Winter 2024 - Montreal Time Series Meetup

Attention Montreal Research Community! Borealis AI and Moov AI present the Fall 2024 instalment of our time series meetup. We're kicking off the new year with two exciting research talks, kindly hosted at ServiceNow's Montreal office.

You're invited to our IN-PERSON meetup on January 18 from 5:30 pm to 7:30 pm ET. Join us at ServiceNow's Office for networking, refreshments, and exciting research discussions.

You'll hear from two industry practitioners at the forefront of Time Series research and business applications. There will be plenty of Q&A time where you can pick our speaker's brains.

Event Agenda:
*All times are in EST

• 5:30 pm - 6:00 pm - Sign in, Grab a name tag, refreshments + Network
• 6:00 pm - 6:05 pm - Welcome / Opening Remarks
• 6:05 pm - 6:35 pm - Transformer-based Models for Time Series by Arjun Ashok, ServiceNow Research, MILA, UdeM*+ Q&A*
• NETWORKING BREAK
• 6:55 pm - 7:25 pm - Lessons Learnt from Time Series Work in Consulting by Alexei Nordell-Markovits, AI Director at Moov AI + Q&A
• 7:25 pm - Close - Networking and Wrap-up

Event Location:
ServiceNow - 6650 Rue Saint-Urbain #500, Montreal, Quebec H2S 3G9

About our Speakers:

• Arjun Ashok, ServiceNow Research, MILA, UdeM

Arjun Ashok is a Visiting Researcher at ServiceNow Research, Montreal and a PhD student at MILA-Quebec AI Institute and Université de Montréal advised by Irina Rish and Alexandre Drouin. His research interests are in time series forecasting, with a focus on designing scalable general-purpose models for time series prediction.

Talk: Towards General-Purpose Models for Time-Series Prediction.

The advent of deep learning-based methods has led to a number of models that excel at a wide range of time series prediction tasks. However, the usability of these models on real-world data is limited by several factors, such as the presence of a large number of time series, irregular and non-uniform sampling frequencies, missing data, and arbitrarily complex data distributions. I will first introduce TACTiS (Transformer-Attentional Copulas for Time Series), a general-purpose model based on transformers that was contributed by our team, and addresses all the stylized facts about real-world time series. I will then address the various challenges in scaling models such as TACTiS, towards building pre-trained foundation models for time-series, and touch up on our team’s ongoing efforts in addressing the same.