EfficientNet: Improving Accuracy and Efficiency through AutoML and Model Scaling
Google did that thing they do again where they make vast steps in artificial intelligence and machine learning. Efficiency of these image recognition networks is up to 10* greater. Most of this gain is because they try to use “scaling coefficients” so that the network scales in a predictable way. I’m just mad because it’s a TensorFlow model and not PyTorch so I can’t drop it into any of my existing image recognition notebooks.
A promising step forward for predicting lung cancer
Another Google blog post about how they are doing incredible things. Man what I wouldn’t do to work for Google Brain. (This research is also being done at the same university I am doing my capstone with, so hey maybe they can sneak me in) Okay so in this article they describe a state of the art result for predicting lung cancer using improved volumetric predictions of CT scans. They instead of looking at individual slices in the image are instead reconstructing 3-d structures to improve the accuracy. This both is and is not a crazy step forward. Being able to use the 3-d structure seems to be truly revolutionary, but some of the radiologists performed equally as well as it. Seems that it will be a useful assistance tool for now.
Moving Camera, Moving People: A Deep Learning Approach to Depth Prediction
I promise this won’t be all Google, but again what they are doing right here is incredibly cool and a little bit scary. They have found a way to approximate the 3-d size, shape, and depth of moving people even when the camera is moving. This work has really cool implications for AR and VR and a little bit terrifying uses for a potential police state. There are many places where face recognition has been banned or people are considering banning it, however, combining a 3-d map of a person with existing effective identification techniques like gait tracking can serve as a proxy for facial recognition in those areas. Combined with facial recognition it could provide an even stronger match limiting false positives, and avoiding false negatives.
Speech2Face: Learning the Face Behind a Voice
Okay we are finally away from Google, but into something even more terrifying. This neural network when fed a small sample of speech is able to generate a qualitatively accurate facial guess. The model seems quite adept at identifying both race and gender. Scary stuff.
This fun parody site created by Joshua Davis, Kyle Gibson, and the pseudonymous Cas Piancey mercilessly lampoons the tomfoolery of Kik’s attempt to challenge the SEC. For the record, I do not think promising an Ethereum public DApp and delivering a one node Stellar fork is a good thing.
If you enjoyed this please subscribe to my newsletter at https://bennettftomlin.substack.com/