Checking in (Give t-shirts and swag)

Q&A

This document provides essential concepts, materials, and examples on Quantum Natural Language Processing. It is a recent topic, and the literature is limited. Therefore, we discuss the main ideas on QNLP and its potential applications. This paper brings quantum computing concepts applied to NLP with different purposes; in particular, we focus on showing basic concepts and implementation. We also discuss some potential QNLP applications.

Unbiased representation learning is still an object of study under specific applications and contexts. Novel architectures are usually crafted to resolve particular problems using mixtures of fundamental pieces. This paper presents different image feature extraction mechanisms that work together with residual connections to encode perceptual image information in an autoencoder configuration. We use data that aims to address further issues regarding criminal activity in consumer-to-consumer online platforms. Preliminary results suggest that the proposed architecture can learn rich spaces using ours and other image datasets.

Within the field of Cultural Heritage, image inpainting is a conservation process that fills in missing or damaged parts of an artwork to present a complete image. Multi-modal diffusion models have brought photo-realistic results on image inpainting where content can be generated by using descriptive text prompts. However, these models fail to produce content consistent with the painter's artistic style and period, being unsuitable for the reconstruction of fine arts and requiring laborious expert judgement. This work presents a methodology to improve the inpainting process by automating the selection process. We propose a discriminator model that processes the output of inpainting models and assigns a probability that indicates the likelihood that the restored image belongs to a certain painter.

Local differential privacy has been gaining popularity in both academic and industrial settings as an effective mechanism to enable the computation of summary statistics by service providers while providing privacy guarantees to end users. In the proposed mechanisms of local differential privacy, the introduction of noise in the user data is key to preserve privacy, but can greatly limit the estimation power on the provider side. In this paper we propose to include pre-filters based on machine learning models to help discard observations that may be too noisy to add value to the estimation process. We test our approach on both synthetic and real datasets, and identify average reductions of up to 31% in the mean squared error.

Q&A for the first block of oral presentations.

AM Break

Q&A

Curriculum Learning is a powerful training method that allows for faster and better training in some settings. This method, however, requires having a notion of which examples are difficult and which are easy, which is not always trivial to provide. A recent metric called C-Score acts as a proxy for example difficulty by relating it to learning consistency. Unfortunately, this method is quite compute intensive which limits its applicability for alternative datasets. In this work, we train models through different methods to predict C-Score for CIFAR-100 and CIFAR-10. We find, however, that these models generalize poorly both within the same distribution as well as out of distribution. This suggests that C-Score is not defined by the individual characteristics of each sample but rather by other factors. We hypothesize that a sample’s relation to its neighbors, in particular, how many of them share the same labels, can help in explaining C-Scores. We plan to explore this in future work.

How to mine opinions from pre-trained and general-purpose models in different domain applications? Answering this question is a recent research challenge since most existing methods fail when applied in domains other than those trained. In this paper, we introduce heterogeneous opinion mining networks as a strategy to combine opinion data from different domains, as well as integrate various models trained for different applications. We discuss how such knowledge results in different nodes and links in heterogeneous networks. We also investigated the mapping of the network nodes to a unified embedding space to allow vector-space-based machine learning models. We evaluate our proposal in a link prediction task to show that nodes and relationships in one domain can predict relevant knowledge in other domains.

Most existing methods focus on sentiment analysis of textual data. However, recently there has been a massive use of images and videos on social platforms, motivating sentiment analysis from other modalities. Current studies show that exploring other modalities (e.g., images) increases sentiment analysis performance. State-of-the-art multimodal models, such as CLIP and VisualBERT, are pre-trained on datasets with the text paired with images. Although the results obtained by these models are promising, pre-training and sentiment analysis fine-tuning tasks of these models are computationally expensive. This paper introduces a transfer learning approach using joint fine-tuning for sentiment analysis. Our proposal achieved competitive results using a more straightforward alternative fine-tuning strategy that leverages different pre-trained unimodal models and efficiently combines them in a multimodal space. Moreover, our proposal allows flexibility when incorporating any pre-trained model for texts and images during the joint fine-tuning stage, being especially interesting for sentiment classification in low-resource scenarios.

Q&A for second block of oral presentations.

Visit the booths from our sponsors

Speech recognition offers promising benefits for business and personal applications. Although automatic speech recognition systems have evolved significantly with deep learning methods, it remains an open research problem. In many languages there is still a shortage of open/public resources, resulting in low quality automatic speech recognition systems. In this talk, a multi-speaker text-to-speech (TTS) system is described for scenarios with few available speakers. Exploring flow-based and multilingual models, it is possible to leverage data from languages with many available speakers and make it viable for those languages with less data. Additionally, we show how this model can be applied to improve automatic speech recognition (ASR) systems in two target languages, simulating a scenario with only one speaker available. This allows for many applications such as the deployment of TTS for people with voice disorders, or ASR in extremely small data available such as native and regional languages.

Q&A for Moacir

Q&A

Afternoon Break

Q&A

Q&A

In person poster session is joint with Women in Machine Learning.

Poster Titles An Introduction to Quantum Natural Language Processing and a Study Case Bottleneck-based Encoder-decoder ARchitecture (BEAR) for Learning Unbiased Consumer-to-Consumer Image Representations Classifier Guided Diffusion for Image Inpainting: Applications to Fine Art A Model-Based Filter to Improve Local Differential Privacy A Study on the Predictability of Sample Learning Consistency Link Prediction from Heterogeneous Opinion Mining Networks with Multi-Domain Applications Transfer Learning with Joint Fine-Tuning for Multimodal Sentiment Analysis Interpretable Process Mining Opportunities and Challenges

Virtual poster session is via GatherTown [ protected link dropped ]