Dev-XP-Session on Machine Learning

Chat session for developers, designers and technology enthusiasts on the future of Machine Learning.

Akshay Sharma

25.04.18

Hello Everyone,

We had a wonderful session yesterday and we thought it best to continue with that awesomeness.

Today, we have Akshay Sharma. We are glad to have you here with us. Today session is centered around Machine Learning but before we get into that, lets know more about you

- Yes sure, I am Akshay Sharma, from India, currently a Software Engineer at Glowing.io YC S18.

- My work revolves around ML and Software Engineering, I also work for the GOVT OF INDIA, to solve problem of graduate employment in India.

Thats sounds interesting! What is it like in YC ?

- It's great to be in YC, being in YC has a lot of benefits. The network you have at your disposal is great.

Lets get down to business πŸ˜„

Question from: Fari @fari

Can you talk about your tech stack you use most frequently in your role? Do you use python ? How much do you do in python? Do you find that you deploy production code in python, or do you have to move to other languages?

There is no particular tech choice of mine, a tech stack is decided based upon the problem at our disposal, for eg: If you are building ML application a python based tech stack makes sense.

Yes, I use python, I love python as a language, at Glowing we use ruby as our tech stack language, not python, so we deploy code using ruby.

Question from: Olaoluwa @olaoluwa

What intel can you give about being successful with YC? Must you have a page to show, or generate revenue, or what exactly can you say take the eye of YC for prospective startup to be funded and successful?

YC, looks at a lot of factors before accepting any startup into the school, the possibility of being a company with a billion dollar valuation is one, YC funds a lot of company in a single batch, being successful with YC, means using the money carefully, using the YC network to ask for help when needed.

Question from: Somto M.Ugeh @somto

What trends are you seeing in the ML/ DL space? What excites you ?

I am particularly excited about capsule networks, published by Hinton et. al in 2017, they seem to out perform our conventional CNN architectures, very excited about them, as well as I see there is a trend in recent years for optimizing the Gradient Descent algorithm to so that it converges more easily as faster.

@dakoto
Capsule nets can't handle segmentation properly.... are there any algorithms upcoming in that space?

I hope someone is already working on the segmentation problem already.

Question from: Olaoluwa @olaoluwa

Based on your Stack/Tech related answer, have you ever been at a cross-road of mastering one particular stack or the other, and confused of which stack to master? And how do you solve that problem of mastering/facing one out of the two as a developer ?

I guess, one should be open to choosing any of the tech stacks available, see which one will make the development process more streamlined on a long scale, regarding the mastery of tech stack, no one can master all the tech stacks as we work on a tech stack we keep learning and become better.

olaoluwa
That’s about the best answer I ever heard on this confusion.

Yes, no one knows everything willingness to accept new ideas drives innovation and development.
It's okay to not know everything.

Question from: Dakoto @dakoto

What Machine Learning tasks does your job require more? Tuning algorithms/hyperparameters, building datasets, cross-validation, or more?

Each task has it's importance, to my experience building data set and hyper parameter tuning require more time and use of resources, but on a larger scale, it depends on the problem you have and the data set.
It's a very iterative as well intuition based process.

Question from: Dakoto @dakoto

When it comes to AI taking over, are you more on the Elon Musk camp(AI can take over if care isn't taken)? or the Mark Zuckerberg camp ("you guys are overreacting) ?

Great question.

Not pointing any fingers.
But I agree with Elon Musk, because on a longer scale if you see the things he said totally makes sense, AI is powerful and should be used in a controlled environment, AI can be dangerous given the pace of AI research, there will be time, when AI can itself get conscious to it's presence which can be dangerous.

Question from: Zainab @zainab

Have you always been in CS, if not how did you make the change and become something of an expert in this subject ? Also how would you advice people with out CS background and would like to dive into ML/AI ? Do they need to go get a degree ?

I am a follower of the notion of DEGREES/MARKS do not matter.

I graduated with a CS degree last year, and I have many friends who are doing great in CS without attending any school, in the end of the day it comes back to your interest in the field and how determined you are to learn and improve.

You have the internet at your disposal, use it to learn more, build projects and you can get into companies like Google with no degree, but saying on the side of a practical scenario a degree has it's own benefits, but if you do not have a degree currently I guess, it's better off to not join a full time CS and rather build your network.

To get into ML/DL, your mathematics intuition should be strong, very strong, because if you are good at maths, you can easily breeze through top research papers in DL/ML and make sense to why this research will work in my problem domain, I myself worked on my math for 1 month, and I revise learn more mathematical concepts to stay relevant.

You can use this to learn more.

Question from: Dakoto @dakoto

Do you have any experience with GANs(generative adversarial networks)? if so, in what kinds of tasks/domains do they prove to be useful ?

Unfortunately, I have not worked on GANs. 😞

Question from: Seun @seun

How long until machine learning algorithms can trigger detections in a new environment? How many algorithms require a learning period, and how long does that take ?

It totally depends on how well training goes and have you found the correct set of hyper parameters for your model, is it overfitting or not, how does it do on the test set. It can take upto 6 hrs to 2 weeks depending upon the model/dataset.

But the best way to try this is to build a model by yourself.

Question from: Afeez @ajay

I read recently on application of regular sql scripts for retaining customers rather than use of AI/ML models? what is your take on this ... How do you think ML/AI affects ecommerce ?

Hi @ajay , I am not in e-commerce space to comment on the sql part, but I know companies like Amazon use ML to price match their products to keep up with the competition.

ajay
Thank you! Use-case and size of data set should maybe guide application

Question from: Dakoto @dakoto

Neural networks (but even with these, you have to constantly parameter-tune, optimize architecture, etc) come pretty close because they're good at both supervised and unsupervised learning ..... but do you think a we can expect a universal algorithm (to rule them all, lol) anytime soon?

Well, it's a very good question, I hope it happens soon, but not at the cost of AI getting dangerous for man kind.

Hi guys, we are gradually coming to the end of the session.

We are grateful to Akshay Sharma for taking out time to be part of us today! if you have an questions, you can ask in the 2 mins we have left.

Thanks everyone for a wonderful session!! until next time! Do have a wonderful day! πŸ˜ƒ

Want to be part
of the conversation ?

Join the community