Date: October 1 at 4 pm
Speaker: Trung Bui – Adobe Research US
Title:
Toward Multimodal Dialog for Conversational Image Editing (View Recording)
Abstract:
Conversational Image Editing is a novel application domain that combines language and vision as well as action execution in the task-oriented scenario. The dialog system application goes beyond traditional slot-filling systems for restaurant search or movie booking.
In this talk, I will first discuss the importance of dialogue incrementality and the models for incremental intent identification based on deep learning and traditional classification algorithms. I will show how our model based on convolutional neural networks outperforms models based on random forests, long short term memory networks, and conditional random fields. By training embeddings based on image-related dialogue corpora, we outperform pre-trained out-of-the-box embeddings, for intention identification tasks. Our experiments also provide evidence that incremental intent processing may be more efficient for the user and could save time in
accomplishing tasks.
Second, I will present a simple multimodal dialogue system prototype for Conversational Image Editing. We formulate our multimodal dialogue system as a Partially Observed Markov Decision Process (POMDP) and trained it with Deep Q-Network (DQN) and a user simulator. Our evaluation shows that the DQN policy outperforms a rule-based baseline policy, achieving a 90% success rate under high error rates in this simple simulated setting.
Date: October 8 at 11 am
Speaker: Andres Suarez-Cetrulo – CeADAR
Title:
Continuous machine learning in concept drifting contexts (View Recording)
Abstract:
In many Artificial Intelligence applications, data is produced incrementally over time and can be observed in the form of a data stream. The data accumulation process creates a set of challenges at pre-processing and model training stages that are often considered unavoidable without process parallelisation. To overcome the performance bottlenecks and avoid retraining models unnecessarily, this talk introduces different predictive AI approaches to deal with new data as it comes: modelling in continuous setups, incremental methods, adaptive learning with forgetting mechanisms, and online learning to predict on-the-fly. I will cover issues to focus on when working with data streams: the detection of context changes, when to train and when to adapt to concept drifts. Finally, different algorithms, artificial datasets and libraries commonly used for learning from data streams will be described.
Date: October 22 at 11 am
Speaker: Lilian Genaro Motti Ader – CeADAR
Title:
Challenges and perspectives in gait assessments using wearable devices (View Recording)
Abstract:
Wearable devices are being explored for health and medical applications, providing data on users’ mobility and activity. My recent study evaluates an approach to enable gait assessment from a small number of steps, identifying digital biomarkers that can provide reliable measures for monitoring symptoms and disease progression in persons with multiple sclerosis (MS). I will discuss the challenges of capturing and analysing data, often fragmented, from non-controlled environments (e.g. community, home settings) and how digital health initiatives could support patients and healthcare systems.
Date: October 29 at 11 am
Speaker: Santos Fernandez Noguerol – TU Dublin
Title:
Applying machine learning to predict appeals after Revaluation at the Valuation Office of Ireland (View Recording)
Abstract:
The Valuation Office is Ireland State property valuation body under the aegis of Department of Housing, Planning and Local Government. Its core business is to provide ratepayers and local authorities with accurate, up-to-date valuations of commercial and industrial properties.
One of the key tasks of the Valuation Office is the National Revaluation Programme, a process by which all rateable properties in a local authority area are valued periodically and, at the same time, by reference to a single valuation date.
For 18 months, a project to predict whether a commercial property will go to tribunal after revaluation applying machine learning algorithms was developed. The presentation will explain the key aspects of the project and its conclusions.
