Getting Started with Tensorflow

By Thomas Weng on February 10, 2017
~1 min. read

I took some time out this past weekend to work through an introductory machine learning talk on building models using Tensorflow. Within a few hours, I coded up a model classifying handwritten digits from the MNIST dataset with 99.5% accuracy! Considering that the accuracy of cutting edge research models currently hover at around 99.7%, my model was a pretty good result for just a few hours of development.

The resource I relied on to ramp up was Google’s crash course through Tensorflow, presented by Martin Gorner at Devoxx. The talk starts with a simple one-layer neural network, building quickly into multi-layer neural nets, convolutional neural nets, recurrent neural nets, and a collection of optimization techniques (e.g. learning rate decay, dropout, batch normalization).

I took the time to write out the code and get things working on my machine. That took longer than simply watching the video, but it deepened my understanding and guaranteed that I didn’t gloss over any important details. The slides from the talk included most of the code required for the models, but I did need to fill in some of the gaps. When I got stuck, I referenced the complete model at the Github repo Gorner created for this talk.

Gorner’s presentation glosses over the math behind concepts presented in order to focus on model building. I think this was a smart tradeoff. Having spent quite a bit of time taking courses and learning the math powering machine learning, the details of how the math works would probably be too involved for a short talk. And when it comes to actually building models, Tensorflow does the heavy lifting and abstracts the mathematical details away anyway.

Of course, the math does become more relevant as one continues working with ML. I’m sure I wouldn’t have picked things up so quickly with Tensorflow had I not already studied the background concepts.

I’m looking forward to playing more with Tensorflow and learning about what’s going on under the hood. I’ll be keeping up with the math through online resources – courses, textbooks, and hopefully papers. Most of all, I’m excited to take the training wheels off soon and start building my own models from scratch!

Robotics Day at Microsoft

By Thomas Weng on February 1, 2017
~3 min. read

When it comes to applying for opportunities, I’m no different from everyone else: I always get anxious. But preparing an application is often a great way to make your work presentable and gauge your progress.

Earlier this month, I heard about Robotics Day, an all-day event at Microsoft consisting of a series of talks, guided tours of Microsoft Research labs, and an expo showcasing robots built by hobbyist employees. The organizers sent out a call for volunteers and an application for prospective expo presenters.

This event was right up my alley. I was excited at the prospect of connecting with other MS folks doing robotics. I had even been building my own robot these past few months, so the expo was the perfect opportunity to show it off! However, I didn’t feel confident about applying for the expo, because I feared my work wouldn’t meet the same bar as other presenters. I didn’t want to stick out and look like I didn’t know what I was doing.

I wavered for a bit about applying, but ultimately I figured that I would give it my best shot and let the organizers decide if I made the cut. I also signed up to be a volunteer, so that in case of rejection, I could at least help out and get to know the organizers.

In the end, I got to do it all: I volunteered, attended the talks, and presented at the expo! I’ll focus on my experience at the expo here, so that this post doesn’t get too long.

At the expo

My presentation naturally focused on Rover, the robot that I have been building and programming for almost half a year. Here is a picture of the bot:

rover

I am building Rover to be an intelligent, mobile agent, kind of like a pet. A big part of the fun is seeing how much intelligence I can build using off-the-shelf hardware and open-source software. I’ll be writing up a more detailed post on building and designing Rover soon.

Rover is still a work in progress, so I put together several materials to aid me with the presentation. Here’s a picture of my booth setup:

IMG_8918

Here’s an online version of the poster on the wall; click for a full page view:

Presentation materials

  • Poster. On the poster, I mapped out the project goals, system design, roadmap, and key challenges, and other topics. It served as a handy reference when people asked questions, as I could point out the relevant details quickly.
  • Video. I played a short video on loop with clips demoing Rover at different stages of development. I also included demos of robots on the market aiming to be intelligent agents for comparison.
  • Contact info. I laid out some paper strips with project links and my contact information for people to take, in case they wanted to connect with me later.

How presenting went

The expo lasted for two hours, and lots of people stopped by. I ended up having virtually no downtime, which was great! It was rewarding to see people checking out my robot, looking at my presentation materials, and asking me questions about what I had built.

It was interesting to observe what drew people’s attention. My robot was powered on but stationary, and many people were drawn to the booth by the spinning LIDAR on the robot. It turned out that few people knew that fairly inexpensive LIDARs were available for hobby robotics. Once people were engaged, I pitched my vision for Rover as an intelligent pet-like robot, and described the progress I was making towards that goal.

There were a lot of other cool robots on display, like battle robots, mobile robots made with bicycle wheels, and even a foosball-playing robot! All the presenters were really friendly, and we all had a great time demoing our creations.

What I learned

My biggest takeaway from this experience is realizing how important it is to practice pitching my ideas to other people. Though I had been working on my robot for months, I had very little to show for it besides the physical robot.

Preparing to show my project to other people forced me to distill the nebulous thoughts I had been working with into well-formed concepts. Showing my work also allowed me to gauge how people responded to Rover and get their feedback. And practicing my pitch helped me keep my communication skills sharp.

Finally, I’m learning to become less anxious about applying to things. I want to treat applications more as learning opportunities rather than tests. If I feel overly anxious, not only is it counter productive, but it probably means that I’m overestimating how much I’ll lose if I get rejected. It’s better to just keep trying, learning, and growing.


Thanks for reading, and best wishes to you on whatever projects or opportunities you are working towards! Stay tuned for more details about Rover.

Tags: robotics

You Should Use a Password Manager

By Thomas Weng on January 23, 2017
~2 min. read

It’s easy to get lazy and use the same password for multiple account logins. But doing so leaves you much less secure than having different passwords.

Still, it’s near impossible nowadays to remember an ever-expanding number of passwords for different accounts. Last year, I found that I was spending a silly amount of time coming up with passwords, remembering them, typing them in incorrectly, and resetting them.

I consequently started hunting for a solution, and ended up trying out a password manager. It solved my password issues so well, I haven’t looked back!

There isn’t much to a password manager, and using one has made logging into accounts so much easier and safer. Here’s the gist of it in less than two minutes.

How do password managers work?

A password manager works like a safe. Your passwords are securely stored in an encrypted database, and you access them by entering a master password to decrypt the database.

It’s much easier to remember one long, secure password than numerous different ones. And if the safe (encrypted database) is stolen, thieves wont be able read your passwords without the key (master password).

When you need to log in to an account, you simply unlock your database, copy your credentials, and paste them into the login fields.1 No more straining to type complicated passwords.

When you need a new password, password managers create new ones for you that are hard to crack. Many of them also remind you to change your passwords if it’s been a while.

Which password manager should I use?

I use KeePass, a free program. KeePass is open-source, meaning that a community of security experts (and anyone else who is interested) is vetting the source code to verify that it really is secure.

KeePass, however, is not the most user-friendly service out there. I’ve played with LastPass as an alternative. LastPass keeps your passwords in a “cloud” so you can access them anywhere.2 LastPass runs on an ad-based model, making it also “free” to use.

There are also other password management programs and services out there, though I haven’t tried any others personally.

What if I screw up with a password manager and lose all of my passwords?

If you’re worried about losing your passwords and/or aren’t sure if a password manager is for you, you always have the option of easing into it. Start by putting one password into the manager for something that isn’t critically important, so that you get the hang of using it. As you get more comfortable, you can add more and more passwords.

If anything goes wrong, you can still reset your passwords to online accounts the regular way.


I hope you take a few minutes out of your day to give a password manager a try. In retrospect, I can’t believe it took me so long to take the jump–using a manager has been so much easier than the old system of remembering passwords that weren’t as safe anyway.

Edited January 28, 2017

  • Added pictures.
  • Edited “How do password managers work?” section.

Footnotes

  1. Some services even offer “Auto-Type”, which fills in the login fields for you with one click or shortcut combo! 

  2. I get around this problem with KeePass by keeping my password database in a cloud service myself (e.g. Dropbox, OneDrive, etc.).