Over the past few years, Federated Learning (FL) has become an emerging machine learning technique to tackle data privacy challenges through collaborative training. In algorithm, clients submit a locally trained model, and server aggregates these parameters until convergence. Despite significant efforts that have been made FL in fields like computer vision, audio, natural language processing, applications utilizing multimodal streams remain largely unexplored. It is known broad real-world...

As the world changes, we need to be able update our models and correct false information without costly retraining. Knowledge-based model editing enables precise modifications weights of large language in order modify encoded within. Recent approaches have seen success enabling recall edited for thousands edits at once. However, these fail produce that account associated contextual information. We present K-Edit, an effective approach generating contextually consistent knowledge edits. By...

Children with Autism Spectrum Disorder (ASD) are known to have difficulty in producing and perceiving emotional facial expressions. Their expressions often perceived as atypical by adult observers. This paper focuses on data driven ways analyze quantify atypicality of children ASD. Our objective is uncover those characteristics gestures that induce the sense Using a carefully collected motion capture database, without ASD compared within six basic emotion categories employing methods from...

Anil Ramakrishna, Victor R. Martínez, Nikolaos Malandrakis, Karan Singla, Shrikanth Narayanan. Proceedings of the 55th Annual Meeting Association for Computational Linguistics (Volume 1: Long Papers). 2017.

Atypical speech prosody is a primary characteristic of autism spectrum disorders (ASD), yet it often excluded from diagnostic instrument algorithms due to poor subjective reliability. Robust, objective prosodic cues can enhance our understanding those aspects which are atypical in autism. In this work, we connect signal-derived descriptors perceptions awkwardness. Subjectively, more awkward less expressive (more monotone) and has perceived rate/rhythm, volume, intonation. We also find...

Violent content in movies can influence viewers’ perception of the society. For example, frequent depictions certain demographics as perpetrators or victims abuse shape stereotyped attitudes. In this work, we propose to characterize aspects violent solely from language used scripts. This makes our method applicable a movie earlier stages creation even before it is produced. complementary previous works which rely on audio video post production. Our approach based broad range features...

We describe a unified active transfer learning framework called Hierarchical Active Transfer Learning (HATL). HATL exploits cluster structure shared between different data domains to perform by imputing labels for unlabeled target and generate effective label queries during learning. The resulting is flexible enough not only adaptive accelerated but also unsupervised semi-supervised derive an intuitive useful upper bound on HATL's error when used infer points. present results synthetic that...

Direct content analysis reveals important details about movies including those of gender representations and potential biases.We investigate the differences between male female character depictions in movies, based on patterns language used.Specifically, we use an automatically generated lexicon linguistic norms characterizing ladenness.We multivariate to correlate them with elements movie production.The proposed metric differentiates utterances exhibits some interesting interactions genres...

Humor is an important social construct that serves several roles in human communication. Though subjective, it culturally ubiquitous and often used to diffuse tension, specially intense conversations such as those psychotherapy sessions. Automatic recognition of humor has been considerable interest the natural language processing community thanks its relevance conversational agents. In this work, we present a model for Motivational Interviewing based We use Long Short Term Memory (LSTM)...

Rahul Sharma, Anil Ramakrishna, Ansel MacLaughlin, Anna Rumshisky, Jimit Majmudar, Clement Chung, Salman Avestimehr, Gupta. Proceedings of the 2022 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies. 2022.

Federated Learning (FL) is a popular algorithm to train machine learning models on user data constrained edge devices (for example, mobile phones) due privacy concerns. Typically, FL trained with the assumption that no part of can be egressed from edge. However, in many production settings, specific data-modalities/meta-data are limited device while others not. For commercial SLU systems, it typically desired prevent transmission biometric signals (such as audio recordings input prompt)...

Decoding methods for large language models (LLMs) usually struggle with the tradeoff between ensuring factuality and maintaining diversity. For example, a higher p threshold in nucleus (top-p) sampling increases diversity but decreases factuality, vice versa. In this paper, we propose REAL (Residual Entropy from Asymptotic Line) sampling, decoding method that achieves improved over by predicting an adaptive of $p$. Specifically, predicts step-wise likelihood LLM to hallucinate, lowers when...

In this work, we introduce the Learnable Response Scoring Function (LARS) for Uncertainty Estimation (UE) in generative Large Language Models (LLMs). Current scoring functions probability-based UE, such as length-normalized and semantic contribution-based weighting, are designed to solve specific aspects of problem but exhibit limitations, including inability handle biased probabilities under-performance low-resource languages like Turkish. To address these issues, propose LARS, a function...

Knowledge graphs (KGs) complement Large Language Models (LLMs) by providing reliable, structured, domain-specific, and up-to-date external knowledge. However, KGs LLMs are often developed separately must be integrated after training. We introduce Tree-of-Traversals, a novel zero-shot reasoning algorithm that enables augmentation of black-box with one or more KGs. The equips LLM actions for interfacing KG the to perform tree search over possible thoughts find high confidence paths. evaluate...

We propose a constraint learning schema for fine-tuning Large Language Models (LLMs) with attribute control. Given training corpus and control criteria formulated as sequence-level on model outputs, our method fine-tunes the LLM while enhancing satisfaction minimal impact its utility generation quality. Specifically, approach regularizes by penalizing KL divergence between desired output distribution, which satisfies constraints, LLM's posterior. This regularization term can be approximated...

Machine unlearning has been used to remove unwanted knowledge acquired by large language models (LLMs). In this paper, we examine machine from an optimization perspective, framing it as a regularized multi-task problem, where one task optimizes forgetting objective and another the model performance. particular, introduce normalized gradient difference (NGDiff) algorithm, enabling us have better control over trade-off between objectives, while integrating new, automatic learning rate...

Contrastive decoding (CD) (Li et al., 2023) improves the next-token distribution of a large expert language model (LM) using small amateur LM. Although CD is applied to various LMs and domains enhance open-ended text generation, it still unclear why often works well, when could fail, how we can make better. To deepen our understanding CD, first theoretically prove that be viewed as linearly extrapolating logits from huge hypothetical We also highlight linear extrapolation unable output most...