Krzysztof Maziarz
@maziarz.bsky.social
Principal Researcher at Microsoft Research AI for Science, ML for Drug Discovery, ex Google AI & tensorflight
In blind comparisons by chemists, the two modules were judged on a similar level, but Chimera was clearly preferred 🧑🔬
December 11, 2024 at 9:30 PM
In blind comparisons by chemists, the two modules were judged on a similar level, but Chimera was clearly preferred 🧑🔬
To our knowledge, we're the first to show large gains from ensembling in this domain.
The trick is to combine full ranked output lists using a strategy based on learning-to-rank.
Gives a significant improvement out-of-the-box for any pair of models 🔥
The trick is to combine full ranked output lists using a strategy based on learning-to-rank.
Gives a significant improvement out-of-the-box for any pair of models 🔥
December 11, 2024 at 9:30 PM
To our knowledge, we're the first to show large gains from ensembling in this domain.
The trick is to combine full ranked output lists using a strategy based on learning-to-rank.
Gives a significant improvement out-of-the-box for any pair of models 🔥
The trick is to combine full ranked output lists using a strategy based on learning-to-rank.
Gives a significant improvement out-of-the-box for any pair of models 🔥
We just preprinted Chimera: our new retrosynthesis prediction model ⚗️
aka.ms/chimera
It has two modules: a large-scale Transformer and a graph-rewriting GNN. Chimera learns to combine predictions from both to leverage their unique strengths & biases.
Thread below 🧵
aka.ms/chimera
It has two modules: a large-scale Transformer and a graph-rewriting GNN. Chimera learns to combine predictions from both to leverage their unique strengths & biases.
Thread below 🧵
December 11, 2024 at 9:30 PM
We just preprinted Chimera: our new retrosynthesis prediction model ⚗️
aka.ms/chimera
It has two modules: a large-scale Transformer and a graph-rewriting GNN. Chimera learns to combine predictions from both to leverage their unique strengths & biases.
Thread below 🧵
aka.ms/chimera
It has two modules: a large-scale Transformer and a graph-rewriting GNN. Chimera learns to combine predictions from both to leverage their unique strengths & biases.
Thread below 🧵
I'm still new here, so I'll just port my badge posting tradition over from X.
If you're at #NeurIPS2024 and want to chat about AI for Science, come to the MSR booth during one of our slots:
- Tuesday 16:30 - 17:00
- Wednesday 16:30 - 17:30
- Thursday 12:00 - 13:00
If you're at #NeurIPS2024 and want to chat about AI for Science, come to the MSR booth during one of our slots:
- Tuesday 16:30 - 17:00
- Wednesday 16:30 - 17:30
- Thursday 12:00 - 13:00
December 10, 2024 at 10:28 PM
I'm still new here, so I'll just port my badge posting tradition over from X.
If you're at #NeurIPS2024 and want to chat about AI for Science, come to the MSR booth during one of our slots:
- Tuesday 16:30 - 17:00
- Wednesday 16:30 - 17:30
- Thursday 12:00 - 13:00
If you're at #NeurIPS2024 and want to chat about AI for Science, come to the MSR booth during one of our slots:
- Tuesday 16:30 - 17:00
- Wednesday 16:30 - 17:30
- Thursday 12:00 - 13:00