TDT99 - Modern AI for Time Series Analysis (FALL 2020)

The course will focus on modern machine learning for the analysis of univariate and multivariate time series (i.e., anomaly detection, forecasting, classification, data imputation) with some focus on "irregular" time series. In particular:

• Use of FNN, LSTM and CNN for time series modelling and forecasting.
• Attention mechanism in LSTM-based architecture for time series forecasting.
• The problem of small data and low-data regime in the time series domain.
• Unsupervised and Self-Supervised Learning for different time series related tasks.
• Transfer Learning and Transformer architecture.
• Few-Shot Learning and TS Classification in low-data regime.
• GAN for time series analysis (i.e. Anomaly Detection, Data Imputation, Data Augmentation, Data Generation, Privacy).
• ((Deep) Echo State Network and Spiking Network for Time Series Analysis.)

Topics

Topic 1: Forecasting Irregular Time Series (4)

• Joint Modeling of Local and Global Temporal Dynamics for Multivariate Time Series Forecasting with Missing Values [Vemund F.]
• Deep and Confident Prediction for Time Series at Uber [Assigned to Christian L.]
• Shape and Time Distortion Loss for Training Deep Time Series Forecasting Models [Eivind S.]
• Self-boosted Time-series Forecasting with Multi-task and Multi-view Learning [Svein Ole M. S.]
• Time-series Extreme Event Forecasting with Neural Networks at Uber [NOT TO PREPARE FOR THE ORAL EXAM]

Topic 2: Attention Mechanism for Time Series Analysis (4)

• Temporal pattern attention for multivariate time series forecasting [Assigned to Henrik G.]
• Attend and Diagnose: Clinical Time Series Analysis Using Attention Models [Assigned to Margrethe G.]
• Multivariate time series forecasting via attention-based encoder–decoder framework [Assigned to Henrik F..]
• Modeling Extreme Events in Time Series Prediction [Assigned to Jens W.]
• DATA-GRU: Dual-Attention Time-Aware Gated Recurrent Unit for Irregular Multivariate Time Series [NOT TO PREPARE FOR THE ORAL EXAM]

Topic 3: GAN for Time Series (4)

• E2GAN: End-to-End Generative Adversarial Network for Multivariate Time Series Imputation [Assigned to Sigurd V.]
• Time-series Generative Adversarial Networks [Assigned to Claus M.]
• Generative Adversarial Networks for Failure Prediction [Assigned to Kristoffer G.]
• MAD-GAN: Multivariate Anomaly Detection for Time Series Data with Generative Adversarial Networks [Assigned to Henrik H.]
• A GAN-Based Anomaly Detection Approach for Imbalanced Industrial Time Series [NOT TO PREPARE FOR THE ORAL EXAM]

Topic 4: Graph NN for Time Series Analysis (4)

• Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting [Assigned to William K.]
• Connecting the Dots: Multivariate Time Series Forecasting with Graph Neural Networks [Assigned to Axel H.]
• GMAN: A Graph Multi-Attention Network for Traffic Prediction [Assigned to Lars B.]
• Dynamic Spatial-Temporal Graph Convolutional Neural Networks for Traffic Forecasting [Assigned to Dario S.]

Topic 5: Unsupervised / Self-Supervised Learning for TS Analysis (3)

• Self-Supervised Learning for Semi-Supervised Time Series Classification [Assigned to Christer B.R.]
• Adversarial Unsupervised Representation Learning for Activity Time-Series [Assigned to Helle G.]
• A Deep Neural Network for Unsupervised Anomaly Detection and Diagnosis in Multivariate Time Series Data [Assigned to Haakon J.J.]
• (Learning Representations for Time Series Clustering) [NOT TO PREPARE FOR THE ORAL EXAM]
• (Deep Multivariate Time Series Embedding Clustering via Attentive-Gated Autoencoder) [NOT TO PREPARE FOR THE ORAL EXAM]

Topic 6: Few-Shot Learning and Transfer Learning for Data Scarcity in TS Analysis (4)

• Reconstruction and Regression Loss for Time-Series Transfer Learning [Assigned to Ask S.]
• TapNet: Multivariate Time Series Classification with Attentional Prototypical Network [Assigned to Herman R.]
• Towards a universal neural network encoder for time series [Assigned to Elen E.K.]
• Meta-Learning for Few-Shot Time Series Classification [Assigned to Erik O.D.]

Teaching approach

The course is structure in form of set of workshops where each student will present one paper followed by group work and discussions. All the student MUST attend the workshops.

Instructions for Students

The student should prepare a 12-15 minutes presentation focusing on the following aspects:

• Motivation (why this problem is important?)
• Description of the addressed problem (and challenges)
• Related work (how is addressed the same or similar problem)
• Methods
• Exaperimental Setting and Results
• Conclusions

The presentation will be followed by a short discussion.

Calendar

When: Fridays, kl 13, Where: Zoom Meeting. Students will receive an invitation

Exam (Oral)

IMPORTANT: In order to access to the exam the student has to present a paper in one of the workshop. When: 27.11.2020 and 30.11.2020 (see calendar here , Where: Zoom Meeting. Students will receive an invitation

List of Students

• Lars B.
• Elen E. K.
• Irene F.
• Vemund F.
• Svein Ole M. S.
• Christian L.
• Helle M. G.
• Haakon J. J.
• Christer B. R.
• Margrethe G.
• Claus M.
• Sigurd V.
• Kristoffer G.
• Henrik H.
• Eivind S.
• Herman R.
• William K.
• Axel H.
• Jens W.
• Henrik G.
• Henrik F.
• Ask S.
• Erik O.D.
• Dario S.