Hacker News Reader

aub3bhaton COST in the land of databases

1Mb vs 1Gb vs 1Tb distinction is a function of information contained in data. If its PII even 1Mb is an orginizational headache from security/risk-management perspective.

Also most commodity compute nodes have allocation around 16 Gb. The big difference comes in 16~256 Gb range where having a single powerful node can make a huge difference.

aub3bhaton COST in the land of databases

The author should spend some time delivering solutions to "real problems". It's easy to be snarky while writing a blogpost with toy datasets but it's difficult to actually deliver production solution in an organization. The real COST is often not the time consumed but costs that arise from maintaining heterogeneous architecture and remaining legally compliant.

A database or hadoop cluster is not just an execution engine, but rather a managable, reproducible execution engine. Sure you can code up a smarter algorithm on your personal i9 workstation. However having something available as part of DB or Hadoop ecosystem enables reproducible execution while taking care of compliance (e.g. privacy), security, access control etc.

Eventually we will have a containerized extensions to DB query execution flow that will allow us to have best of both worlds but for 99% practicing data scientists downloading a Gigabyte sized subset to their own workstation is not a viable option.

The real opportunity lies in ability to add arbitrary compute/memory capacity to DB query execution flows.

aub3bhaton PixelNN – Example-Based Image Synthesis

Perceptual hashing is useful for copy detection. Its not robust to changes/transformations nor do the hashes encode any semantic information.

aub3bhaton PixelNN – Example-Based Image Synthesis

I am developing a Visual Search engine with pluggable indexing models.

https://www.deepvideoanalytics.com

aub3bhaton Relicensing React, Jest, Flow, and Immutable.js

PyTorch itself isnt, But the multi-GPU version requires Gloo which is BSD+PATENTS

https://github.com/facebookincubator/gloo/blob/master/PATENT...

aub3bhaton Relicensing React, Jest, Flow, and Immutable.js

https://github.com/facebookresearch/faiss/blob/master/PATENT...

aub3bhaton Our Approach to Privacy

On Android you can Sideload VPN apps, iOS on the otherhand banned them in China. Apple mounted the most successful attack on General Computing with it's walled garden. The whole Apple is good for privacy is marketing.

aub3bhaton ML Basics: K Nearest Neighbors in Ruby

Umm that's a beautiful theoretical result no doubt. But in practice the choice of the "space" that determines the neighborhood, matters more than anything else on the top of that space. Also as another commenter pointed out the result is applicable when infinite labeled samples are present, which is almost never.

IMHO the reason Machine Learning is NOT Stats or Information theory because these beautiful fundamental result were found to be futile in practice.

aub3bhaton Putin: Leader in artificial intelligence will rule world

Unless you are traveling in subway/Uber/Car-pool/commuter-rail and catching a quick nap.

What does this have to do with AI. In fact having an alarm that randomly gets disbabled is Stupid design.

aub3bhaton Putin: Leader in artificial intelligence will rule world

Artificial intelligence == ML/CV/NLP (not the AGI) will transform the world over the next two decades. People vastly underestimate importance of recent advances in detection/recognition/NLU tasks.

Sure it sounds boring by some sci-fi standards but when it comes to warfare ability to deploy cheap drones that can do near real time enemy identification will transform not just the battelfield but the entire society. Look at the current heated debate on Police stops leading to shooting. In less than a decade we will have autonomous cop car capable of disabling vehicles and firing tasers.

The world will be completely transformed just with todays state of the art tech. All it currently lacks are "platforms" for deployments.

aub3bhaton PhD Productivity Strategy: Do Deep Work in Less Hours

    or revolutionizing the field

aub3bhaton A camera that snaps a GIF and ejects a cartridge that displays it

"printing" to a "removable screen" might actually be more eco-friendly than printing a "per-frame polaroid flip-book" which what I assumed initially.

aub3bhaton PhD Productivity Strategy: Do Deep Work in Less Hours

Hedge funds, prop trading etc. do hire from math heavy background. But I agree.

My advice is only good for CS or similar (ECE, Data Science) PhD's.

aub3bhaton A camera that snaps a GIF and ejects a cartridge that displays it

GIF are a great example of a medium that succeeded specifically DUE to it's shortcoming/constraints. In fact most "gifs" on highly visited sites are mp4 videos (which can have audio and longer length). But they pretend to be gif in order to convey original set of affordances of gif format.

aub3bhaton PhD Productivity Strategy: Do Deep Work in Less Hours

This is stupid advice with sole purpose of keeping students engaged in publishing incremental research. So that no one questions vanity metrics like H-index and the ridiculous condition of academic research.

As a PhD student the real secret strategy is to ignore academia, work only on very high impact projects ( with potential of becoming startups or revolutionizing the field), intern with production teams (in core areas) at great companies. And enjoy lack of 9 to 5 scheduled life.

The deep work in less hours is meaningless if the work itself is merely publishing "Minimum Publishable Units" couple of times every year. It's just fast track to mediocrity.

Though I agree that not all environments are hospitable to risk taking. But then dropping out is a better deal.

aub3bhaton Which job skills are most employable?

>As well as looking at income, we rated skills on satisfaction, risk of automation, and breadth of applicability.

Translation: We cooked up data by adding our biases, but since we teamed up with a "data science fellow", thus our findings must scientific!

This is worse than the average feel good pop pscyhology that appears in NYtimes.

Reminds me of this good bit from an EJMR post:

"Any approach that claims to recover the distribution of individual utility parameters from aggregate data based on arbitrary distributional and functional form assumptions is dubious at best,"

aub3bhaton I’m Peter Roberts, immigration attorney who does work for YC and startups. AMA

A CEO, CTO, CFO, CXO etc. are all "employees", technically an employee is someone who works for the company (unpaid work also counts) the company.

E.g. You can have 100% equity, be sole shareholder, pay money to a manager/employee to conduct work/business. You can even derive a dividend from the said business. however you cannot "engage in business" which means that you cannot write code whose IP gets assigned to the company. You also cannot have day-to-day management responsibilities. If I remember correctly planning a business and/or attending meetings of board of directors is allowed.

At end of the day shareholding pattern trumps everything else.

aub3bhaton I’m Peter Roberts, immigration attorney who does work for YC and startups. AMA

"cofounder" is a legally meaningless SV term. There are three ways you can be associated with a company

1. Shareholder (anyone in the world, except few countries with sanctions)

2. On board of directors (varies from case to case)

3. Employee (requires immigration status)

aub3bhaton Don't Over-React to the Facebook Patents License

I think the real concern is that there is nothing in PATENTS that restricts itself to software patents, if tomorrow say a University patents a new AR/VR Chip and Facebook/Oculus infringes on that patent there the PATENTS would affect all researchers/students at that University.

If I am mistaken please show me where it states that only the agreement is restricted to software patents alone.

I am completely fine with Facebook's choice at end of the day it's their hard work and right. However it's important to understand broader implications.

aub3bhaton Google TPU Performance Analysis

Floating point operations (TFlop) != Tensor Operations (TOp)

aub3bhaton Nvidia-Docker – Build and Run Docker Containers Leveraging Nvidia GPUs

Caffe 1 is developed by Berkeley and I think Apache or BSD licensed. The Patents.txt issue occurs with Caffe 2 which is developed by Facebook.

aub3bhaton Nvidia-Docker – Build and Run Docker Containers Leveraging Nvidia GPUs

Certain algorithms/models are implemented in PyTorch or Caffe and typically it's huge amount of work to convert them to TensorFlow while ensuring correctness / Parity. Also I personally like design of PyTorch.

aub3bhaton Nvidia-Docker – Build and Run Docker Containers Leveraging Nvidia GPUs

I use NVidia-Docker extensively in my Open Source project Deep Video Analytics [1] when combined with TensorFlow (which allows explicit GPU memory allocation) its unbeatable in running multiple inference models on a single GPU in a reliable manner. Combining this setup with docker volumes on AWS EFS allows simple multi machine deployments.

[1] https://github.com/AKSHAYUBHAT/DeepVideoAnalytics

aub3bhaton Reidentification ban is not a solution

The author is clearly mistaken. There are several things that are possible in the physical world yet illegal, e.g. forging signatures, breaking doors open, breaking into parked car, sending spam emails etc. Specifying reindentification as illegal is a great step since it let's legal machinery to do its job.

The reality of data privacy is that it's impossible to guarantee anonymity while keeping data useful.

Reindentification ban enshrine coherent guidelines into the law. It's a good step forward.

I am saying this as a researcher who has signed several agreements with US government agencies which had reindentification ban clause and penalty of felony offense if found violated.

aub3bhaton Prodigy: A new tool for radically efficient machine teaching

You might be interested in Deep Video Analytics, its a Visual Data Analytics platform that I am building. [1]

[1] https://github.com/AKSHAYUBHAT/DeepVideoAnalytics

aub3bhaton Show HN: Anonymous interview evals of strong software/ML engineering candidates

The person interviewed isn't senior. The one interviewing is.

I think her solution for array larger than memory is very inefficient. Even in smaller case one might achieve speed up by excluding unnecessary read/temp/move in else clause by just by writing two values which will likely be in registers.

aub3bhaton Show HN: Anonymous interview evals of strong software/ML engineering candidates

For the interviewer 1 question of sorting an array with only two unique elements. Am I missing something? The solution just requires counting lower value and writing it number of times it occurs and then writing the other value until the end.

https://en.m.wikipedia.org/wiki/Counting_sort

aub3bhaton How Fear and Outrage Are Sold for Profit

>> You're also short-changing those liberal arts majors. After all, many of them go on to law school, with the result that Congress is about 2/3 liberal arts majors.

Congress the beacon of efficiency, progress and wisdom.

aub3bhaton Launch HN: D-ID (YC S17) – Protecting Identities from Face Recognition Tech

I don't think so, I tested and neither detector nor recognizer were fooled. Here are the results with FaceNet, Inception, MTCNN face detector and MobileNet SSD detector. Facenet also retrieved similar looking faces from LFW.

[1] https://www.dropbox.com/s/weasqs8pxn9q7qv/screencapture-loca...

[2] https://www.dropbox.com/s/i2nii4biy8aioyo/screencapture-loca...

[3] https://www.dropbox.com/s/lmq8ldqujyk3pa9/screencapture-loca...

aub3bhaton How Fear and Outrage Are Sold for Profit

Microprocessors, Internet, Green Revolution, T-Cell therapy, China & India over the last 2 decades ..... You know the stuff that actually required knowledge and power. Not some feel goodery truthiness.