Mahdi Namazifar on LinkedIn: We still have 1 open internship position in Amazon AGI. If someone you… | 12 comments (2024)

Learning from human feedback is critical in training LLMs. RLHF is focused on using binary human feedback. However human feedback does not have to be limited to such binary feedback (preference). In our most recent paper we show learning from human feedback that is in natural language is highly efficient and effective. In this work that we call Critique and Revise (CnR) we show that with 1000 critique examples a 40b parameter LLM could be trained to improve the responses from even ChatGPT!https://lnkd.in/ggpKghYN

203

If you're interested in topics related to controllability of LLMs, make sure to check this workshop out!

19

Our KILM paper was accepted at ACL 2023! In this collaboration with Yan XU, Devamanyu Hazarika, Di Jin, Aishwarya Padmakumar, Yang Liu, and Dilek Hakkani-Tur we show how atomic pieces of knowledge could be injected into a pre-trained large language model. Paper: https://lnkd.in/gRS9up3p Code: https://lnkd.in/gApv3zdz Amazon Science

152

10 Comments

At #Alexa AI, with Devamanyu Hazarika and Dilek Hakkani-Tur we recently published a new paper that studies the role of bias terms in dot-product attention (the one used in Transformers). In this work we show that the bias term for the "key" linear transformation is redundant. Moreover, we show that the bias term for the "value" linear transformation has a more prominent role compared to the "query" linear transformation. #amazonsciencehttps://lnkd.in/gTDrTGe6

75

3 Comments

At #Alexa AI we recently published two of our papers that are accepted at EMNLP2022. The first work studies prefix tuning from the lens of kernels and proposes an efficient adaptation approach for transformers. The second work studies the role of language instructions in action learning in human robot interactions. I'm grateful for such amazing collaborators! #amazonscience https://lnkd.in/gNGw4ynxhttps://lnkd.in/gyZ3uH9p

51

2 Comments

We recently published two of our papers that are accepted at #NAACL22. In the first work we propose a novel approach to capture the contents of a long document in a graph and later use that graph in response generation. In the second work we propose a more efficient attention adaptation approach based on kernels theory. I'm immensely thankful for being a part of such an amazing team led by Dilek Hakkani-Tur at #Amazon #Alexa AI! Also I'm grateful for the opportunity of working with each and every one of the co-authors of these two works! #AmazonScience1) https://lnkd.in/eicBa5ns2) https://lnkd.in/egM9T4Xq

83

8 Comments

Our work on zero-shot control of Natural Language Generation (NLG) was accepted at AAAI 2022. Extended preprint: https://lnkd.in/gRc-EK4In this work we show that 1) attention could be biased in zero-shot without catastrophic forgetting and 2) NLG could be controlled in zero-shot with attention biasing and context augmentation. It's been great learning experience to work with Devamanyu Hazarika and Dilek Hakkani-Tur on this paper as part of the big #AlexaAI family! #AmazonScience

66

2 Comments

In our most recent work we show how a trained Natural Language Generation (NLG) model could be manually manipulated for zero-shot control! We also report some unexpected behaviors from Transformers that could be leveraged for efficient training of NLG models. With Devamanyu Hazarika and Dilek Hakkani-Turhttps://lnkd.in/gRc-EK4

57

In our most recent work with @JohnMalik LI Erran Li Gokhan Tur Dilek Hakkani-Tur we show how Warped Language Models (WLM) could be used for correcting audio transcriptions. We show that even human transcription could be improved by up to 10% using WLMs. https://lnkd.in/gjsYT2K

57