MPAI holds presentation of all its activities and results
By the 31st of March 2023 MPAI will have been in operation for 30 months. A short time for a standard developing organisation, but a long one if we see what MPAI has achieved in 30 months: a consolidated standard development process; 12 official documents approved; 3 standards adopted without modification by IEEE; more work in the pipeline to extend existing documents and develop new ones; new standards and projects on health, autonomous vehicles, video coding, online gaming and XR venues. A gigantic effort that has involved tens of experts in 1,000 + hours of online meetings.
MPAI thinks it is time to expose the result of this incredible amount of work to the industry. Realising this task, however, is challenging: there are 16 topics to present for a reasonable amount of time – say, 20 minutes – and we must cover some 19 time zones.
This is the format adopted: 16 presentations will be made in two different events, one starting at 06:00 UTC and lasting 5h20m and another, starting at 17 UTC with the same duration.
We welcome those who will stay for the entire duration of 5h20m, but we will not be surprised if somebody may not be interested in the full range of MPAI activities. For this reason, there are two schedules: the first is directed to Australia/Asia/Europe and the second to Europe/America. For logistic reasons, the order of the presentations and the presenters will be slightly different. All presentations will last for 20 minutes, and we guarantee that the schedule will be strictly enforced. You can safely tune-in at the scheduled time and be sure to hear the presentation you are interested in.
To register, please use the following URL: https://bit.ly/3Z0K9nm. Date is 31st of March 2023.
These are the time-wise friendly Asia-Europe presentation sessions:
|L. Chiariglione||CH||A standards body for AI||06:00|
|M. Bosi||US||Enhancing audio with AI||06:20|
|S. Dukes||US||Connecting with standards organisations||06:40|
|E. Lantz||US||AI-powered XR Venues||07:00|
|D. Schultens||US||Connected Autonomous Vehicles||07:20|
|S. Casale-Brunet||CH||MPAI Metaverse Model||07:40|
|J. Yoon||KR||Avatar interoperability||08:00|
|M. Choi||KR||Humans and computers converse||08:20|
|M. Mitrea||FR||Watermarking Neural Networks||08:40|
|G. Perboli||IT||Predicting company performance||09:00|
|A. Basso||IT||Multi-sourced AI apps||09:20|
|A. De Almeida||PT||Federated AI for Health||09:40|
|P. Ribeca||UK||MPAI Ecosystem Governance||10:00|
|C. Jia||CN||End-to-End Video Coding||10:20|
|R. Iacoviello||IT||AI-Enhanced Video Coding||10:40|
|M. Mazzaglia||IT||Better and fairer online games with AI||11:00|
These are the time-wise friendly Europe-America presentation sessions:
|L. Chiariglione||CH||A standards body for AI||17:00|
|C. Jia||CN||End-to-End Video Coding||17:20|
|P. Ribeca||UK||MPAI Ecosystem Governance||17:40|
|A. Basso||IT||Multi-sourced AI apps||18:00|
|S. Casale-Brunet||CH||MPAI Metaverse Model||18:20|
|A. Bottino||IT||Avatar interoperability||18:40|
|M. Bosi||US||Enhancing audio with AI||19:00|
|M. Seligman||US||Human and computers converse||19:20|
|G. Perboli||IT||Predicting company performance||19:40|
|M. Mitrea||FR||Watermarking Neural Networks||20:00|
|M. Breternitz||PT||Federated AI for Health||20:20|
|R. Iacoviello||IT||AI-Enhanced Video Coding||20:40|
|M. Mazzaglia||IT||Better and fairer online games with AI||21:00|
|D. Schultens||US||Connected Autonomous Vehicles||21:20|
|E. Lantz||US||AI-powered XR Venues||21:40|
|S. Dukes||US||Connecting with standards organisations||22:00|
Neural Network Watermarking
Three targets have been identified for the neural network watermarking standard.
- Imperceptibility evaluation. If you add a watermark to an item, the result is different than the original one. As in the case of media, the question is: how much is it different and how much does the watermark impact the functionality of the item? In the case of a neural network, the watermark can be added not only to a trained model, but can also be added while the network is being trained. The standard specifies a testing process that enables a measure of watermark imperceptibility.
- Robustness evaluation. What happens to the performance of a neural network when it has been modified? The standard specifies a process that measures the performance of the robustness of the watermark against a set of modifications for the detector (“is there a watermark?”) and the decoder (“what is the payload?”). The modifications include: Gaussian noise addition, L1 pruning, Random pruning, Quantisation, Fine tuning/Transfer learning, Knowledge distillation, and Watermark ovewriting.
- Computational cost evaluation. The standard specifies the process to evaluate the computational cost of 1) watermark injection, in terms of memory footprint, time to process an epoch, and of 2) detecting or decoding, in terms of both memory footprint and time taken by the detector or decoder to produce the expected result.
Meetings in the coming January meeting cycle
|Group name||24-feb||27 feb – 3 Mar||6-10 Mar||13-17 Mar||20-24 Mar||Time(UTC)|
|AI-based End-to-End Video Coding||15||14|
|AI-Enhanced Video Coding||8||22||14|
|Artificial Intelligence for Health Data||10||14|
|Avatar Representation and Animation||2||9||16||13:30|
|Connected Autonomous Vehicles||22||13|
|Context-based Audio enhancement||28||7||14||21||17|
|Governance of MPAI Ecosystem||28||14||16|
|Industry and Standards||3||17||16|
|MPAI Metaverse Model||24||3||10||17||24||15|
|Neural Network Watermaking||28||7||14||21||15|
|Server-based Predictive Multiplayer Gaming||2||9||16||14:30|
|General Assembly (MPAI-30)||22||15|