Thread by @TivadarDanka

An exciting result came out from @GoogleAI recently, which raises several questions about how deep network architectures should be.

Here is their announcement, including a very interesting post. I would like to unpack this a bit.

https://twitter.com/GoogleAI/status/1389680797985161218

Suppose that you have a trained network and a set of samples 𝑋. You take this data and run it through the network, storing all intermediate results.

The output of the 𝑖-th layer is denoted by 𝑋ᵢ. These encode the intermediate internal representations of the data.

In general, the further you go, the higher level these representations become.

For a convolutional network, filters in earlier layers detect edges, while later activations represent objects.

Check the fantastic article below for more details!

distill.pub/2017/feature-v…

Each layer should learn something new, isn't it?

According to the latest results by Google AI (referenced in the first tweet), this is not the case!

There are quantitative measurements that can tell if representations 𝑋ᵢ and 𝑋ⱼ are similar. In this particular paper, the Centered Kernel Alignment was used. (CKA paper: arxiv.org/pdf/1905.00414…)

It turns out that the representations 𝑋ᵢ for many consecutive layers are often similar. Visualizing the CKA similarities on a heatmap, a block structure emerges.

This means that there are large sections that are doing nothing.

(Image source: arxiv.org/pdf/2010.15327…)

This observation raises several interesting questions, like

• Is the emergence of block structure dependent on the dataset?
• Can redundant layers be removed?

These and many more are answered in the aforementioned paper by Thao Nguyen, Maithra Raghu, and Simon Kornblith.

IMO this is an extremely important research area. Since the size of networks can be absolutely crazy (see GPT-3), reducing them can make a significant impact in the long run, both on the applicability of the technology and on its environmental footprint.