Welcome

Latest Posts

• Kronecker Structure in Composed Convolutions

  This post is a technical note I’ve been sitting on since around 2019. At the time I was exploring what happens when you look at stacked convolutional layers as a single linear operation, and I noticed that the resulting virtual matrix has a very specific algebraic structure: it is a Kronecker product. Not because you choose to decompose it that way, but because it just is one, by construction.

• Prediction Markets Could Create the Missing Incentive for Climate Action

  We know climate change is real. We know it’s happening now. And yet, we do nothing about it. We treat it as something we prepare for rather than something we actively avoid and nobody thinks they can make a difference. But what if there was a way to change that, and even profit from climate success?

• Why Are Kronecker Products So Effective?

  A few days ago, the list of ICLR 2021’s Outstading Paper Awards was announced. One paper immediately caught my attention from the list: Beyond Fully-Connected Layers with Quaternions: Parameterization of Hypercomplex Multiplications with \(1/n\) Parameters.