goodmattg

Intriguing post! Back in my undergrad days I wrote a course thesis on Syntactic Tree Kernels applied to the problem of determining duplicate question on Quora. Like you, I am convinced that using graphs to link semantics with structure is under-explored:

https://github.com/goodmattg/quora_kaggle

https://goodmattg.github.io/articles/2017-09/Syntactic-Tree-...

Further, as someone who came out of the the graph signal processing world (Ribeiro, Gama, et. al.), and with recent experience in types and functional programming (Pierce, etc.), I too admire the beauty of using types and graphs, but as you note, hardware implementation is a fundamental constraint to production. Perhaps the IPU units from Graphcore are a move towards this? I could use clarification on your points regarding next steps in low-level languages and hardware.

I've had PyTorch Lightning recommended before, and heard it works great! My main concern is that it isn't officially supported, while Ignite is part of PyTorch.

Ignite is still in alpha and needs to catch up, but I get the feeling that PyTorch / FB will commit resources rather than folding in Lightning (unlike TF did w/ Keras). The new releases are already promising (w.r.t. distributed training).

Always looking to collaborate, and please lmk if anything above is incorrect.

a) Likewise!

b) This repo contains a sample research repository. The sample we provide is training MNIST for classification.

To use: (1) install using the installation instructions in README.md, then (2) train using "python research/train.py".

When you train: (a) a timestamped folder will show up in (/outputs) containing training logs (b) a Visdom server will auto spinup on port 8000 (localhost:8000) showing training plots

Correct, the config comes from (research/configs/default.yaml). The magic of Hydra (thanks to Omry Yadan @omry) is that you hierarchically define yaml configs. I.e. in default.yaml, when you say "backbone: vgg19", that pulls in the config "vgg19.yaml" and merges it with the default.yaml config.

Specific things we include: - "LogDirector" that lets you log to multiple modalities by setting up a recipe - example of correct Hydra usage - example of Ignite usage - auto Visdom server spinup - lots of helper functions common in computer vision research

We're still working on docs, they are not clear enough!

Hi HN, we want to present open-source boilerplate for computer vision research! The goal of this project is to provide a lightweight boilerplate repository that researchers, Ph.D. students, or generally curious individuals can clone to kick off their research.

Our main idea is that researchers waste too much time managing their projects because tooling is hard to set up. We believe that project plumbing should "just work in a sane way", so we set you up to do YAML configuration management, logging to multiple modalities (image files, Visdom, log files, stdout, etc.) with no changes.

Thanks to the UPenn Computer Vision Group and several individuals in GRASP for their generous feedback.

Now get back to research!

-Mia & Matt

Great project! Obviously a stretch, but would be curious to see a breakdown by workshop, general conference, orals, etc. The umber of publications is not the same as their quality.

Also, do you disambiguate corporate affiliations in academia? E.g. many papers will be +1's for a university and a corporate research group (Stanford + Google)

The comparison to Freetown Christinia is thought provoking, but the key difference seems to be that Capital Hill has been a trendy part of town with new apartments going up and rents increasing for the last decade, whereas Christinia was originally an abandoned site.

I don't question the resolve or peaceful intentions of the people currently occupying CHAZ, but instead the patience of property managers, local residents, and local businesses (though closed due to COVID-19).

The description of Caro's books as "Harry Potter for adults" is accurate. Highly accessible, no preparation necessary

I'm currently in Prof. Pierce's class at UPenn! Can confirm this is a great book, and we literally work through the exercises. Taking this class AFTER using Scala in industry has really put a lot into context. It's so easy to start too - download CoqIDE or ProofGeneral and get to it!

This whole "how much math is required to work in ______" always feels like reading Rashomon. Everyone is looking at the same situation, but no one sees it the same way. Data Science teams range in size and sophistication across various industries, so obviously "what it takes to be a data scientist" varies based on the company.

Want to be a data scientist in a large org. with limited data science sophistication? You may not need an advanced mathematics background to make an impact; out-of-the-box models/techniques might produce actionable results if you have access to clean data.

Try doing the same thing at sophisticated tech companies (FANG, BAT, etc.). The datasets are massive, heterogenous, and often unstructured. Add to that, large teams full of former Ph.Ds and clever engineers have already given the common problems a go, so any better solution will need to be 'more clever'. It's difficult to be 'more clever' without advanced math.

What is hilarious about these articles is none of them ever say "Math as a SENIOR data scientist", because we all acknowledge that senior scientists know the math. Almost every job listing for senior data scientist will prefer a Ph.D. On larger teams, the Ph.D (advanced math) crowd will come up with the general approach that junior data scientists will work on. So yes, getting in the door as a data scientist requires having enough math/programming so you can at least understand what a senior scientist wants you to do. But if you want to lead a team in a technical capacity, you can't possibly imagine doing that without really knowing the math or having a brilliant track record.

And if you are the rockstar 10x guru Data Scientist who consistently delivers actionable insights without a strong math background, you deserve the respect of everyone here and will probably be my boss someday.

Is there some inside joke to the blog name? I really don't get it...

Does anyone else see "work in progress" at the top of the page?

To elaborate on why the job security is so good, it's because of the clearance system. During the internship, my manager said that if I wanted to come back next summer I would have to get top-secret (TS) clearance. Almost everyone working there has at least secret (S) or TS clearance. Getting clearance takes ~1 year WITH an employer backing you, and it may take years (a.k.a impossible) without one. So once you're in the system with clearance, you can bounce around the def-sec world forever.

I had the same experience! Sophomore EE (signals) internship at Raytheon. While I had the same reaction as you at the time, my thinking has evolved a bit. I can see why a subset of people enjoy working there. The problems are difficult, the work is 'meaningful' (many of my coworkers had family in the military), and the pace is deliberate. I met 60 year-old guys writing firmware for fighter jets and they loved it. The culture is the exact opposite of "move fast and break things" - you may write 10 lines of code a week, but those are some exceedingly important lines.

This is my issue with the phrase "AI/ML" as a catch all. You have an excellent general skillset, but "AI/ML" encapsulates a wide spectrum of jobs.

ML-SWE: SWE with ML focus - building architecture around models, feature engineering, distributed training, etc. Relatively limited ML knowledge needed (IMO). The math won't be helpful for this role. Much more important to have SWE background. If you want this, keep building your programming knowledge (Python) and read books. Would focus on understanding the popular frameworks PyTorch and TensorFlow b/c your work will likely interface with those.

Research engineer: Mostly for MS/PHD background. Farther away from the product and closer to actual research. This doesn't sound like what you want to do.

Data Scientist: ML is a subset of the knowledge needed. Applied statistics as important, if not more so. Doesn't sound like you want this.

A path forward:

(1) Program a lot. On what? Anything at all, b/c you need programming skill to work as a SWE.

(2) If you want to do ML-SWE, program with an eye towards ML applications. Maybe do a simple cloud project that leverages ML - Google Cloud makes this particularly easy for classification tasks. Focus on breadth here, not depth. No sane person outside of academia can keep up with state-of-the-art and truly understand it. Far too much material, so focus on fundamentals.

(3) Work towards your strengths. You aren't some hotshot kid out of college proclaiming to be an AI guru. That would be silly and no competent recruiter would believe it. You know hardware - and AI (neural networks) leverages a lot of hardware. Why not focus on the hardware side of AI? Demonstrate your knowledge of how/why TensorFlow is so effective across distributed hardware, or how CUDA accelerates NN computation, or why TPU claims vs. Nvidia may be up to interpretation, etc. This should be a natural transition given your background.

TLDR; Know what you really want to do. Your background is valuable. Play to your strengths. Don't ring the bell.

"no end in sight" is false. It took ~5 seconds to Google this.

https://www.barrons.com/articles/netflix-will-be-cash-flow-p...

This is literally what trading firms try to do. If you want to know more, ask a trader.

Fair point. Looked back in my notes and can't find a percentile breakdown of the use of options in markets. For the derivatives markets as a whole less than 7% are even options.