Very nice results from a @GoogleAI research effort on a general purpose time-series prediction model that gives good zero-shot performance to new forecasting tasks.
@jeffdean
-
Google Maps Largest Human Brain Connectome Achievement
By
–
It's been amazing watching the progress of this @GoogleResearch collaboration w/HHMI, @MCB_Harvard, … over the last decade. Starting w/fly brains years ago, progressing through parts of mouse brain, and now the largest-ever connectome for part of a human brain! 🧠 https://t.co/RSWfpgdH3z
— Jeff Dean (@JeffDean) 10 mai 2024It's been amazing watching the progress of this @GoogleResearch collaboration w/HHMI, @MCB_Harvard
, … over the last decade. Starting w/fly brains years ago, progressing through parts of mouse brain, and now the largest-ever connectome for part of a human brain! -
DistBelief Training System: Clarifying the Cat Detector Paper
By
–
There was a bit of confusing back and forth between myself and @ylecun recently, because in the tweet below, 'it' was referencing DistBelief, but @ylecun thought DistBelief referred to the ICML 2012 "cat detector" paper (
https://
icml.cc/2012/papers/73
.pdf
…), not to the training system called -
Med-Gemini Achieves State-of-the-Art Medical Imaging Results
By
–
This new paper on Med-Gemini research from @GoogleResearch and @GoogleDeepMind shows how the multimodal and long context capabilities of Gemini 1.5 can be used to get state-of-the-art results on a variety of 2D and 3D medical imaging and genomic risk score tasks.
-
DistBelief Legacy: Model and Data Parallelism in Modern Neural Networks
By
–
I'm still a bit surprised that you think the system (DistBelief) that first combined large scale model- and data- parallelism with backprop-based training of very large neural nets is a "dead end", given that nearly every large-scale model that is trained today combines these
-
DistBelief convolutions development shared weights 2012
By
–
We supported shared weights from very early in DistBelief development, and we started adding support for various kinds of convolutions in March of 2012 (because convolutions are a very sensible idea for some things!), and you visited in July, I believe.
-
DistBelief Convolutions Backpropagation Training Before ICML 2012
By
–
DistBelief supported convolutions and end-to-end training with convolutions via backprop before the ICML 2012 paper appeared (even though that particular paper didn't use convolutions).
-
DistBelief: The Distributed Neural Network System Behind Modern AI
By
–
The name actually was "DistBelief" for a few reasons. The first was a play on "disbelief", because many people didn't seem to believe the approach of large-scale distributed neural networks would work. The second was because Deep Belief Networks was one of many algorithms we
-
DistBelief Distributed Training System Framework Comparison
By
–
DistBelief is the distributed training system used for that paper & 1000s of other things. Your statement is equivalent to implementing an unsupervised algorithm in PyTorch, seeing modest results (like 70% relative improvement in SoTA on Imagenet 20k) & declaring PyTorch a dead
-
Neural Network Training Scaled 30X Beyond Previous Literature Records
By
–
This system successfully scaled up training of a neural net 30X larger than previously reported in the literature. You're somehow not very knowledgeable & very opinionated about it, referring to it as a "dead end". Even your "Apologies for the confusion" tweet is incorrect.