JADS Doctoral Symposium

Presentation title
Efficient Focus for Autonomous Agents: on Generalization in Deep RL

Abstract 
The world provides an abundance of information, but often many inputs to a reinforcement learning agent are irrelevant or even distracting to its current task. We cover two approaches aimed at enhancing performance in challenging environments. First, the Automatic Noise Filtering (ANF) method is introduced, which learns to focus its efficient sparse connectivity on task-relevant features, outperforming dense algorithms with significantly fewer weights. Second, we highlight MaDi, an algorithm that learns to mask distractions in pixel-based reinforcement learning, improving focus and generalization.

Presentation title
Predicting Body Measurements from Images: Malnutrition In Children 

Abstract 
Approximately 800 million people in the world go to bed hungry; 200 million of them are children and 5.6 million avoidable infant deaths are caused by malnutrition (according to statistics of the WFP, the UN, and the World Bank ). Timely detection of child malnutrition through adequate measurements is crucial for treatment. World Health Organisation statistics show that only 35% of malnourished children are adequately measured. Manually measuring height and weight is time consuming and often inaccurate. Precisely estimating a child’s body measurements and weight from a single image will allow us to overcome the challenges of manually measuring children. However, training a model with children’s images and body measurements comes with ethical responsibility in terms of careful handling and processing of data, as well as explainability of these models and how the decisions inside these models are made. The first last of the presentation will cover the current results and progress in terms of data acquisition and model training and will present a detailed plan on how we intend to provide some insights into how these models work. The second part of the presentation will be reporting on our current ongoing work on explainable AI, to look at how these models work and make decisions. The third and final presentation will serve as an open forum to discuss and gather new thoughts and how we can go about such challenging technical task of predicting body measurements from images of vulnerable subjects, which is ethically concerning, but if performed successfully can have a huge social impact to save millions of lives worldwide.

Presentation title

A population-based study of the Netherlands on the effects of poverty on child and maternal perinatal health.

Abstract 

Nearly 16% percent of children around the world live in poverty and those born into it are especially susceptible to its effects. Although poverty is consistently linked to adverse health outcomes, it is a heterogenous phenomenon and depending on the context its associated risk factors can differ both in nature and in severity. Interventions during the first years of life can be both effective and cost efficient. Given the complexity of context-specific poverty risk factors and their interactions with each other, the current project takes a data-driven approach to identifying intervention targets that can aid policymakers. The goal of the study is to investigate the association between family poverty and maternal and infant health in The Netherlands, as well as potential mediating intervention targets. Method: Data used for this study is part of a large population-based dataset consisting of records from official healthcare providers and the Central Bureau for Statistics in The Netherlands. It covers a time period of 10 years (2010-2020) with 1.8 million pregnancies. We take a two-step approach to the analysis to address the goals of the research question – 1) a descriptive analysis of relevant poverty-related factors in regards to child health; 2) causal pre-registered confirmatory mediation analyses of theory-informed mediators of the effects of poverty on child birth outcomes. After a conceptual pre-registration of the hypotheses we split 10% of the data for feasibility assessment of the planned analyses. Following the 10% holdout sample, an analytic pre-registration of the methods will be submitted, after which analyses will be conducted on the remaining 90% of the cases. Additional robustness tests will also be conducted, as well as analyses replicating the results using different variables to account for income, such as education, which can often be the only proxy for socioeconomic status that researchers have access to.

Presentation title
ChatGPT (Cover) Letter for Me: The Role of LLMs in Recruitment 

Abstract 
We examine the effect of Large Language Models (LLMs) on job application processes. Through two experiments, we determine whether LLMs enhance the quality of applicants’ cover letters and how recruiters respond to their use. The first experiment involves university students and recruiters from four multinational companies, showing that LLMs increase the quality of cover letters, benefiting especially applicants with lower abilities. Despite these improvements, the likelihood of securing an interview does not increase, as LLMs improve less essential parts of cover letters. During the second experiment, recruiters assess a sample of cover letters, some crafted with LLM assistance and others without. Awareness of LLM involvement in writing does not change the cover letter evaluations. However, recruiters reward high-quality applicants when it is clear that LLMs were not employed. Finally, we develop a theoretical model to study the general equilibrium implications of our findings. 

Presentation title
Learning from machines: can CEO vocal cues help to predict future firm performance? 

Abstract 
“I apply tailored deep learning models on CEO voiceprints of earnings conference calls to predict the firm’s future performance as measured by cumulative gross returns and cumulative abnormal returns. The out-of-sample evaluations of the models achieve a prediction accuracy ranging from 53.2% to 61.1%. This study documents new empirical evidence that audio recordings of conference calls contain valuable information about a firm’s returns, incremental to qualitative “soft” information conveyed by textual content, and quantitative earnings information. To achieve some level of explainability of the learned voiceprints’ nuances, I refer to a three-dimensional vocal emotion classifier – arousal, valence, and dominance – which contributes to improving and refining vocal sentiment analysis in the existing literature. I find a positive correlation between the valence of a CEO’s speech and the firm performance. Finally, the novel empirical results of predicting market returns using the CEO’s vocal cues indicate that the CEO’s vocal cues may capture underlying macroeconomic sentiment.(JEL D83, G14, G41).”

Presentation title
Computer Vision and Semantic Segmentation applied to Precision Agriculture.

Abstract 

In the context of precision agriculture, drones equipped with computer vision systems are becoming essential for efficient crop management. This talk explores the task of Weed Mapping, which involves the detection and spatial mapping of weed infestations using drone-acquired data. Key challenges in this domain include handling multispectral data, overcoming the scarcity of annotated datasets, and addressing the computational limitations of drones. To tackle these issues, the presentation will examine the use of transfer learning and few-shot learning techniques to mitigate data scarcity, as well as the application of knowledge distillation to optimize models for deployment on resource-constrained devices. These approaches aim to advance the scalability and effectiveness of weed mapping, providing actionable insights to enhance agricultural productivity and sustainability.