... where Deep Learning happens in Toulouse
You are not logged in.
Pages: 1
Hi dear Tiddlers!
I really hope you are all good and safe during this pandemics.. and you all have a good connection to a bunch of powerful GPUs to pass this long quarantine until the next deepBeer!
I'd like to know what you guys think about a web-talk Bengio recently gave at the Neuromatch conference* (live from his place office..). It was very interesting, and about his (new?) line of research in Artificial Consciousness.
https://www.crowdcast.io/e/neuromatch/22
I'd be very curious to know what do you think about it overall?
For example, do you think he is too vague in his system1 - system2 starting point (from 'thinking fast and slow' - Daniel Kahneman book), or do you think it's a reasonable approach from a ML point of view?
What do you think of the inductive biases for system2 he talks about?
I just dropped some questions on top of my head, but I'd be very curious what you guys think about it overall, whether you have any different insights or ideas about it at all.
Looking forward to hear your opinion on that!
Cheers,
Andrea
*you may need to register (for free) to access the conference. Otherwise he gave a very similar talk at NeurIPS19, just longer : https://www.youtube.com/watch?v=T3sxeTgT4qc&t=296s
Please find here the sildes of the presentation:
https://docs.google.com/presentation/d/ … sp=sharing
Dear TIdDLers,
thanks for your interest in and positive responses to the NeurIPS meetup we are organizing in collaboration with TDS. Registrations are officially closed (mostly to make sure the catering --generously sponsored by ANITI-- will be enough for everyone), and we are glad to finally share the final program (see below).
If you wanted to join but didn't register yet, you may still email andrea.alamia@cnrs.fr to make a request, we'll try our best to accommodate you. Similarly, if you registered but finally won't be able to come, please let us know so we can offer your spot to somebody else.
We are looking forward to seeing you on Monday.
All the best,
Andrea (for the Animation Committee)
NeurIPS meetup Toulouse program:
[Monday, Dec 16th, the Village by Ca!, 31 Allees Jules Guesdes, Toulouse]
9:30 - 10:00 : Welcome and Introduction + Coffee and Breakfast
10:00 - 10:15 : Projection 1: "Uniform convergence may be unable to explain generalization in deep learning"
10:15 - 10:45 : "Non-Stationary Markov Decision Processes, a Worst-Case Approach using Model-Based Reinforcement Learning"
by Emmanuel Rachelson and Erwan Lecarpentier (ISAE/SUPAERO)
10:45 - 11:00 : Projection 2: "Brain-Like Object Recognition with High-Performing Shallow Recurrent ANNs"
11:00 - 11:30 : "Bio-inspired few-shot learning on the edge with Akida" by Sebastien Crouzet and Douglas McLelland (BrainChip)
11:30 - 12:00 : "Neural Networks for Poetry and Song Generation" by Tim Van de Cruys (IRIT/ANITI)
12:00 - 13:30 : Buffet lunch
13:30 - 13:45 : Projection 3: "Causal Confusion in Imitation Learning"
13:45 - 14:15 : "Training convolutional spiking neural networks with backpropagation through time & application to speech processing"
by Thomas Pellegrini and Timothée Masquelier (IRIT, CerCo)
14:15 - 14:30 : Projection 4: "Geometry-Aware Neural Rendering"
14:30 - 15:00 : "3D trajectories analysis using spiking cameras" by Guillaume Debat Robin Baurès, Benoit Cottereau, Timothée Masquelier, Tushar Chauhan, Clément Joly, Olivier Brousse, Michel Paindavoine (CerCo)
15:00 - 15:30 : --- Coffee break ---
15:30 - 15:45 : Projection 5: "A neurally plausible model learns successor representations in partially observable environments"
15:45 - 16:15 : "Asymptotically exact data augmentation: models, algorithms and theory"
by Maxime Vono, Nicolas Dobigeon and Pierre Chainais (IRIT/ENSEEIHT)
16:15 - 17:00: Round table / General discussion
17:00 - 18:00: Cocktail
It's time to kick off our first seminar, which will take place the 2nd of December from 14:30 to 15:30 at the IRIT Auditorium in the Paul Sabatier Campus. We will have two speakers, and each presentation will last 20 minutes plus an additional 10 minutes for questions. The presentations will be in English and everybody is welcome to join!
The speakers are Bhavin Choksi from CerCo who will present his current work on biologically inspired neural networks, and Vincent Roger from IRIT who will provide an overview about few-shot learning techniques.
Here the two abstracts:
- Bhavin Choksi (CerCo): "Building biologically inspired robust neural networks"
"Adversarial perturbations in computer vision can easily fool neural networks to misclassify images from a category as belonging to a completely different category. We hypothesize that this is due to the poor representations learned by the networks while training on one-hot targets that do not convey any information about semantic relations between categories. We also use predictive mechanisms thought to be used in the brain to build more robust architectures. In this talk, I will discuss our approach and the progress we have made thus far."
- Vincent Roger (IRIT): "Overview of few-shots learning techniques"
"Actual state-of-the-art models use Big Datasets to achieve their results. These models are mainly Deep Neural Networks and are not well adapted for small datasets. By small datasets we mean more or less 5 examples per class. In this talk we will review few-shot techniques/frameworks that try to tackle this problem."
The idea of the Deep Beer Meeting is to have a monthly informal event which gathers Tiddle members in front of a (few) beer(s). The goal is to get to know each other without necessarily talking about Deep things. Everyone is welcome to join, so feel free to spread the invitation to friends and colleagues!
The first event will take place the 18th of November at 18h00 at LePery (https://www.lepery.com/).
We will try to have a monthly meeting, changing location and day of the week. Any suggestions or recommendation is welcome!
Looking forward to see you soon!
Andrea (on behalf of the TIdDLe Organizing Committee)
Pages: 1