Organized by the See.4C European Project
Saturday 22nd, April 2017 (right before ICLR’2017)
Vacanciel Carqueiranne – French Riviera, SALLE GIENS
[PHOTO Gallery of the Event] - (Credits: Sébastien Treguer)
Investigating and improving machine learning methods for spatio-temporal forecasting and related application areas.
This workshop is part of See.4C competition program, whose aim is to prepare an upcoming challenge of predicting power flows in the French electricity grid. We will practice making submissions on the challenge platform using data from another task: video forecasting.
Instructions to participate:
2) Meet your coach:
Himalaya: Coach Stephane ( email@example.com )
Appalaches: Coach Julio ( firstname.lastname@example.org )
Alpes: Coach Lisheng ( email@example.com )
Oural: Coach Diviyan ( firstname.lastname@example.org )
3) Develop your code on GPUs: Install the [key] in your key chain. Change permissions (chmod 400 hackathon.pem). Then connect to the AWS instance with the instructions from your coach to connect to AWS.
5) Cheat sheet
- Download the [STARTING KIT]. Download the [PUBLIC_DATA] (optional).
- Process video num 0 (change num 0 for 1, 2, etc. to test other videos):
python predictSpatioTemporal.py 0 sample_data/ results/ `pwd`
- Test all videos and compute RMSE:
python utilities/score.py sample_data/ results/ `pwd`
- Change example classifier:
Replace predictSpatioTemporal.py by one of the files in predictSpatioTemporal/.
* Zip your submission: zip all the contents of your directory.
- Use more data to train/test:
Change sample_data/ to public_data/.
6) Test prednet
Download a sample submission using prednet [HERE].
Form the starting kit, this corresponds to replacing predictSpatioTemporal_prednet.py instead of predictSpatioTemporal.py.
- No training: Change in predictSpatioTemporal.py to DEBUG_MODE=1.
- Pre-trained net: Copy the contents of pretrained_models in the cache directory. Use DEBUG_MODE=1.
- Fast training: Change parameters nb_epoch': 1, 'samples_per_epoch':5
9:00 am: Breakfast and registration. Group forming.
9:30 am - 12:30 pm: MORNING SESSION - Tutorials and invited talks
9:30 am (10 min): Isabelle Guyon (UPSud Paris-Saclay, France) and Cecile Capponi (U. Aix-Marseille). Welcome.
9:40 am (10 min): Florin Popescu (Fraunhofer Institute, Germany). Project presentation.
9:50 am (20 min): Sergio Escalera, Xavier Baro, and Julio Jacques (U. Barcelona, Spain). Sneak peek at the data and deep learning benchmarks.
10:10 am: Break (30 min). Coffee and poster viewing.
10:40 am (40 min): INVITED. Mehdi Mirza (Google DeepMind): Applications of CNNs to spatio-temporal data. [SLIDES]
11:20 am (40 min): INVITED. Bill Lotter (Harvard Univ. USA), Deep Predictive Coding Networks for Video Prediction and Unsupervised Learning [SLIDES][HP param overview]
12:00 - 12:30 pm (30 min): Lisheng Sun and Diviyan Kalainathan. Presentation of the "starting kit" and beat-persistence benchmarks.
12:30 pm: Lunch break (free lunch for participants having completed the prerequisites). Poster viewing. People are free to start hacking.
12:30 am - 17:30 pm: AFTERNOON - HACKATHON
18:00 am: Break (1 h). APERO
18:40 pm - 21:00 pm: EVENING SESSION - More invited talks!
18:40 am (40 min): INVITED. Danny Silver (Acadia University, Canada): Forecasting using machine learning techniques in energy and agriculture.
19:20 pm (40 min): INVITED. Benjamin Donnot (RTE: Réseau de transport d’électricité, France). Forecasting for Electrical Transmission Grid.
20:00 pm: Break (30 min). Snacks and poster viewing.
20:50 pm: Announcement of the winners. Stephane Ayache and Cecile Capponi (U. Aix-Marseille).
21:00 pm: Adjourn
The participants are welcome to bring a poster, which will be viewed during breaks and lunch time, on one of the following topics of interest.
- Novel methods related to spatio-temporal forecasting: image parametrization and analysis such as dimensionality reduction, time series prediction methods over large state spaces, feature selection for forecasting, causal feature selection in time series
- Novel methods to deal with wide data bandwidths in forecasting
- Novel methods for adaptive, quick turnaround deep learning in forecasting
- Benchmarks for forecasting video sequences
- Benchmarks for forecasting of time series
- Incremental learning and flexible resource-constrained learning
- Automatic machine learning for forecasting
- Spatio-temporal datasets for forecasting in agriculture, epidemiology, geosciences, economics, animate motion
- Statistical comparison of quality in forecasting systems
- Stochastic forecasting evaluation
- Combinations of loss types for increased forecasting reliability
- Robust tail-estimation and distribution representation of stochastic forecasts
- Computational languages, platforms and toolboxes for forecasting
- Adaptation of computer vision toolboxes for video understanding and forecasting
- Big open data pipelines for forecasting
- Energy sector, grid, renewables
- Video-based applications, compression, predictive control systems using video streams
- Stochastic predictions in geosciences and meteorology using ML
- Multi-disciplinary and multi-modal convergence for event and trajectory forecasting
- For speakers: They are invited at the workshop from friday night 21st to sunday morning 23rd. Travel reimbursement will be available based on need, hackathon results, timeliness and quality of contribution, within constraints.
- For pre-selected hackathon participants: The attendance to the workshop is free during the scheduled time, the mi-day lunch is offered if pre-requisite "homework" is completed. Homework infomation will be send to people registered to the meetup.
How to get there
Address : Club Vacanciel, 600 Avenue de la Valérane, 83320 Carqueiranne
GPS position : Lat. : 43.0896700 – Long. : 6.093463897
From the train station of Toulon : direct bus line n°39 (every 30mn, 45mn duration, bus stop station = Valérane)
Co-organizers and coaches
Julio Jacques Jr.