The OpenAI Microscope is a collection of visualizations of every significant layer and neuron of eight important vision models like VGG19, AlexNet, ResNet etc.

https://microscope.openai.com/models

#ml

Darket YOLOv4 is faster and more accurate than real-time neural networks Google TensorFlow EfficientDet and FaceBook Pytorch/Detectron RetinaNet/MaskRCNN on Microsoft COCO dataset.

https://medium.com/@alexeyab84/yolov4-the-most-accurate-real-time-neural-network-on-ms-coco-dataset-73adfd3602fe

#ml

A uniform interface to run deep learning models from multiple frameworks from Uber. Now only supports TensorFlow, PyTorch, TorchScript, Keras and Ludwig.

Generalization in ML continues, I guess we will see more soon

https://github.com/uber/neuropod

#ml

😢

https://youtu.be/ePhAj8ktiEA

Maybe somebody doesn't know but there is a good collection of SOTA papers with accuracy scores splitted by areas. Check this out: https://paperswithcode.com/sota

#ds

In Amazon, books are slowly appearing in official co-authorship with GPT-2
https://www.amazon.com/s?rh=p_27%3AOpenAI+GPT+2

#ml

How do you know if the code is well-written? Ask junior developer what this code does. If he doesn't understand, the code isn't written well.

#dev

Eventual consistency in comics

#big_data

IBM Research created Fully Homomorphic Encryption (FHE), which keeps sensitive data encrypted at all times, securing the cloud for highly regulated industries. They announced a new FHE toolkit for Linux, available today, bringing FHE to multiple Linux distributions for IBM Z and x86 architectures.

This is a missing link of end-to-end encryption. And will be very(very) much in demand of developers on distributed platforms. But unfortunately it takes 10-20 times more memory and 40-50 times more CPU.

https://fhe-website.mybluemix.net/

#privacy #big_data

Git is a super complex piece of software, there's so much inside. Check out this article on how to get git to remember previously resolved merge conflicts with git rerere.

https://levelup.gitconnected.com/the-git-rerere-command-automate-solutions-to-fix-merge-conflicts-d501a9ab9007.

#dev

This guy created a testing tool for distributed systems called Jepsen. So he tested Riak, MongoDB, Redis, Cassandra, NuoDB, Kafka, Zookeeper, etcd, Consul, Elasticsearch, RabbitMQ, Aerospike, etc. So, Zookeeper is passed his tests, and that's it, just Zookeeper.

https://youtu.be/tRc0O9VgzB0

#big_data

Old post about non-practicality of CAP-theorem by Martin Kleppmann

https://martin.kleppmann.com/2015/05/11/please-stop-calling-databases-cp-or-ap.html

#dev

M-Motivation, M-money, M-Mastery, M-Mystery

https://luminousmen.com/post/m-motivation

In depth explanation of a new features in Apache Spark 3.0

- The new Adaptive Query Execution (AQE) framework within Spark 3.0 can yield query performance gains. Based on a 3TB TPC-DS benchmark, two queries had more than a 1.5x speedup, and another 37 queries had more than 1.1x speedup.
- With Dynamic Partition Pruning (DPP), we can significantly speed up performance by pruning partitions based on the joins between the fact and dimension tables common in star schema design.
- Accelerator-aware Scheduling helps Spark take advantage of GPU and hardware accelerators for certain workloads (e.g deep learning). This release enhances the scheduler and makes the cluster manager accelerator-aware.
- Spark 3.0 also introduces new Pandas UDF types and new Pandas function APIs for improved performance and usability.
- Enhanced monitoring capabilities including the new UI for Structured Streaming, enhanced EXPLAIN command, and observable metrics.


https://youtu.be/g-qZslQsOuE

#spark #big_data

A large database of datasets for computer vision projects. Can be sorted by popularity, addition and publication times. Please use whose interested: https://www.visualdata.io/discovery.

#cv #ds

Procedural content creation pipeline importing OpenStreetMap data into Unity using Houdini to generate real time environments based on real cities. Everything in the video placed or generated procedurally.

https://twitter.com/stinastinzen/status/1287692890253271041

#ml

Data & AI Landscape 2019, Source: https://mattturck.com/data2019/

Hard to see anything, right?

#big_data

Thoughts on management

The manager is like a system administrator. The better he works, the more free time he has (which can be used for larger tasks rather than firefighting and micromanaging). If you get fired because of it, then this is a stupid company, and there is no point in catching something there. Or you just outgrew the company, and it's time to look for new tasks on the side.

If you set up the processes and organize the work so that it doesn't take 100% of your time, you can take more employees, run more projects. Lead a department or a team.

You probably know that you can do a lot more with other people's hands. More people are bigger tasks and a bigger contribution to the company's success.

#management

The AI bot defeated a human pilot(5:0) in a series of virtual air battles that unfolded in the sky, albeit in an air simulator, during a competition conducted by the American military research unit DARPA. This is a detailed 5-hour video of AI actions. All the fun starts at 4:40.

Well, now we've got a machine that kills people better than any human...

https://youtu.be/NzdhIA2S35w

#ml

Lex Fridman explains the number of GPT-3 parameters and explains how much will it cost to train a language model the size of the human brain.

https://youtu.be/kpiY_LemaTc

#ml

Define "production-ready"

I came across this interesting question on one of the stackoverflow-like sites. I've never thought about it in terms of definition. Interesting read.

https://softwareengineering.stackexchange.com/questions/61726/define-production-ready.

#dev