Research

Do you have time for a quick chat? Designing a conversational interface for sexual harassment prevention training

Sexual harassment (SH) incidents are increasing and call into question the effectiveness of traditional SH prevention training. In this paper, we introduce a proof-of-concept design of a conversational interface (CI) for understanding SH cases. Key features of the interface include that it engages the learner in a dyadic conversation, prompts the learner for guidance, and tells a story of SH from a first-person perspective. From a mixed-methods study (N=32), learners experiencing a SH vignette using the conversational interface reported feeling less overwhelmed with the content, more engaged with the situation, and more comfortable discussing the topic compared to reading the same vignette online. Participants also reported that using a first-person narrative made the vignette feel realistic and relatable. However, there was no difference in empathy between the conditions.

Evaluating How Fine-tuning on Bimodal Data Effects Code Generation

Despite the increase in popularity of language models for code generation, it is still unknown how training on bimodal coding forums affects a model's code generation performance and reliability. We, therefore, collect a dataset of over 2.2M StackOverflow questions with answers for finetuning. These fine-tuned models have average $pass@k$ improvements of 54.64% and 85.35% on the HumanEval (Chen et al., 2021) and Mostly Basic Program Problems (Austin et al., 2021) tasks, respectively. This regime further decreases the number of generated programs with both syntax and runtime errors. However, we find that at higher temperatures, there are significant decreases to the model's ability to generate runnable programs despite higher $pass@k$ scores, underscoring the need for better methods of incorporating such data that mitigate these side effects.

Performance Evaluation of MaxViT in Cosmological Parameter Inference

The goal of this work is to compare the performance of two models on the task of cosmological parameter inference using CAMELS, a large-scale dataset of cosmological simulations. The first model is a Learnable Scattering Transform[1]. The second model is a Multi-Axis Vision Transform (MaxViT), which consists in a combination of convolutional neural networks and vision transformers. In the work, the images used are two-dimensional CAMELS maps, which are slices taken from the original three-dimensional maps. The parameters are ΩM, σ8, ASN1, ASN2, AAGN1, AAGN2. The first two are cosmological parameters and the next four are astrophysical parameters that control the efficiency of the supernova and Active Galactic Nuclei (AGN) feedback. In order to test the performance of MaxViT, the trainable parameters were optimized to estimate ΩM and σ8. The scattering network hybrid model utilizing wavelets we are trying to improve upon had an error of 2.199% for ΩM and 1.670% for σ8. Meanwhile, MaxViT had an error of 2.016% for ΩM and 1.518% for σ8- not improving upon the Wavelet model, but getting rather close with a possibility of surpassing with further tuning. Similar to the Wavelet model, the MaxViT model poorly constrained the four efficiency/feedback parameters, with further improvement possibly coming from a combination of a different algorithm specifically for those four inferences.

the link to paper here

Click consonant phonotactic learning: An investigation of phonotactic learning with click consonants in English speakers

Many African languages use clicks as consonants and in this study, we investigated phonotactic learning by native English speakers of first-order constraints with two types of clicks, dental and alveolar. Specifically, we are interested in whether people can implicitly learn phonotactic constraints involving sounds that do not occur in their native language. Learning was measured through the types of speech errors produced. If these errors tend to obey the experiment’s rule, learning has taken place. However, if they don’t, learning has not taken place. The experiment involved participants reciting consonant-vowel-consonant syllables on two different days and the rules involved whether a click can occur in the beginning or end of a syllable. The first day was a practice/ training session and the second day was the actual experiment. The results show that English speakers did learn the click positions when compared to the regular-consonants through restricted and unrestricted legality errors. But clicks do not stick to their positions as much as the regular consonants do. This implies that English speakers do not treat clicks like other consonants.

the link to paper here

Pruning Pre-Trained BERT For Efficient Transfer Learning

While a plethora of network pruning algorithms have been developed for the vision domain, pruning bulky Large Language Models (LLMs) has been generally overlooked, partly due to the novelty of these models. At the same time, finetuning pre-trained LLMs has become a state-of-the-art approach for many downstream language tasks, such as text classification. In this study, we develop a flexible PyTorch codebase with clear API for experimenting with pruning LLMs and finetuning them on text classification tasks. Using this module, we benchmark severa recent pruning algorithms applied to a pretrained BERT for CoLa, SST-2 and IMDb classification tasks. We extend our investigation to higher than usual compression rates (100°ø and above) and utilize the effective sparsity framework to ensure accurate results.

the link to paper here

Page updated

Google Sites

Report abuse