Pitch Workshop - Tuesday 6 and 13 July 9h00

5 julho 2021, 07:51 • Ana Almeida Matos

We will have two sessions of our Pitch Workshop:

Tuesday 6 July 9h00

Title: Search-Oriented Conversational Assistant
Presenter: Gonçalo Raposo
Abstract:
State-of-the-art dialogue models often produce factually inaccurate responses. Transformers may be fine-tuned for tasks such as response generation, and are able to produce fluent and well-written results, due to the very large amount of text they are exposed to during pre-training. However, generated responses tend to suffer from factual incorrectness and knowledge hallucination. These problems often arise because the models only consider the given conversation, and thus any knowledge present in the generated response comes implicitly from the model parameters. This work aims to introduce a retrieval step that will search for passages related to the given utterance and explicitly use them to generate a response. The PEGASUS model, i.e. a state-of-the-art Transformer for text summarization, is fine-tuned to address answer generation as a task of summarizing the retrieved passages, conditioned on the current conversation. A few conversational datasets are considered for experiments, as well as a community support dataset, in order to evaluate the system in a customer support scenario. The obtained results show that the system is able to make use of the retrieved knowledge to generate consistent and factually accurate responses. Moreover, by relying on a retrieval stage, the system also provides more interpretable responses.

Title: Towards Dataset Comparability: An Approach based on User Behavior
Presenter: João Góis
Abstract:
A current concern in today's society is to mitigate the risk of global climate change. Although there have been several initiatives to achieve more sustainable worldwide energy distribution, improper energy use remains an issue. In this work, a new methodology is proposed for detecting and analyzing energy consumption in buildings. For illustration, the methods are applied to some appliances of the REFIT dataset. The proposed methods enable a straightforward and rigorous distinction of different consumption patterns and, consequently, the definition of user profiles for each building throughout the seasons of the year.

Title: Learning prognostic biomarkers from three-dimensional biomedical data of psychiatric disorders
Presenter: Leonardo Duarte Rodrigues Alexandre
Abstract:
The number of patients diagnosed with a mental disorder (depression, attention deficit and hyperactivity disorder, anxiety, bipolar disorder, and/or schizophrenia) is considerably rising. Due to the isolation of the population during the pandemic, this number is especially high. Thus, the need for prognostic markers is essential to place appropriate diagnostics and treat patients with the appropriate therapies in accordance with their unique neurobiological profile. This treatment can be critical to prevent morbidity and, in some cases mortality. Despite being essential to diagnose these patients correctly, this task is hampered due to many symptoms overlapping between mental disorders. Thus, my PhD’s aim is to develop machine learning approaches to identify prognostic biomarkers in psychiatric disorders and support therapeutic choices from cohorts with available neuroimaging, cognitive and molecular data. This type of data is typically consolidated using a three-dimensional tensor representation with the dimensions being patients-variables-time. Techniques such as triclustering and temporal pattern mining, which remain largely unexplored within mental disorder research papers, will be used to explore this three-dimensional space to discover meaningful patterns. We will then proceed to: 1) understand the extent to which three-dimensional patterns discovered from neurobiological data assist the understanding of complex neurophysiological relationships, such as boundaries and overlaps between mental disorders, to better understand disease progression, and response to stimuli after drug admission, 2) assess the impact that different methods have in finding, classifying, and exhaustively searching the three-dimensional space for meaningful biomarkers, as well as provide statistical tests to guarantee the statistical significance of the discovered biomarkers, 3) extend the proposed machine learning approaches towards predictive tasks as to place new diagnostics, prognostics, and therapy recommendations for new patients, using their neurobiological profile against the found patterns.

Title: Persistent Memory for Data-intensive applications
Presenter: Ilia Kuzmin
Abstract:
Real-world applications have complex constraints on the hardware they run on. Many of them require intensive computations to process large amounts of data and to achieve decent performance enough resources (like CPU and RAM) should be supplied. Yet any resources go for a cost, and finding optimal configuration could be a challenge by itself. Furthermore, having millions of source code lines exist, it is practically impossible to adjust each of them to use cutting-edge technology features, thus abstraction layers (like Operations Systems) should provide an opportunity to employ full hardware capacity transparently.
My current work is focused primarily on non-uniform memory access technologies. In particular, it focused on incorporating large amounts of cheap, energy-efficient, yet slow random access memory to the data-intensive applications on the system level, to enable performance boost without changing particular application implementation.

Tuesday 13 July 9h00

Title: Cooperation Dilemmas on Imperfect Information in Hybrid Populations
Presenter: Henrique Fonseca
Abstract:
The mechanism of Indirect Reciprocity (IR) provides an elegant solution to the cooperation dilemma by arguing that reputations and social norms are core elements of human social decision making. This has been studied in the fields of ecology, psychology or economy, both mathematically or computationally. However, little has been researched regarding the dynamics forced by imperfect information, i.e. when agents have diversified opinions on the same matters. Moreover, those that tackle this problem often assume that individuals have binary non-null opinions on all agents in a population, something that tends not to scale well with population sizes. Here I tackle the problem of IR with imperfect information by changing the commonly used computational models to include a third reputation besides Good and Bad: the Unknown reputation. This leads to new unexplored dynamics in Evolutionary Game-Theory based simulations capable of tackling questions such as: How to regard strangers for cooperation to arise?; What are the roles of gossip, empathy or social conformity in providing consensus to chaos driven populations?; and What are the impacts on information dissemination of having different cognitive capabilities?

Abstract:
Title: Concurrency-induced crash-consistency bugs in Persistent Memory systems
Presenter: João Gonçalves
Title: Procedural Content Generation for Cooperative Games