Hi there 👋

This code will download all CVPR 2023 papers to your local machine. I haven’t included any filters, but those are easy to add with this template in place. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 #!...

Setting up an Nvidia GPU enabled system is the ante for deep learning research at home. Without a dedicated GPU, users are forced to either use Google Colab, which has great performance for a free service but is a notebook environment, or a cloud ML service that costs money (~$0.9 GPU/hour on AWS). After playing with those options for awhile, I decided to build my own system. In doing so, I learned how confusing the process can be, especially if it’s your first PC build or first build with a dedicated GPU system....

Community is everything. Building a network effect through your community is becoming a key marker of success. This much is obvious and written to death in thinkpieces. Tools to actually support that community, to make it more enjoyeable for them to participate are still growing. Think about how Discord went from gaming chat, to the defacto community management platform for web3 / bleeding-edge software projects. As developers we build these things we think are great… and then they go nowhere....

This post is going to cover the state of deep learning in 2021 with particular emphasis on AI infrastructure and deep learning tooling. If you’re coming from a theory-based university classroom, get ready for an exhausting amount of detail. The classic pattern of “just feed some data through the network and backpropogate” is still the truth, but it takes ~10x more effort beyond the network to get anything useful....

The Demand: smart people who want lucrative jobs correlated with “smartness” I still don’t get crypto, but I kind of get one of the needs for crypto. If you run in my “elite” 1 post-grad circle you’ll notice that a lot of people talk about finance. Every quantitatively oriented software engineer I meet will introduce themselves and be like “I’m John. In my free time I like hiking, hanging out with friends, and I do some quantitative trading just for fun....

When we talk about type safety in the context of statically typed languages, we mean that the language builds in a typing checking mechanism into the compiler. No type-check, no compilation. But in the context of research / algorithms oriented software that has no uptime requirements, the bias coming out of academia is to get the math into code, these days Python, and put it in production. Python is dynamically typed and has no type inference… so the typing is optional....

The authors present a succinct and powerful manifesto of the importance of graph networks in artificial intelligence research. The paper is a move to synthesize the previously disjoint field of graph neural networks into a single framework: “Graph Networks (GN)”. Thoughts We have a much needed survey paper that unifies graph networks (GNs) and refocuses attention away from the technical details of network composition/functionality back to the marquee goal of building networks that achieve combinatorial generalization through flexible computation on structured representation....

Mendeley is a great tool to keep track of papers and citations across multiple devices if you can figure out how to actually enable file (i.e. PDF) syncing across devices. The UX designers were on a lunch break when this one was cooked up. The good news: enabling syncing is easy and you’ll never have to think about it again: Step 1) Open Mendeley Desktop To my knowledge, the sync files across devices setting can only be configured from Mendeley Desktop....

I was reading Mohammed AlQuraishi’s post1 “AlphaFold @ CASP13: What just happened?” recently about DeepMind’s advance in protein folding. His point about research siloing holding the whole field back resonated with me in light of my current work identifying malignant tumors in ultrasound images. To put it bluntly, it isn’t helpful that I can quickly find ten good papers claiming state-of-the-art performance on tumor segmentation if I can’t quickly download and verify the model performance on a non-toy dataset....

Prove Newton’s Theorem for $t \in \mathbb{R}$ for $|x|<1$. If $t=n$ is any positive integer obtain the standard binomial formula $(a+b)^n=\sum_{k=0}^n\binom{n}{k}a^{n-k}b^k$. $$(1+x)^t=1+\binom{t}{1}x+\binom{t}{2}x^2+\binom{t}{3}x^3\cdots=\sum_{k=0}^{\infty}\binom{t}{k}x^k$$ My calculus is rusty, so my first and only instinct when I see an infinite series like this with incrementing exponents is to use Taylor’s approximation theorem with remainder. $$f^{n+1}(x) = \underbrace{\sum_{k=0}^{n}f^{(k)}(0)\frac{x^k}{k!}}\text{T(x)}~+~\underbrace{\frac{1}{n!}\int_0^xt^nf^{n+1}(x-t)~dt}{R_n(x)}$$ Looking this over, we can tell this is probably a good approach. That initial term matches up almost perfectly to our proof objective....