I have been working and thinking about data and ML models for many years and this is probably the most concise and lucid summary of a post NN philosophy I have heard so far. It also resonates with me. Thanks for sharing.
Unlike some of those top comments, I actually like your curation a lot. Prefer it to The Economist, New York Book Review, and the other sites ppl linked in the comments. Good job!
I know you have an RSS feed, and the crowd here probably is all for RSS, but I would love it if I could leave my email somewhere and get a notification when a new submit was posted.
I was thinking along those lines in the first half of my twenties. Now in the second half I feel lucky that I found a topic that I am going deep on and I am using that to go wide in some other connected areas.
In hindsight there was some time where I had to actively force myself to stick with one topic longer than usual to go deep. Once that hurdle is taken I have less of a problem now integrating adjacent topics.
I also agree with the advice that focusing on one topic improves professional success. Long-term is TBD though.
Quick question: What's the reason for making "Transform" part of the Feature Store definition. I've been evaluating a couple of feature stores (incl. Tecton and Feast - great job by the way willempienaar) and I'm wondering if that doesn't complicate things. Especially if you already have your own data processing pipes.
"Operation Transformation" = "a system that supports collaboration functionalities by separating the high-level transformation (or integration) control from the low-level transformation functions"
Source: OT's Wikipedia article
But I felt the same. Never heard of "Operation Transformation" before and both OT and its alias were equally opaque to me.
Hearing friends talking about flaking, I would say it is a very common theme. I would also think, that you may not want to meet up with that person, because either not in the same city, or because you may match with a lot of different people which also may change over time.
I also thought of it more like a self-validation thing, a source of more information, recommendations, etc. What you would get back may be an online conversation or a feed of relevant websites.
I haven't tried Mixnode yet, but the way I understand it, it lets you query websites and retrieve their HTML content that you can then parse - without you having to crawl the site. Looking at their Github, they seem to utilize WARC, so they may also allow you to request the website for certain timestamps?
That being said, I find this highly interesting, if it works like that. We are working on a peer-to-peer database that lets you query a semantic database, popularized mostly by public web data, but with strong guarantees of accurate and timely data, and this could be a great way to write more robust linked-data converters.
I am wondering if the technique described in this Nature publication from yesterday [1] to potentially operate quantum computers at room temperature, could be used for the atomic storage as well. Does anyone know?
Yes, we had trouble to attract the right community and convert the stars into recurrent contributions, as there was less overlap between the Rust and the general ML community than we anticipated. This bothered us a lot and played a role in our decision to suspend Leaf.
We think it is more beneficial for collaboration if we stick to the common naming of layers, functions and concepts than using metaphors. We provided two links, which help to get you started with that.
But with Leaf it becomes very easy to create modules (Rust crates) that expose layers/networks/concepts, which can have a metaphorical name.
I think CNTK and Tensorflow and Theano, too, have a declarative approach, representing the computation via a computational graph. Which in my opinion is beneficial for research. But for a hacker or a software developer who wants to build an application, this creates an unnecessarily steep learning curve and feels unintuitive. (I have the feeling, that this an important reason why Keras, Lasagne and co. exists)
Leaf takes an imperative approach and explores an easier API (only Layers (Functions)[1] and Solvers (Optimizer Algorithms)), reusability through modularity and abstractions that keep the implementation and concepts to a minimum or rather abstractions that feel as familiar to a hacker as possible.
For future versions e.g., we want to explore what is practically possible with auto-differentiation via dual numbers and differentiable programming.
Well, it has a bias towards NN and deep learning for now, as this was our initial focus for the proof of concept, but the architecture of Layers and Solvers should allow it to express any machine learning concept/algorithm. We are actually working on it (verifying it) with James from rusty-machine[1][2].
We would love to have them and compare their performance with recurrent layers of other frameworks[1]. There exists an issue for the implementation of recurrent layers in Leaf (#73)[2].
We would love to test that and release the performance on Deep Learning Benchmarks[1] or convnet-benchmarks[2]. But so far no convolution/NN-related OpenCl kernels are linked[3], although it might be quite easy, if the kernels exist.