Profile cover photo
Profile photo
Lukasz Stafiniak
368 followers
368 followers
About
Lukasz's posts

Post has attachment

Post has attachment

Post has shared content
Transfer/multi-modality learning: "One Model To Learn Them All", Kaiser et al 2017:

"Deep learning yields great results across many fields, from speech recognition, image classification, to translation. But for each problem, getting a deep model to work well involves research into the architecture and a long period of tuning. We present a single model that yields good results on a number of problems spanning multiple domains. In particular, this single model is trained concurrently on ImageNet, multiple translation tasks, image captioning (COCO dataset), a speech recognition corpus, and an English parsing task. Our model architecture incorporates building blocks from multiple domains. It contains convolutional layers, an attention mechanism, and sparsely-gated layers. Each of these computational blocks is crucial for a subset of the tasks we train on. Interestingly, even if a block is not crucial for a task, we observe that adding it never hurts performance and in most cases improves it on all tasks. We also show that tasks with less data benefit largely from joint training with other tasks, while performance on large tasks degrades only slightly if at all."

Impressive. Not SOTA, but steps toward a generic NN which can do zero-shot, one-shot, or just plain learn faster by drawing on informative priors from other domains. One reason that deep NNs are data-hungry is that they can't borrow information from other domains, but that's never been fundamental - after all, humans have good finite-sample performance in part thanks to drawing on prior knowledge from all other tasks. And computer NNs don't 'age'.

Post has attachment
Vicarious picked up Gary Drescher's work (published 1991), they build a generative causal model "Schema Network, an object-oriented generative physics simulator capable of disentangling multiple causes of events and reasoning backward through causes to achieve goals."

https://www.vicarious.com/img/icml2017-schemas.pdf

Post has attachment

Post has attachment

Post has attachment

Post has attachment

Post has attachment
http://www.space.com/36892-darpa-xs-1-space-plane-contract-decision.html

"In Phase 1 of the XS-1 program, DARPA awarded prime contracts to three companies, each of which will work with a commercial launch provider: Boeing (working with Blue Origin), Masten Space Systems (working with XCOR Aerospace) and Northrop Grumman (working with Virgin Galactic)."

Wait while more posts are being loaded