The authors describe the preferred stimuli of the lower layers of InceptionV1. Simple cells & color contrast cells pool into complex cells.pic.twitter.com/xazqnHZCfg
You can add location information to your Tweets, such as your city or precise location, from the web and via third-party applications. You always have the option to delete your Tweet location history. Learn more
The authors describe the preferred stimuli of the lower layers of InceptionV1. Simple cells & color contrast cells pool into complex cells.pic.twitter.com/xazqnHZCfg
Starting in layer conv2d2 we start seeing more complicated features that look fairly V2 like - line and curve detectors, as well as various texture unitspic.twitter.com/1Vw0uRZmWh
Lots of specialized detectors for circles starting in layer 3 and above. It starts to look very V4 like.pic.twitter.com/02XevCXwfb
Starting in mixed3b, the network sure starts to like eyeballs and other complex featurespic.twitter.com/VOcbwqXVP3
There's a lot of interesting parallels between different layers of the model and what we know about the early (and not-so-early) visual cortex. It would be nice to see fine-grained measurements of alignments between ANNs and BNNs sorted by feature type (i.e. micro-BrainScores)
At the same time, there's a lot of representation of features that haven't been well-studied: second-order edges defined by spatial frequency, joint color and shape tuning, eyes, pattern selectivity in V4, etc.
I would love to see if there's an eye space in visual cortex in the same way that there's face spacepic.twitter.com/3rIhv9OL7O
The preferred stimulus representation used by the authors doesn't show nonlinear feature integration or invariant subspaces: how can we do that with this type of visualization?
I would love to see a video showing highlighting the invariant subspaces of a given neuron. And I think nonlinear feature integration (and/like) could be highlighted using model reductions (fitting a model to a model), perhaps with canonical nonlinearitieshttps://www.ncbi.nlm.nih.gov/pubmed/18254695
I was skeptical, but I think this approach is turning out pretty fruitful! Excited for the next batch of articles to come out.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.