Talks on David's Blog

Global Modelling for Structural Health Monitoring

Thu, 26 Mar 2026 10:30:00 -0400

In this seminar, I presented updates on using a global TAGI-LSTM to model reversible effects in structural health monitoring (SHM). I demonstrated how the global model improves predictive performance over the local TAGI-LSTM, using either zero-shot filtering or fine-tuning on a target time series.

I also shared recent results from integrating the global model into the SKF framework while highlighting the current challenges we are addressing to unlock the model’s full potential for anomaly detection.

Introduction to Continual Learning

Tue, 03 Feb 2026 18:00:00 -0500

This talk is a practical introduction to continual learning, covering the core problem, common strategies, and why probabilistic, uncertainty-aware models can be better suited for real-world evolving data.

The event was hosted by the Montreal AI ML Meetup Group at Grande Bibliotheque - BAnQ, room 1.110.

Source: Meetup event

LLMs for Time Series Forecasting

Fri, 30 Jan 2026 14:00:00 -0500

This seminar explored foundation models for time series forecasting, focusing on the use of large language models as global, cross-domain forecasters. I discussed the shift from local, dataset-specific models to large pre-trained approaches, along with key challenges such as time-series tokenization, scaling behavior, and probabilistic zero-shot forecasting.

The presentation included a hands-on tutorial using Chronos and a comparison between LLM-based forecasters and global deep learning time series models, highlighting where these methods succeed and where they still fall short.

Global Modelling of SHM Data

Thu, 23 Oct 2025 10:30:00 -0400

This seminar followed up on the progress of the Global TAGI-LSTM. Specifically, I compared the forecasting capability of the local model with the global model with and without embeddings, on a dataset of 127 time series made up of detrended real SHM data.

I also discussed the next steps of using Global TAGI-LSTM with SKF for anomaly detection.

Source: CIV-ML’s Seminars

Structured Embeddings with Global TAGI-LSTM

Thu, 31 Jul 2025 10:30:00 -0400

This seminar introduced an enhanced embedding approach for Global TAGI-LSTM. Previously, embeddings were learned individually per time series, improving performance but limiting interpretability and usability.

The proposed method employs structured embeddings, learned per category or feature within each time series. Advantages of this approach were demonstrated through toy examples, along with a discussion on future applications in Structural Health Monitoring (SHM).

Source: CIV-ML’s Seminars

Regime Switch Detection with Global LSTM-SSM

Thu, 08 May 2025 10:30:00 -0400

In this seminar, I presented a global LSTM-SSM approach for detecting regime switches in data streams using the Switching Kalman Filter.

While the existing method relies on locally trained LSTM models, I showed how to leverage additional time series data to train a single global LSTM, enabling more scalable and generalizable anomaly detection.

Source: CIV-ML’s Seminars

Online and Continual Learning for Structural Health Monitoring

Thu, 20 Mar 2025 10:30:00 -0400

In this thesis proposal, I explored a framework for online continual learning in structural health monitoring (SHM). By leveraging global modelling, the approach captures richer representations of time-series data.

A key component is the Recurrent Rauch-Tung-Striebel (R-RTS) smoother, which enables online parameter updates and real-time learning. I also presented preliminary advances in continual learning through parameter intervention.

Source: CIV-ML’s Seminars

Advancing Online Continual Learning for Time Series Modelling

Thu, 27 Feb 2025 10:30:00 -0500

In this seminar, I explored online continual learning using the recurrent RTS smoother with TAGI-LSTM. Experiments on synthetic data were presented as a proof of concept for methods that enable BRNN-based architectures to learn online and adapt continuously to dynamic environments.

The approach uses interventions on model parameters through the addition of noise. Future directions included optimizing the smoothing window length and calibrating noise injection into the parameters.

Source: CIV-ML’s Seminars

Continual Learning with TAGI

Thu, 16 Jan 2025 10:30:00 -0500

In this seminar, I explored TAGI’s capability to facilitate continual learning in neural networks. One widely studied approach in the literature involves applying regularization to model updates to prevent catastrophic forgetting.

I discussed how TAGI inherently incorporates such regularization within its parameter distribution updates, and examined whether adding noise to model parameters can improve the stability-plasticity trade-off in Bayesian neural networks.

Source: CIV-ML’s Seminars

Overview of Continual Learning in Neural Networks

Thu, 21 Nov 2024 10:30:00 -0500

This seminar offered an overview of existing approaches to continual learning, with an emphasis on how they address the stability-plasticity dilemma.

It also explored how continual learning can be approached using TAGI, building on insights from the literature.

Source: CIV-ML’s Seminars

Learning Time Series Embeddings with TAGI

Thu, 24 Oct 2024 10:30:00 -0400

In this seminar, I explored the use of embeddings for time series analysis in a global framework. I examined how TAGI enables inference of information from the analyzed time series and its capacity to distinguish between different time series.

I also discussed how embeddings can enhance the performance of the Global TAGI-LSTM model and how this contribution can be linked to the online learning capabilities of TAGI-LSTM with the recurrent fixed-lag smoother.

Application of the Modified Fixed-Lag Smoother in TAGI-LSTM

Thu, 12 Sep 2024 10:30:00 -0400

In this seminar, I presented the application of the modified fixed-lag smoother in TAGI-LSTM, enabling the model to learn without traditional epochs.

This approach allows TAGI-LSTM to transition from offline to online learning, making it more adaptable to real-time data. I also discussed challenges such as optimizing the fixed-lag window size and addressing potential overfitting.

Source: CIV-ML’s Seminars

Global Forecasting with TAGI-LSTM

Wed, 31 Jul 2024 10:00:00 -0400

This seminar explored the application of global modelling with TAGI-LSTM, how it compares to state-of-the-art models, and its potential advantages.

The talk contrasted local forecasting, where a distinct predictive model is trained for each time series, with global forecasting, where a single predictive model is trained using multiple time series.

Source: CIV-ML’s Seminars

Modified Fixed-Lag Smoother

Wed, 17 Jul 2024 10:00:00 -0400

This seminar presented a modified version of the fixed-lag smoother, traditionally used to update states within a specific interval as new observations are received.

The proposed modification operates within a fixed-lag window and incorporates the RTS smoothing equation, enabling online smoothing in nonlinear settings such as neural networks.

Source: CIV-ML’s Seminars

Tutorial on Using the Fixed-Lag Smoother

Wed, 29 May 2024 10:00:00 -0400

This seminar offered an introduction to the theory and practice of the fixed-lag smoother, a tool for online smoothing in time series analysis and signal processing.

The fixed-lag smoother operates on a defined interval of previous observations, providing a balance between incorporating historical data and remaining responsive to new information.

Source: CIV-ML’s Seminars

Characterization of the Construction Period of Buildings in Beirut Using Multi-Modal Machine Learning

Mon, 19 Feb 2024 15:00:00 -0500

This seminar presented an artificial-intelligence framework for automated characterization of building construction periods in Beirut, supporting city-scale seismic risk and vulnerability assessment.

The work compared street-view image modelling, tabular modelling, and a late-fusion multi-modal model. The multi-modal model achieved the best performance and was used in a toy example to predict building construction periods and assess earthquake vulnerability in a Beirut neighborhood.

Source: CIV-ML’s Seminars