anand kulkarni 🇺🇸

Tensorflow.js came out last week, which lets you run existing models and train new models directly in-browser. Expect to see a lot more impressive NN experiences cropping up on the web. medium.com/tensorflow/int…

D3 version 5.0 is out, from @mbostock (and others). Still one of my favorite visualization frameworks and among the most popular at @Crowdbotics. #changes-in-d3-50" target="_blank" rel="nofollow noopener">github.com/d3/d3/blob/mas…

Had a great time speaking at @googleorg last night about @Crowdbotics and our work hiring + deploying top (female) software engineers from @Laboratoriala.

This past year, Nabila Hassan used her earnings as a LeadGenius researcher to build a borehole well in her home village of Bomani, #Kenya that provides more than 500 families with safe drinking water. hubs.ly/H09zlGf0

Learned approaches for common datab structures outperform standard, non-ML ones. Neat and practical result. twitter.com/chrmanning/sta…

In particular, this implies that for these applications, it's the structure of the neural network that's essential – not the training data. The solution is found by searching through the parameter space of the network, starting with a random point, not image space.

This deep learning result is phenomenal (and totally surprising). The authors show convolutional neural nets can achieve amazing visual results *without* training data: dmitryulyanov.github.io/deep_image_pri…

Four questions to answer about marketplace startups: @jgolden/four-questions-every-marketplace-startup-should-be-able-to-answer-defb0590e049" target="_blank" rel="nofollow noopener">medium.com/@jgolden/four-…

Tensorflow Lite is out. It can execute trained neural nets efficiently on mobile devices: developers.googleblog.com/2017/11/announ…

FB is estimated to have grown revenue 11% per week (close to YC's 10% recommendation) between its seed and Series A. Those round sizes & valuations look proportional to revenue growth. (source: quora.com/What-were-Face…)

Founders sometimes point to Facebook's early success and conclude they should ignore revenue. This is inaccurate. FB's revenues grew exponentially from year one: a $300K run rate in its first year/seed round, a $6M run rate at Series A, and a $52M run rate at Series B.

This software can read new deep-learning papers, analyze the figures, and then automatically attempt to build a neural net following the approach of the paper. Nice idea. Still rough, but could be a useful approach to a self-improving software system. twitter.com/Deep_Hub/statu…

Important explanation of how software 2.0, instantiated as deep neural nets, is fundamentally different from previous models of software development: @karpathy/software-2-0-a64152b37c35" target="_blank" rel="nofollow noopener">medium.com/@karpathy/soft…

A nifty little Vue plugin from @fahrenq at @crowdbotics. twitter.com/MadeWithVueJS/…

Google's Tangent lets you differentiate functions written in Python, and gives you results in Python. Useful for those of us who spend a lot of time with NNs. github.com/google/tangent

In Japan, the Family Romance corporation provides (for years) purported fathers, brides, even deceased relatives. theatlantic.com/health/archive…

@fahrenq @_StartupTracker @ProductHunt I’m curious! Where do they get their source data? Is it simply aggregating Crunchbase and AngelList?

The NYT discusses efforts to teach ML systems to figure out parameters & topologies currently being tuned by hand by ML practicioners: nytimes.com/2017/11/05/tec…

Excellent article on of the state of the art in feature visualization. distill.pub/2017/feature-v…