One Paragraph Reviews, Vol. I

Going to try and see if this format helps me get through the backlog of reviews I’ve been meaning to write. The schema I’ll try is: (1) why it’s interesting (2) the most interesting insight.

• 3D Gaussian Splatting for Real-Time Radiance Field Rendering by Kerbl, Kopanas, Leimkühler, and Drettakis (2023). The authors reconstruct 3D scenes from 2D images and render much faster than before (≥ 100fps) by representing them as clouds of blurry balls. No neural networks–backpropagation is used to position the blurry balls (really “anisotropic Gaussians”; anisotropic just means they are rotated/skewed); errors are propagated all the way back from image-space pixels to world-space Guassians! Another interesting non-neural network use of backprop is Constrain by Prof. Andrew Myers at Cornell, which uses backprop for constraint-based 2D graphics.
• John Calvin’s Anxiety by William J. Bouwsma (1984). Calvinism is mostly known to others¹ for the doctrine of predestination; essentially the idea that God alone chooses who to save: not even the saved get a choice! This sounds fatalistic, so it’s interesting that Bouwsma, who was a history professor at Berkeley, observes that: “Anxiety is a motif that beats through almost everything Calvin wrote.” Bouwsma thinks that Calvin’s anxiety about “fragile [physical] world” is connected to Calvin’s conception of a “constantly active” God so powerful as to appear “arbitrary.”
• Compiler and Runtime Support for Continuation Marks by Flatt and Dybvig (2020). Continuation marks are similar to dynamically-scoped variables, like UNIX shell environment variables. In languages like Scheme with first-class continuations, they can be used to efficiently implement features like exceptions, which in e.g. Java you could never implement outside of the compiler. The paper is mostly about the efficient implementation of continuation marks in Chez Scheme, but it has a good overview of continuations and continuation marks.
• The Remains of the Day by Kazuo Ishiguro (1989). An English butler reminisces as he leaves on a road trip from Darlington Hall, the English aristocratic mansion in which he’s worked his whole life. You come to find that he is not being entirely honest with himself. But Ishiguro skill as a writer allows emotion to pours through the pages in spite of this. The best book I’ve read in 2024 so far: a true masterpiece² of show-don’t-tell.
• Introduction to H.264 Advanced Video Coding by Chen, Kao, and Lin (2006). A nice introduction to H.264, the preeminent video codec today: it’s probably built into the hardware of the computer you’re using to read this! Frames are broken up into macroblocks; each macroblock is encoded as a prediction plus a residual. There are three prediction modes: P(redicted)-type macroblocks are based on a macroblock from a previous frame, B(idirectional)-type macroblocks are based on a weighted average of macroblocks from multiple frames, and I(ntracoded)-type macroblocks are based on neighboring macroblocks in the same frame. P and B-type macroblocks have motion vectors that describe the offset from which to take their reference frame. For example, if a ball is rolling around, it could ideally be encoded as one I-type macroblock then a series of P-type macroblocks with motion vectors and small residuals. The residual is compressed using the discrete cosine transform, dropping higher-frequency signals (e.g. fine textures), similar to JPEG. It was interesting to learn that many more bits are used for luminance/Y (256) than chroma (64 for blue/U, 64 for red/V).

… and that’s it for now!