Wojciech Jaśkowski
@WJaskowski
Followers
53
Following
525
Media
7
Statuses
45
Maxine Learning / Senior Research Scientist @ Snowflake. Opinions are mine own.
Joined August 2012
Dear @Google, "Full house"? What is that supposed to mean? Should I kick off one of my children on the street? Not only did I find this offensive but also exclusive! How can I create a supervised gmail account for my fifth child?
1
0
1
Training this model was a lot of fun. As usual, lots of data + lots of compute = great results. This one is special however due to its small size and 400k context. The power of encode-decoder models has been underappreciated these days.
We introduce Arctic-TILT, which achieves accuracy on par with models 1000x its size for #LLM workloads involving answering questions grounded on PDF or scan content. 🧵 .
0
0
2
RT @AgainstRevisio1: ONLY IN POLAND!. Most people have no idea about the differences between the countries occupied by the Germans. The Po….
0
3K
0
RT @kawecki_maciej: @elonmusk space travels era with 90’s electronics, really? Check out innovation from @mkratajczak. Saves space in stars….
0
238
0
A fascinating debunking of Dunning-Kruger Effecr.
Nothing is sacred to these people.
0
0
0
RT @mmagierowski: Words matter, dear @CNN. There was no such thing as „Poland’s concentration camps”. Those were German Nazi camps in Germa….
0
1K
0
simple & solid improvement.
Deep learning has seen huge gains when you increase the number of layers, but what about Deep RL? . Introducing D2RL! Changing how you parameterize your policy + Q function boosts performance. Co-led with @mangahomanga. @AravSrinivas @animesh_garg . Link:
0
0
0
0
527
0
RT @nnaisense: In a recent technical report, we have developed theoretical connections between energy functions of noisy data learned using….
0
1
0
RT @nnaisense: It's live! Visit our completely new website to learn about how we're bringing real value through cutting-edge Machine Learni….
0
8
0
We also demonstrate how accurate are the value gradients w.r.t actions predicted by MAGE's critic when compared to classical value-based TD-learning (log-scale, lower is better)
1
0
2
TD3-MAGE significantly outperforms model-based and model-free TD3-based baselines.
1
0
0
We propose MAGE, which explicitly trains the critic for value gradients w.r.t to actions using temporal difference learning. To obtain the *gradient* targets, MAGE backprops through the learned dynamics model. This leads to a critic tailored explicitly for policy improvement.
1
0
2
How to learn a useful critic in RL? We provide an answer in (@nnaisense). Typically, the critic learns to predict expected return even though the theory says it not the value but the value gradient w.r.t. to actions that is used in policy improvement
1
9
24
RT @nnaisense: We're honored to be featured by Forbes among the AI30 2020 list of AI startups making waves in the DACH region. https://t.c….
0
5
0
Nice. It is nice to see that still so many people find our #vizdoom useful in their research.
Typically, we think of intrinsic motivation as _maximizing_ surprise. But agents in complex worlds with unexpected events can learn meaningful behaviors by _minimizing_ surprise, leading to behaviors that seek out homeostasis: Can learn vizdoom w/o reward
0
0
2
RT @nnaisense: Here is a list of all of our hot-off-the-press research to be presented at NeurIPS 2019 workshops in Vancouver! https://t.co….
0
7
0
The results of our recent work @nnaisense.
Our MAX algorithm for pure, active exploration now scales to high-dimensional continuous environments to build task-agnostic models. Check it out:
0
0
2