Yuan, Barua, Aditya, Ungureanu, Victor, Zhang, Yuan, Batsaikhan, Bat-Orgil, Wirth, Mateo, Qin, James, Danihelka, Ivo, Doshi, Tulsee, Chadwick, Martin, Chen, Jilin, Jain, Sanil, Le, Quoc, Kar, Arjun, Gurumurthy, Madhu, Li, Cheng, Sang, Ruoxin, Liu, Fangyu, Lamprou, Lampros, Munoz, Rich, Lintz, Nathan, Mehta, Harsh, Howard, Heidi, Reynolds, Malcolm, Aroyo, Lora, Wang, Quan, Blanco, Lorenzo, Cassirer, Albin, Griffith, Jordan, Das, Dipanjan, Lee, Stephan, Sygnowski, Jakub, Fisher, Zach, Besley, James, Powell, Richard, Ahmed, Zafarali, Paulus, Dominik, Reitter, David, Borsos, Zalan, Joshi, Rishabh, Pope, Aedan, Hand, Steven, Selo, Vittorio, Jain, Vihan, Sethi, Nikhil, Goel, Megha, Makino, Takaki, May, Rhys, Yang, Zhen, Schalkwyk, Johan, Butterfield, Christina, Hauth, Anja, Goldin, Alex, Hawkins, Will, Senter, Evan, Brin, Sergey, Woodman, Oliver, Ritter, Marvin, Noland, Eric, Giang, Minh, Bolina, Vijay, Lee, Lisa, Blyth, Tim, Mackinnon, Ian, Reid, Machel, Sarvana, Obaid, Silver, David, Chen, Alexander, Wang, Lily, Maggiore, Loren, Chang, Oscar, Attaluri, Nithya, Thornton, Gregory, Chiu, Chung-Cheng, Bunyan, Oskar, Levine, Nir, Chung, Timothy, Eltyshev, Evgenii, Si, Xiance, Lillicrap, Timothy, Brady, Demetra, Aggarwal, Vaibhav, Wu, Boxi, Xu, Yuanzhong, McIlroy, Ross, Badola, Kartikeya, Sandhu, Paramjit, Moreira, Erica, Stokowiec, Wojciech, Hemsley, Ross, Li, Dong, Tudor, Alex, Shyam, Pranav, Rahimtoroghi, Elahe, Haykal, Salem, Sprechmann, Pablo, Zhou, Xiang, Mincu, Diana, Li, Yujia, Addanki, Ravi, Krishna, Kalpesh, Wu, Xiao, Frechette, Alexandre, Eyal, Matan, Dafoe, Allan, Lacey, Dave, Whang, Jay, Avrahami, Thi, Zhang, Ye, Taropa, Emanuel, Lin, Hanzhao, Toyama, Daniel, Rutherford, Eliza, Sano, Motoki, Choe, HyunJeong, Tomala, Alex, Safranek-Shrader, Chalence, Kassner, Nora, Pajarskas, Mantas, Harvey, Matt, Sechrist, Sean, Fortunato, Meire, Lyu, Christina, Elsayed, Gamaleldin, Kuang, Chenkai, Lottes, James, Chu, Eric, Jia, Chao, Chen, Chih-Wei, Humphreys, Peter, Baumli, Kate, Tao, Connie, Samuel, Rajkumar, Santos, Cicero Nogueira dos, Andreassen, Anders, Rakićević, Nemanja, Grewe, Dominik, Kumar, Aviral, Winkler, Stephanie, Caton, Jonathan, Brock, Andrew, Dalmia, Sid, Sheahan, Hannah, Barr, Iain, Miao, Yingjie, Natsev, Paul, Devlin, Jacob, Behbahani, Feryal, Prost, Flavien, Sun, Yanhua, Myaskovsky, Artiom, Pillai, Thanumalayan Sankaranarayana, Hurt, Dan, Lazaridou, Angeliki, Xiong, Xi, Zheng, Ce, Pardo, Fabio, Li, Xiaowei, Horgan, Dan, Stanton, Joe, Ambar, Moran, Xia, Fei, Lince, Alejandro, Wang, Mingqiu, Mustafa, Basil, Webson, Albert, Lee, Hyo, Anil, Rohan, Wicke, Martin, Dozat, Timothy, Sinha, Abhishek, Piqueras, Enrique, Dabir, Elahe, Upadhyay, Shyam, Boral, Anudhyan, Hendricks, Lisa Anne, Fry, Corey, Djolonga, Josip, Su, Yi, Walker, Jake, Labanowski, Jane, Huang, Ronny, Misra, Vedant, Chen, Jeremy, Skerry-Ryan, RJ, Singh, Avi, Rijhwani, Shruti, Yu, Dian, Castro-Ros, Alex, Changpinyo, Beer, Datta, Romina, Bagri, Sumit, Hrafnkelsson, Arnar Mar, Maggioni, Marcello, Zheng, Daniel, Sulsky, Yury, Hou, Shaobo, Paine, Tom Le, Yang, Antoine, Riesa, Jason, Rogozinska, Dominika, Marcus, Dror, Badawy, Dalia El, Zhang, Qiao, Wang, Luyu, Miller, Helen, Greer, Jeremy, Sjos, Lars Lowe, Nova, Azade, Zen, Heiga, Chaabouni, Rahma, Rosca, Mihaela, Jiang, Jiepu, Chen, Charlie, Liu, Ruibo, Sainath, Tara, Krikun, Maxim, Polozov, Alex, Lespiau, Jean-Baptiste, Newlan, Josh, Cankara, Zeyncep, Kwak, Soo, Xu, Yunhan, Chen, Phil, Coenen, Andy, Meyer, Clemens, Tsihlas, Katerina, Ma, Ada, Gottweis, Juraj, Xing, Jinwei, Gu, Chenjie, Miao, Jin, Frank, Christian, Cankara, Zeynep, Ganapathy, Sanjay, Dasgupta, Ishita, Hughes-Fitt, Steph, Chen, Heng, Reid, David, Rong, Keran, Fan, Hongmin, van Amersfoort, Joost, Zhuang, Vincent, Cohen, Aaron, Gu, Shixiang Shane, Mohananey, Anhad, Ilic, Anastasija, Tobin, Taylor, Wieting, John, Bortsova, Anna, Thacker, Phoebe, Wang, Emma, Caveness, Emily, Chiu, Justin, Sezener, Eren, Kaskasoli, Alex, Baker, Steven, Millican, Katie, Elhawaty, Mohamed, Aisopos, Kostas, Lebsack, Carl, Byrd, Nathan, Dai, Hanjun, Jia, Wenhao, Wiethoff, Matthew, Davoodi, Elnaz, Weston, Albert, Yagati, Lakshman, Ahuja, Arun, Gao, Isabel, Pundak, Golan, Zhang, Susan, Azzam, Michael, Sim, Khe Chai, Caelles, Sergi, Keeling, James, Sharma, Abhanshu, Swing, Andy, Li, YaGuang, Liu, Chenxi, Bostock, Carrie Grimes, Bansal, Yamini, Nado, Zachary, Anand, Ankesh, Lipschultz, Josh, Karmarkar, Abhijit, Proleev, Lev, Ittycheriah, Abe, Yeganeh, Soheil Hassas, Polovets, George, Faust, Aleksandra, Sun, Jiao, Rrustemi, Alban, Li, Pen, Shivanna, Rakesh, Liu, Jeremiah, Welty, Chris, Lebron, Federico, Baddepudi, Anirudh, Krause, Sebastian, Parisotto, Emilio, Soricut, Radu, Xu, Zheng, Bloxwich, Dawn, Johnson, Melvin, Neyshabur, Behnam, Mao-Jones, Justin, Wang, Renshen, Ramasesh, Vinay, Abbas, Zaheer, Guez, Arthur, Segal, Constant, Nguyen, Duc Dung, Svensson, James, Hou, Le, York, Sarah, Milan, Kieran, Bridgers, Sophie, Gworek, Wiktor, Tagliasacchi, Marco, Lee-Thorp, James, Chang, Michael, Guseynov, Alexey, Hartman, Ale Jakse, Kwong, Michael, Zhao, Ruizhe, Kashem, Sheleem, Cole, Elizabeth, Miech, Antoine, Tanburn, Richard, Phuong, Mary, Pavetic, Filip, Cevey, Sebastien, Comanescu, Ramona, Ives, Richard, Yang, Sherry, Du, Cosmo, Li, Bo, Zhang, Zizhao, Iinuma, Mariko, Hu, Clara Huiyi, Roy, Aurko, Bijwadia, Shaan, Zhu, Zhenkai, Martins, Danilo, Saputro, Rachel, Gergely, Anita, Zheng, Steven, Jia, Dawei, Antonoglou, Ioannis, Sadovsky, Adam, Gu, Shane, Bi, Yingying, Andreev, Alek, Samangooei, Sina, Khan, Mina, Kocisky, Tomas, Filos, Angelos, Kumar, Chintu, Bishop, Colton, Yu, Adams, Hodkinson, Sarah, Mittal, Sid, Shah, Premal, Moufarek, Alexandre, Cheng, Yong, Bloniarz, Adam, Lee, Jaehoon, Pejman, Pedram, Michel, Paul, Spencer, Stephen, Feinberg, Vladimir, Xiong, Xuehan, Savinov, Nikolay, Smith, Charlotte, Shakeri, Siamak, Tran, Dustin, Chesus, Mary, Bohnet, Bernd, Tucker, George, von Glehn, Tamara, Muir, Carrie, Mao, Yiran, Kazawa, Hideto, Slone, Ambrose, Soparkar, Kedar, Shrivastava, Disha, Cobon-Kerr, James, Sharman, Michael, Pavagadhi, Jay, Araya, Carlos, Misiunas, Karolis, Ghelani, Nimesh, Laskin, Michael, Barker, David, Li, Qiujia, Briukhov, Anton, Houlsby, Neil, Glaese, Mia, Lakshminarayanan, Balaji, Schucher, Nathan, Tang, Yunhao, Collins, Eli, Lim, Hyeontaek, Feng, Fangxiaoyu, Recasens, Adria, Lai, Guangda, Magni, Alberto, De Cao, Nicola, Siddhant, Aditya, Ashwood, Zoe, Orbay, Jordi, Dehghani, Mostafa, Brennan, Jenny, He, Yifan, Xu, Kelvin, Gao, Yang, Saroufim, Carl, Molloy, James, Wu, Xinyi, Arnold, Seb, Chang, Solomon, Schrittwieser, Julian, Buchatskaya, Elena, Radpour, Soroush, Polacek, Martin, Giordano, Skye, Bapna, Ankur, Tokumine, Simon, Hellendoorn, Vincent, Sottiaux, Thibault, Cogan, Sarah, Severyn, Aliaksei, Saleh, Mohammad, Thakoor, Shantanu, Shefey, Laurent, Qiao, Siyuan, Gaba, Meenu, Chang, Shuo-yiin, Swanson, Craig, Zhang, Biao, Lee, Benjamin, Rubenstein, Paul Kishan, Song, Gan, Kwiatkowski, Tom, Koop, Anna, Kannan, Ajay, Kao, David, Schuh, Parker, Stjerngren, Axel, Ghiasi, Golnaz, Gibson, Gena, Vilnis, Luke, Yuan, Ye, Ferreira, Felipe Tiengo, Kamath, Aishwarya, Klimenko, Ted, Franko, Ken, Xiao, Kefan, Bhattacharya, Indro, Patel, Miteyan, Wang, Rui, Morris, Alex, Strudel, Robin, Sharma, Vivek, Choy, Peter, Hashemi, Sayed Hadi, Landon, Jessica, Finkelstein, Mara, Jhakra, Priya, Frye, Justin, Barnes, Megan, Mauger, Matthew, Daun, Dennis, Baatarsukh, Khuslen, Tung, Matthew, Farhan, Wael, Michalewski, Henryk, Viola, Fabio, Quitry, Felix de Chaumont, Lan, Charline Le, Hudson, Tom, Wang, Qingze, Fischer, Felix, Zheng, Ivy, White, Elspeth, Dragan, Anca, Alayrac, Jean-baptiste, Ni, Eric, Pritzel, Alexander, Iwanicki, Adam, Isard, Michael, Bulanova, Anna, Zilka, Lukas, Dyer, Ethan, Sachan, Devendra, Srinivasan, Srivatsan, Muckenhirn, Hannah, Cai, Honglong, Mandhane, Amol, Tariq, Mukarram, Rae, Jack W., Wang, Gary, Ayoub, Kareem, FitzGerald, Nicholas, Zhao, Yao, Han, Woohyun, Alberti, Chris, Garrette, Dan, Krishnakumar, Kashyap, Gimenez, Mai, Levskaya, Anselm, Sohn, Daniel, Matak, Josip, Iturrate, Inaki, Chang, Michael B., Xiang, Jackie, Cao, Yuan, Ranka, Nishant, Brown, Geoff, Hutter, Adrian, Mirrokni, Vahab, Chen, Nanxin, Yao, Kaisheng, Egyed, Zoltan, Galilee, Francois, Liechty, Tyler, Kallakuri, Praveen, Palmer, Evan, Ghemawat, Sanjay, Liu, Jasmine, Tao, David, Thornton, Chloe, Green, Tim, Jasarevic, Mimi, Lin, Sharon, Cotruta, Victor, Tan, Yi-Xuan, Fiedel, Noah, Yu, Hongkun, Chi, Ed, Neitz, Alexander, Heitkaemper, Jens, Sinha, Anu, Zhou, Denny, Sun, Yi, Kaed, Charbel, Hulse, Brice, Mishra, Swaroop, Georgaki, Maria, Kudugunta, Sneha, Farabet, Clement, Shafran, Izhak, Vlasic, Daniel, Tsitsulin, Anton, Ananthanarayanan, Rajagopal, Carin, Alen, Su, Guolong, Sun, Pei, V, Shashank, Carvajal, Gabriel, Broder, Josef, Comsa, Iulia, Repina, Alena, Wong, William, Chen, Warren Weilun, Hawkins, Peter, Filonov, Egor, Loher, Lucia, Hirnschall, Christoph, Wang, Weiyi, Ye, Jingchen, Burns, Andrea, Cate, Hardie, Wright, Diana Gage, Piccinini, Federico, Zhang, Lei, Lin, Chu-Cheng, Gog, Ionel, Kulizhskaya, Yana, Sreevatsa, Ashwin, Song, Shuang, Cobo, Luis C., Iyer, Anand, Tekur, Chetan, Garrido, Guillermo, Xiao, Zhuyun, Kemp, Rupert, Zheng, Huaixiu Steven, Li, Hui, Agarwal, Ananth, Ngani, Christel, Goshvadi, Kati, Santamaria-Fernandez, Rebeca, Fica, Wojciech, Chen, Xinyun, Gorgolewski, Chris, Sun, Sean, Garg, Roopal, Ye, Xinyu, Eslami, S. M. Ali, Hua, Nan, Simon, Jon, Joshi, Pratik, Kim, Yelin, Tenney, Ian, Potluri, Sahitya, Thiet, Lam Nguyen, Yuan, Quan, Luisier, Florian, Chronopoulou, Alexandra, Scellato, Salvatore, Srinivasan, Praveen, Chen, Minmin, Koverkathu, Vinod, Dalibard, Valentin, Xu, Yaming, Saeta, Brennan, Anderson, Keith, Sellam, Thibault, Fernando, Nick, Huot, Fantine, Jung, Junehyuk, Varadarajan, Mani, Quinn, Michael, Raul, Amit, Le, Maigo, Habalov, Ruslan, Clark, Jon, Jalan, Komal, Bullard, Kalesha, Singhal, Achintya, Luong, Thang, Wang, Boyu, Rajayogam, Sujeevan, Eisenschlos, Julian, Jia, Johnson, Finchelstein, Daniel, Yakubovich, Alex, Balle, Daniel, Fink, Michael, Agarwal, Sameer, Li, Jing, Dvijotham, Dj, Pal, Shalini, Kang, Kai, Konzelmann, Jaclyn, Beattie, Jennifer, Dousse, Olivier, Wu, Diane, Crocker, Remi, Elkind, Chen, Jonnalagadda, Siddhartha Reddy, Lee, Jong, Holtmann-Rice, Dan, Kallarackal, Krystal, Liu, Rosanne, Vnukov, Denis, Vats, Neera, Invernizzi, Luca, Jafari, Mohsen, Zhou, Huanjie, Taylor, Lilly, Prendki, Jennifer, Wu, Marcus, Eccles, Tom, Liu, Tianqi, Kopparapu, Kavya, Beaufays, Francoise, Angermueller, Christof, Marzoca, Andreea, Sarcar, Shourya, Dib, Hilal, Stanway, Jeff, Perbet, Frank, Trdin, Nejc, Sterneck, Rachel, Khorlin, Andrey, Li, Dinghua, Wu, Xihui, Goenka, Sonam, Madras, David, Goldshtein, Sasha, Gierke, Willi, Zhou, Tong, Liu, Yaxin, Liang, Yannie, White, Anais, Li, Yunjie, Singh, Shreya, Bahargam, Sanaz, Epstein, Mark, Basu, Sujoy, Lao, Li, Ozturel, Adnan, Crous, Carl, Zhai, Alex, Lu, Han, Tung, Zora, Gaur, Neeraj, Walton, Alanna, Dixon, Lucas, Zhang, Ming, Globerson, Amir, Uy, Grant, Bolt, Andrew, Wiles, Olivia, Nasr, Milad, Shumailov, Ilia, Selvi, Marco, Piccinno, Francesco, Aguilar, Ricardo, McCarthy, Sara, Khalman, Misha, Shukla, Mrinal, Galic, Vlado, Carpenter, John, Villela, Kevin, Zhang, Haibin, Richardson, Harry, Martens, James, Bosnjak, Matko, Belle, Shreyas Rammohan, Seibert, Jeff, Alnahlawi, Mahmoud, McWilliams, Brian, Singh, Sankalp, Louis, Annie, Ding, Wen, Popovici, Dan, Simicich, Lenin, Knight, Laura, Mehta, Pulkit, Gupta, Nishesh, Shi, Chongyang, Fatehi, Saaber, Mitrovic, Jovana, Grills, Alex, Pagadora, Joseph, Petrova, Dessie, Eisenbud, Danielle, Zhang, Zhishuai, Yates, Damion, Mittal, Bhavishya, Tripuraneni, Nilesh, Assael, Yannis, Brovelli, Thomas, Jain, Prateek, Velimirovic, Mihajlo, Akbulut, Canfer, Mu, Jiaqi, Macherey, Wolfgang, Kumar, Ravin, Xu, Jun, Qureshi, Haroon, Comanici, Gheorghe, Wiesner, Jeremy, Gong, Zhitao, Ruddock, Anton, Bauer, Matthias, Felt, Nick, GP, Anirudh, Arnab, Anurag, Zelle, Dustin, Rothfuss, Jonas, Rosgen, Bill, Shenoy, Ashish, Seybold, Bryan, Li, Xinjian, Mudigonda, Jayaram, Erdogan, Goker, Xia, Jiawei, Simsa, Jiri, Michi, Andrea, Yao, Yi, Yew, Christopher, Kan, Steven, Caswell, Isaac, Radebaugh, Carey, Elisseeff, Andre, Valenzuela, Pedro, McKinney, Kay, Paterson, Kim, Cui, Albert, Latorre-Chimoto, Eri, Kim, Solomon, Zeng, William, Durden, Ken, Ponnapalli, Priya, Sosea, Tiberiu, Choquette-Choo, Christopher A., Manyika, James, Robenek, Brona, Vashisht, Harsha, Pereira, Sebastien, Lam, Hoi, Velic, Marko, Owusu-Afriyie, Denese, Lee, Katherine, Bolukbasi, Tolga, Parrish, Alicia, Lu, Shawn, Park, Jane, Venkatraman, Balaji, Talbert, Alice, Rosique, Lambert, Cheng, Yuchung, Sozanschi, Andrei, Paszke, Adam, Kumar, Praveen, Austin, Jessica, Li, Lu, Salama, Khalid, Kim, Wooyeol, Dukkipati, Nandita, Baryshnikov, Anthony, Kaplanis, Christos, Sheng, XiangHai, Chervonyi, Yuri, Unlu, Caglar, Casas, Diego de Las, Askham, Harry, Tunyasuvunakool, Kathryn, Gimeno, Felix, Poder, Siim, Kwak, Chester, Miecnikowski, Matt, Dimitriev, Alek, Parisi, Aaron, Liu, Dangyi, Tsai, Tomy, Shevlane, Toby, Kouridi, Christina, Garmon, Drew, Goedeckemeyer, Adrian, Brown, Adam R., Vijayakumar, Anitha, Elqursh, Ali, Jazayeri, Sadegh, Huang, Jin, Carthy, Sara Mc, Hoover, Jay, Kim, Lucy, Kumar, Sandeep, Chen, Wei, Biles, Courtney, Bingham, Garrett, Rosen, Evan, Wang, Lisa, Tan, Qijun, Engel, David, Pongetti, Francesco, de Cesare, Dario, Hwang, Dongseong, Yu, Lily, Pullman, Jennifer, Narayanan, Srini, Levin, Kyle, Gopal, Siddharth, Li, Megan, Aharoni, Asaf, Trinh, Trieu, Lo, Jessica, Casagrande, Norman, Vij, Roopali, Matthey, Loic, Ramadhana, Bramandia, Matthews, Austin, Carey, CJ, Johnson, Matthew, Goranova, Kremena, Shah, Rohin, Ashraf, Shereen, Dasgupta, Kingshuk, Larsen, Rasmus, Wang, Yicheng, Vuyyuru, Manish Reddy, Jiang, Chong, Ijazi, Joana, Osawa, Kazuki, Smith, Celine, Boppana, Ramya Sree, Bilal, Taylan, Koizumi, Yuma, Xu, Ying, Altun, Yasemin, Shabat, Nir, Bariach, Ben, Korchemniy, Alex, Choo, Kiam, Ronneberger, Olaf, Iwuanyanwu, Chimezie, Zhao, Shubin, Soergel, David, Hsieh, Cho-Jui, Cai, Irene, Iqbal, Shariq, Sundermeyer, Martin, Chen, Zhe, Bursztein, Elie, Malaviya, Chaitanya, Biadsy, Fadi, Shroff, Prakash, Dhillon, Inderjit, Latkar, Tejasi, Dyer, Chris, Forbes, Hannah, Nicosia, Massimo, Nikolaev, Vitaly, Greene, Somer, Georgiev, Marin, Wang, Pidong, Martin, Nina, Sedghi, Hanie, Zhang, John, Banzal, Praseem, Fritz, Doug, Rao, Vikram, Wang, Xuezhi, Zhang, Jiageng, Patraucean, Viorica, Du, Dayou, Mordatch, Igor, Jurin, Ivan, Liu, Lewis, Dubey, Ayush, Mohan, Abhi, Nowakowski, Janek, Ion, Vlad-Doru, Wei, Nan, Tojo, Reiko, Raad, Maria Abi, Hudson, Drew A., Keshava, Vaishakh, Agrawal, Shubham, Ramirez, Kevin, Wu, Zhichun, Nguyen, Hoang, Liu, Ji, Sewak, Madhavi, Petrini, Bryce, Choi, DongHyun, Philips, Ivan, Wang, Ziyue, Bica, Ioana, Garg, Ankush, Wilkiewicz, Jarek, Agrawal, Priyanka, Guo, Danhao, Xue, Emily, Shaik, Naseer, Leach, Andrew, Khan, Sadh MNM, Wiesinger, Julia, Jerome, Sammy, Chakladar, Abhishek, Wang, Alek Wenjiao, Ornduff, Tina, Abu, Folake, Ghaffarkhah, Alireza, Wainwright, Marcus, Cortes, Mario, Liu, Frederick, Maynez, Joshua, Terzis, Andreas, Samangouei, Pouya, Mansour, Riham, Kępa, Tomasz, Aubet, François-Xavier, Algymr, Anton, Banica, Dan, Weisz, Agoston, Orban, Andras, Senges, Alexandre, Andrejczuk, Ewa, Geller, Mark, Santo, Niccolo Dal, Anklin, Valentin, Merey, Majd Al, Baeuml, Martin, Strohman, Trevor, Bai, Junwen, Petrov, Slav, Wu, Yonghui, Hassabis, Demis, Kavukcuoglu, Koray, Dean, Jeffrey, and Vinyals, Oriol

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence
Abstract

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Published
2024

5. Beyond Human Data: Scaling Self-Training for Problem-Solving with Language Models

Author

Singh, Avi, Co-Reyes, John D., Agarwal, Rishabh, Anand, Ankesh, Patil, Piyush, Garcia, Xavier, Liu, Peter J., Harrison, James, Lee, Jaehoon, Xu, Kelvin, Parisi, Aaron, Kumar, Abhishek, Alemi, Alex, Rizkowsky, Alex, Nova, Azade, Adlam, Ben, Bohnet, Bernd, Elsayed, Gamaleldin, Sedghi, Hanie, Mordatch, Igor, Simpson, Isabelle, Gur, Izzeddin, Snoek, Jasper, Pennington, Jeffrey, Hron, Jiri, Kenealy, Kathleen, Swersky, Kevin, Mahajan, Kshiteej, Culp, Laura, Xiao, Lechao, Bileschi, Maxwell L., Constant, Noah, Novak, Roman, Liu, Rosanne, Warkentin, Tris, Qian, Yundi, Bansal, Yamini, Dyer, Ethan, Neyshabur, Behnam, Sohl-Dickstein, Jascha, and Fiedel, Noah

Subjects
Computer Science - Machine Learning
Abstract

Fine-tuning language models~(LMs) on human-generated data remains a prevalent practice. However, the performance of such models is often limited by the quantity and diversity of high-quality human data. In this paper, we explore whether we can go beyond human data on tasks where we have access to scalar feedback, for example, on math problems where one can verify correctness. To do so, we investigate a simple self-training method based on expectation-maximization, which we call ReST$^{EM}$, where we (1) generate samples from the model and filter them using binary feedback, (2) fine-tune the model on these samples, and (3) repeat this process a few times. Testing on advanced MATH reasoning and APPS coding benchmarks using PaLM-2 models, we find that ReST$^{EM}$ scales favorably with model size and significantly surpasses fine-tuning only on human data. Overall, our findings suggest self-training with feedback can substantially reduce dependence on human-generated data., Comment: Accepted to TMLR. Camera-ready version. First three authors contributed equally

Published
2023

6. Frontier Language Models are not Robust to Adversarial Arithmetic, or 'What do I need to say so you agree 2+2=5?

Author

Freeman, C. Daniel, Culp, Laura, Parisi, Aaron, Bileschi, Maxwell L, Elsayed, Gamaleldin F, Rizkowsky, Alex, Simpson, Isabelle, Alemi, Alex, Nova, Azade, Adlam, Ben, Bohnet, Bernd, Mishra, Gaurav, Sedghi, Hanie, Mordatch, Igor, Gur, Izzeddin, Lee, Jaehoon, Co-Reyes, JD, Pennington, Jeffrey, Xu, Kelvin, Swersky, Kevin, Mahajan, Kshiteej, Xiao, Lechao, Liu, Rosanne, Kornblith, Simon, Constant, Noah, Liu, Peter J., Novak, Roman, Qian, Yundi, Fiedel, Noah, and Sohl-Dickstein, Jascha

Subjects
Computer Science - Computation and Language, Computer Science - Artificial Intelligence, Computer Science - Computers and Society, Computer Science - Machine Learning
Abstract

We introduce and study the problem of adversarial arithmetic, which provides a simple yet challenging testbed for language model alignment. This problem is comprised of arithmetic questions posed in natural language, with an arbitrary adversarial string inserted before the question is complete. Even in the simple setting of 1-digit addition problems, it is easy to find adversarial prompts that make all tested models (including PaLM2, GPT4, Claude2) misbehave, and even to steer models to a particular wrong answer. We additionally provide a simple algorithm for finding successful attacks by querying those same models, which we name "prompt inversion rejection sampling" (PIRS). We finally show that models can be partially hardened against these attacks via reinforcement learning and via agentic constitutional loops. However, we were not able to make a language model fully robust against adversarial arithmetic attacks.

Published
2023

7. Can Neural Network Memorization Be Localized?

Author

Maini, Pratyush, Mozer, Michael C., Sedghi, Hanie, Lipton, Zachary C., Kolter, J. Zico, and Zhang, Chiyuan

Subjects
Computer Science - Machine Learning, Computer Science - Computer Vision and Pattern Recognition
Abstract

Recent efforts at explaining the interplay of memorization and generalization in deep overparametrized networks have posited that neural networks $\textit{memorize}$ "hard" examples in the final few layers of the model. Memorization refers to the ability to correctly predict on $\textit{atypical}$ examples of the training set. In this work, we show that rather than being confined to individual layers, memorization is a phenomenon confined to a small set of neurons in various layers of the model. First, via three experimental sources of converging evidence, we find that most layers are redundant for the memorization of examples and the layers that contribute to example memorization are, in general, not the final layers. The three sources are $\textit{gradient accounting}$ (measuring the contribution to the gradient norms from memorized and clean examples), $\textit{layer rewinding}$ (replacing specific model weights of a converged model with previous training checkpoints), and $\textit{retraining}$ (training rewound layers only on clean examples). Second, we ask a more generic question: can memorization be localized $\textit{anywhere}$ in a model? We discover that memorization is often confined to a small number of neurons or channels (around 5) of the model. Based on these insights we propose a new form of dropout -- $\textit{example-tied dropout}$ that enables us to direct the memorization of examples to an apriori determined set of neurons. By dropping out these neurons, we are able to reduce the accuracy on memorized examples from $100\%\to3\%$, while also reducing the generalization gap., Comment: Accepted at ICML 2023

Published
2023

8. The Role of Pre-training Data in Transfer Learning

Author

Entezari, Rahim, Wortsman, Mitchell, Saukh, Olga, Shariatnia, M. Moein, Sedghi, Hanie, and Schmidt, Ludwig

Subjects
Computer Science - Computer Vision and Pattern Recognition, Computer Science - Machine Learning
Abstract

The transfer learning paradigm of model pre-training and subsequent fine-tuning produces high-accuracy models. While most studies recommend scaling the pre-training size to benefit most from transfer learning, a question remains: what data and method should be used for pre-training? We investigate the impact of pre-training data distribution on the few-shot and full fine-tuning performance using 3 pre-training methods (supervised, contrastive language-image and image-image), 7 pre-training datasets, and 9 downstream datasets. Through extensive controlled experiments, we find that the choice of the pre-training data source is essential for the few-shot transfer, but its role decreases as more data is made available for fine-tuning. Additionally, we explore the role of data curation and examine the trade-offs between label noise and the size of the pre-training dataset. We find that using 2000X more pre-training data from LAION can match the performance of supervised ImageNet pre-training. Furthermore, we investigate the effect of pre-training methods, comparing language-image contrastive vs. image-image contrastive, and find that the latter leads to better downstream accuracy

Published
2023

9. Leveraging Unlabeled Data to Track Memorization

Author

Forouzesh, Mahsa, Sedghi, Hanie, and Thiran, Patrick

Subjects
Computer Science - Machine Learning
Abstract

Deep neural networks may easily memorize noisy labels present in real-world data, which degrades their ability to generalize. It is therefore important to track and evaluate the robustness of models against noisy label memorization. We propose a metric, called susceptibility, to gauge such memorization for neural networks. Susceptibility is simple and easy to compute during training. Moreover, it does not require access to ground-truth labels and it only uses unlabeled data. We empirically show the effectiveness of our metric in tracking memorization on various architectures and datasets and provide theoretical insights into the design of the susceptibility metric. Finally, we show through extensive experiments on datasets with synthetic and real-world label noise that one can utilize susceptibility and the overall training accuracy to distinguish models that maintain a low memorization on the training set and generalize well to unseen clean data.

Published
2022

10. Layer-Stack Temperature Scaling

Author

Khalifa, Amr, Mozer, Michael C., Sedghi, Hanie, Neyshabur, Behnam, and Alabdulmohsin, Ibrahim

Subjects
Computer Science - Machine Learning, I.2.6, I.2.10
Abstract

Recent works demonstrate that early layers in a neural network contain useful information for prediction. Inspired by this, we show that extending temperature scaling across all layers improves both calibration and accuracy. We call this procedure "layer-stack temperature scaling" (LATES). Informally, LATES grants each layer a weighted vote during inference. We evaluate it on five popular convolutional neural network architectures both in- and out-of-distribution and observe a consistent improvement over temperature scaling in terms of accuracy, calibration, and AUC. All conclusions are supported by comprehensive statistical analyses. Since LATES neither retrains the architecture nor introduces many more parameters, its advantages can be reaped without requiring additional data beyond what is used in temperature scaling. Finally, we show that combining LATES with Monte Carlo Dropout matches state-of-the-art results on CIFAR10/100., Comment: 10 pages, 7 figures, 3 tables

Published
2022

11. Teaching Algorithmic Reasoning via In-context Learning

Author

Zhou, Hattie, Nova, Azade, Larochelle, Hugo, Courville, Aaron, Neyshabur, Behnam, and Sedghi, Hanie

Subjects
Computer Science - Machine Learning, Computer Science - Artificial Intelligence, Computer Science - Computation and Language
Abstract

Large language models (LLMs) have shown increasing in-context learning capabilities through scaling up model and data size. Despite this progress, LLMs are still unable to solve algorithmic reasoning problems. While providing a rationale with the final answer has led to further improvements in multi-step reasoning problems, Anil et al. 2022 showed that even simple algorithmic reasoning tasks such as parity are far from solved. In this work, we identify and study four key stages for successfully teaching algorithmic reasoning to LLMs: (1) formulating algorithms as skills, (2) teaching multiple skills simultaneously (skill accumulation), (3) teaching how to combine skills (skill composition) and (4) teaching how to use skills as tools. We show that it is possible to teach algorithmic reasoning to LLMs via in-context learning, which we refer to as algorithmic prompting. We evaluate our approach on a variety of arithmetic and quantitative reasoning tasks, and demonstrate significant boosts in performance over existing prompting techniques. In particular, for long parity, addition, multiplication and subtraction, we achieve an error reduction of approximately 10x, 9x, 5x and 2x respectively compared to the best available baselines.

Published
2022

12. REPAIR: REnormalizing Permuted Activations for Interpolation Repair

Author

Jordan, Keller, Sedghi, Hanie, Saukh, Olga, Entezari, Rahim, and Neyshabur, Behnam

Subjects
Computer Science - Machine Learning, Computer Science - Artificial Intelligence, Computer Science - Computer Vision and Pattern Recognition, Statistics - Machine Learning
Abstract

In this paper we look into the conjecture of Entezari et al. (2021) which states that if the permutation invariance of neural networks is taken into account, then there is likely no loss barrier to the linear interpolation between SGD solutions. First, we observe that neuron alignment methods alone are insufficient to establish low-barrier linear connectivity between SGD solutions due to a phenomenon we call variance collapse: interpolated deep networks suffer a collapse in the variance of their activations, causing poor performance. Next, we propose REPAIR (REnormalizing Permuted Activations for Interpolation Repair) which mitigates variance collapse by rescaling the preactivations of such interpolated networks. We explore the interaction between our method and the choice of normalization layer, network width, and depth, and demonstrate that using REPAIR on top of neuron alignment methods leads to 60%-100% relative barrier reduction across a wide variety of architecture families and tasks. In particular, we report a 74% barrier reduction for ResNet50 on ImageNet and 90% barrier reduction for ResNet18 on CIFAR10.

Published
2022

13. Understanding the effect of sparsity on neural networks robustness

Author

Timpl, Lukas, Entezari, Rahim, Sedghi, Hanie, Neyshabur, Behnam, and Saukh, Olga

Subjects
Computer Science - Computer Vision and Pattern Recognition
Abstract

This paper examines the impact of static sparsity on the robustness of a trained network to weight perturbations, data corruption, and adversarial examples. We show that, up to a certain sparsity achieved by increasing network width and depth while keeping the network capacity fixed, sparsified networks consistently match and often outperform their initially dense versions. Robustness and accuracy decline simultaneously for very high sparsity due to loose connectivity between network layers. Our findings show that a rapid robustness drop caused by network compression observed in the literature is due to a reduced network capacity rather than sparsity.

Published
2022

14. Leveraging Unlabeled Data to Predict Out-of-Distribution Performance

Author

Garg, Saurabh, Balakrishnan, Sivaraman, Lipton, Zachary C., Neyshabur, Behnam, and Sedghi, Hanie

Subjects
Computer Science - Machine Learning, Statistics - Machine Learning
Abstract

Real-world machine learning deployments are characterized by mismatches between the source (training) and target (test) distributions that may cause performance drops. In this work, we investigate methods for predicting the target domain accuracy using only labeled source data and unlabeled target data. We propose Average Thresholded Confidence (ATC), a practical method that learns a threshold on the model's confidence, predicting accuracy as the fraction of unlabeled examples for which model confidence exceeds that threshold. ATC outperforms previous methods across several model architectures, types of distribution shifts (e.g., due to synthetic corruptions, dataset reproduction, or novel subpopulations), and datasets (Wilds, ImageNet, Breeds, CIFAR, and MNIST). In our experiments, ATC estimates target performance $2$-$4\times$ more accurately than prior methods. We also explore the theoretical foundations of the problem, proving that, in general, identifying the accuracy is just as hard as identifying the optimal predictor and thus, the efficacy of any method rests upon (perhaps unstated) assumptions on the nature of the shift. Finally, analyzing our method on some toy distributions, we provide insights concerning when it works. Code is available at https://github.com/saurabhgarg1996/ATC_code/., Comment: Accepted at ICLR 2022

Published
2022

Catalog

Books, media, physical & digital resources
See catalog results