Forem: Charlie Harrington

An Afternoon with Arduino

Charlie Harrington — Fri, 29 Mar 2019 18:25:20 +0000

For the last eight years or so, I've been carrying around an Arduino (not literally on my person, but, you know, amongst my treasures), waiting for just the right time to start tinkering with it.

That time, it turns out, was yesterday afternoon. This post outlines some of things I've learned so far.

Ard-what-now?

An Arduino is a microcontroller board. The board contains a CPU (central processing unit) along with some I/O (input / output) connections. You can think about it as a small circuit - a circuit that happens to contain a programmable computer on a chip. A chip that you can learn to easily program.

Here's a picture of my Arduino (of the Uno varietal):

I love the MADE IN ITALY mark in the upper left.

The microprocessor (the computer chip) is the long flat black rectangle near the bottom-right corner of the device. It has 28 pins (14 on each side). Along the top edge of the Arduino you can see a strip of black input pins - these are the Arduino's "digital" pins (meaning that they can either be "on" or "off"). Below the CPU on the bottom right you can see another 5 "analog" pins (meaning that they can receive analog / continuous signals), and then a few "power" pins to the left to provide voltage, ground, and some other stuff that I don't know about yet.

Arduino vs Raspberry Pi

At this point, you might be asking yourself, "How is an Arduino different from a Raspberry Pi?"

It's a good question, since both are affordable, adorable, tiny little computers that you can buy for about $30 bucks or less. But the Raspberry Pi is a full-on Linux computer. An Arduino is... not. Instead, the Arduino computer holds just one program at a time. It stores this program in durable memory, so that you can turn the Arduino on and off, and it will still "remember" its latest program. It's more like a single-purpose device -- except you can dream up and build that single-purpose as many times as you want.

Open-source roots

The company behind Arduino is a non-profit and the Arduino itself is open-source - which means that anyone can build an Arduino board themselves. The original idea behind Arduino was to make a simple device that designers and artists could use for rapid prototyping of physical computing projects that use sensors (aka inputs like a keyboard or mouse or motion detectors) and actuators (aka outputs like a display or printer or lights) to interact and communicate with us human beans. Cool.

Arduino is closely tied to the Processing community. In fact, that's a bit of an understatement, since you actually write Processing code when writing programs for the Arduino -- and, just like in Processing, these programs are also called Sketches. I was happy to see this, since one of my earliest computing classes was Jer Thorp's Introduction to Processing course (which I unabashedly recommend, by the way).

As I mentioned, anyone can download the open-source schematics for Arduino and build a board themeselves with basic components. But if you'd like to make your tinkering lifestyle easier, then I suggest picking up a pre-assembled Arduino from a retailer like Makershed or Adafruit. The kit I bought (eight years ago) is the MAKE: Getting Started With Arduino Kit -- the current version (v3) appears to be retailing for $79.99 bucks. IMHO, it's definitely worth it -- the kit comes along with a bunch of goodies that help you get started right away, like a breadboard, colorful wires, clickable switches, LEDs, sensors, and a friendly introductory book. I also picked up Arduino: A Quick-Start Guide by Maik Schmidt, and I've been enjoying this book as well.

Fun with LEDs

I believe you're legally required to write a program that blinks an LED on and off as your first project with Arduino.

If you asked me a few days ago about LEDs - yeah, sure, I know about LEDs. Those little red lights in things like my Game Boy. Stands for... light emitting... diode.

Great, you continue, what's a diode?

Um.

This is already one of the fun things about playing with Arduino. There are all sorts of basic electronics stuff that I sorta know about, but couldn't explain to a five-year-old or to a rubber duck on my desk. Or just don't know at all. But Arduino is helped me tackle these topics in a practical, tangible way.

So, let's take a look at an LED together.

See the longer pin / leg sticking out of the red part? That's the anode terminal. The anode is the positive end of the LED. The shorter leg is the cathode - the negative side. Electrons will flow from the anode to the cathode when connected. You'll want to connect the positive end to something providing voltage, and the negative end needs to be connected to ground. All diodes are polarized, meaning they have these distinct positive and negative sides. And LEDs (light emitting diodes) happen to provide illumination when they're connected to an active circuit.

And they can be lots of pretty colors, too.

Okay, so here's our legally-required sketch for blinking an LED connected to digital pin 13 every half-second.

const int LED = 13;

void setup() {
    pinMode(LED, OUTPUT);
}

void loop() {
    digitalWrite(LED, HIGH);
    delay(500);
    digitalWrite(LED, LOW);
    delay(500);
}

Pretty simple, right? We first declare a constant variable for the pin we're using. The setup() function will run once per program, right before the loop() kicks off its infinite loop, so we'll just let the Arduino know that we want to set pin 13 to OUTPUT mode. And then during our infinite loop, we'll toggle the voltage to the pin by passing HIGH (5 volts) or LOW (0 volts) to our pin 13 using the digitalWrite function, pausing 500 milliseconds between these operations.

If you're coming from the Processing world, then this program structure of setup() and loop() should look very familiar, since it's literally the same.

The Arduino IDE provides an easy way to verify your programs compile before flashing them over to your actual Arduino, so I suggest clicking the Verify button first. This will catch syntax errors, like pesky missing semi-colons.

Next, we can set up our physical device.

I'm going to stick the LED into the Arduino, with the anode leg going into pin 13 and the cathode leg into ground. Note here that pin 13 is a special pin on the Arduino that has a resister built-in. If you try this with any other pin the Arduino, then the LED will burn out.

Finally, we can send our program from our computer to the Arduino over a USB connection by clicking the Upload button in the IDE. Your Arduino should flash happily once its complete, and then it's off to the infinite races.

Look at that blinker. Pretty great, huh? Note that this gif definitely speeds things up a bit.

Putting the "S" in USB

So, as I continued building stuff, I inevitably found myself wanting to console.log the heck out of a program that wasn't working.

Let's talk about printing stuff with Arduino.

Your Arduino is connected to your computer via a USB cable. USB. USB. That has to stand for something, right? It does. It stands for "Universal Serial Bus." USB is a quote "industry standard" for communications between computers and peripherals. If you think back really hard to the time of Captain Marvel or even earlier, you might remember other ways that we connected peripherals to our computers -- like an dot matrix printer's parallel port or a PS/2 keyboard port. Well, in the time since Carol Danvers left us here to fend for ourselves, USB has taken over our hearts, minds, and wallets. But we're still using a "serial connection" when we're using USB devices - so we'll need to use the serial protocol to communicate with our Arduino.

In other words, if we want to send or receive info from our Arduino program, we need to establish a serial connection with the device. Here's how you do that in a Processing sketch:

const unsigned int BAUD_RATE = 9600;

void setup() {
    Serial.begin(BAUD_RATE);
}

void loop() {
    Serial.println("Hello, world!");
}

Baud rate, huh? I know this baud term, too. Modems had baud rates, IIRC. Some Wikipedia-ing and Google-ing reveal that baud rates are the rates at which information is transferred in a serial channel. In this case, with a baud rate of 9600, we're transferring a max of 9600 bits per second. 9600 happens to be the standard baud rate for Arduinos, but I believe you can choose a different rate.

To view your "console", you can click the "Serial Monitor" button in the IDE.

In addition to viewing received information, you can also send messages back to the Arduino in this monitor using the text input on the top panel and the Send button. For example, you might write a program that toggled an LED on or off based on a specific input key.

What if you don't want to use the Serial Monitor in the Arduino IDE? Maybe it's time to let the old ways die. I agree. If you're on a Mac, then you can try running the screen command from your terminal, specifying both the name of your serial connection to your Arduino and the baud rate.

screen <name_of_serial_connection> 9600

In my case, the name of the connection was /dev/cu.usbmodem14101, which you can find in the Tools/Port menu of the Arduino IDE.

Word of warning, however. If you close this terminal window, it won't close the sesssion, and you'll be unable to Upload new programs to your Arduino. This is called a "detached screen" and it's annoying. You need to quit the screen somehow, and you can use this command to do so:

screen -X -S <name_of_session> quit

Oh, to get the name of the detached session, you can type:

screen -ls .

This whole serial communications thing opens up some interesting ideas, since you can have two way comms between your Arduino and something else. Forget Alexa. Not-okay, Google. Go away, Siri. Now you can build your own talking robotic best friend, instead. Hopefully gets some gears turning for you, too.

Counting in binary with LEDs

In general, life-goal-wise, I've been trying to get better at thinking and counting in binary, so I decided to build a little binary counter for my next Arduino project.

const unsigned int LED_BIT0 = 12;
const unsigned int LED_BIT1 = 11;
const unsigned int LED_BIT2 = 10;
const unsigned int LED_BIT3 = 9;

long result = 0;

void setup() {
  pinMode(LED_BIT0, OUTPUT);
  pinMode(LED_BIT1, OUTPUT);
  pinMode(LED_BIT2, OUTPUT);
  pinMode(LED_BIT3, OUTPUT);
}

void loop() {
  result++;
  if (result == 16) {
    result = 0;
  }
  output_result(result);
  delay(500);
}

void output_result(const long result) {
  digitalWrite(LED_BIT0, result & B0001);
  digitalWrite(LED_BIT1, result & B0010);
  digitalWrite(LED_BIT2, result & B0100);
  digitalWrite(LED_BIT3, result & B1000);
}

I'm not sure why the red LED isn't as bright as the other three LEDs. I tried swapping it out with another LED to no avail. But, hey, other than that, this thing works!

I also learned that breadboards are great. Being able to run all the cathode sides of the LEDs to the bottom negative row of the breadboard, and then only connecting that row once to the Arduino's ground port is pretty darn helpful. I have more to learn and appreciate here, for sure.

Also, this is the first time that I've really leveraged the power of the bitwise-and operator. I'm taking my result and bitwise-and it with a binary number that represents a binary digit (1's, 2's, 4's, 8's) for each of the LEDs. The bitwise-and operation returns true if result and our binary number both contain a 1 for the given binary digit. For example, let's look at the number 3

3 & B0001; // true
3 & B0010; // true
3 & B0100; // false
3 & B1000; // false

The final trick here is that digitalWrite function transforms true boolean values into HIGH (turn on the LED) and false into LOW (turn off the LED). So, for the number 3 the LED for the 1's digit and the 2's digit should be lit, and the 4's and 8's should be off.

That's pretty awesome and makes this code very concise. There's much more to explore here for me.

More tinkering

So, after a mere afternoon, I've learned a ton and had quite a bit of fun along the way.

What's next, you ask? Well, resistors are still perplexing. I'm not sure yet how to determine what level of resistence is needed for a given situation. I've already fried an LED (a delightful puff of smoke wisps out during its last gasp of life), likely for this very reason. It's also really hard to read those colorful bands to try to determine their resistance level. This seems like it could be a great little computer vision / deep learning app. Or perhaps I should just use my multimeter more regularly.

I'm also thinking more about the difference between analog and digital signals. Digital is binary (either on or off), whereas analog is continuous. Most of what we observe in life is an analog signal. So when we choose to digitize them, we need to choose specific moments to "sample" the values of the continuous signal. The Schmidt book explained that an audio CD takes a sample every 44,100 per second (or 44.1 kHz). Maybe this is why vinyl is back.

I thinking that my obvious next project here with Arduino is to make an alarm clock with binary numbers. There are tons of neat examples of this project across the web, and I think it could be a good way to learn / improve my soldering skills, as well as my quick mental binary counting, especially while groggy in the middle of the night.

Generating your own graduation speeches with Markov Chains

Charlie Harrington — Tue, 25 Sep 2018 18:37:59 +0000

Imagine this. You're the founder slash CEO slash product-visionary of a three month-old electric scooter startup. In between transforming the global transportation market and dabbling in your first angel investments (you know, just to get your feet wet), you've been asked by your beloved alma mater to deliver this year's commencement address to the graduating class.

That's right, you unfortunately didn't drop out of college, so you've already lost that convenient narrative thread. You agree to give the speech and assure yourself you'll be fine. You've got no problem waxing on about local city scooter politics or the unit-economics of the secondary charging market. That should give you a good five minutes or so. Maybe you can even ride a scooter on stage towards the podium for a laugh? Write that down, you mutter to no one in particular.

But as the graduation date approaches, the pressure's rising. Your plucky Chief-of-Staff slash travel agent slash only real friend anymore asks you how the speech draft's going, and you just smile and nod, "Great, Sam. It's about connecting the dots. In reverse."

"You mean like that Steve Jobs one? Stanford, 2005. I've watched the YouTube video like a million times."

"Oh, no, not like that. It's more about the benefits of failure and, you know, the importance of imagination."

"J.K. Rowling, Harvard, 2008. C'mon, you're not going to make that world's largest Gryffindor reunion joke, too. Are you? Say no, please. Say no right now."

"No, course not. Anyway, isn't it about time for my daily transcendental gratitude journaling? You almost made me miss it again. Give me one of your pens. Not that one, the other one."

You sit down at the communal lunch table, a glass of ginger green-tea kombucha and a terrifying blank piece of paper in front of you.

Right as – you swear – you were about to start writing the greatest commencement address of all time, one of those newfangled data scientists walks over and sits down, directly across from you. Can't they see that you're in-the-zone? And why are you paying them so much if they're just sitting around all the time?

"I think I can help you."

You glance up at them, barely.

"I overheard your conversation with Sam. Don't give me that look, it's an open office layout – probably your idea, too. Anyway, I think I can help you out. You need a speech. A good one. And quickly."

You drop your (Sam's) pen on the table, cross your arms, and lean backwards.

"I'm listening."

"Have you heard of Markov chains?"

A little bit of Markov in your life

In this post, I'll show you how you can easily generate your own overwrought and highly-sentimental commencement address clichés using a simple Markov chain library in Python and an open dataset of commencement speeches on FloydHub.

In just a few minutes, you'll be spouting gems like:

The secret to success in start-ups, or any other collaboration, is to stick an old head on a motorcycle weaving his way down the hall, he passed a door – it was empty.

Or perhaps this wisdom nugget:

Think of your ancestors: Among them, for everybody here, among your ancestors and I even thought about running away from the old ones.

Try it now

Click this button to open a Workspace on FloydHub where you can train a Markov chain model to generate "commencement speech style" sentences in a live JupyterLab environment. The commencement address dataset of ~300 famous speeches will be automatically attached and available in the Workspace. Just follow along with the speech_maker Jupyter notebook. It's that easy, folks.

But what's a Markov chain?

A Markov chain, named after this bearded devil, is a model describing a sequence of states (these states which could be some situation, or set of values, or, in our case, a word in a sentence) along with the probability of moving from one state to any other state.

The best explainer I've found for Markov chains is the visual explainer produced by Victor Powell and Lewis Lehe. Stop what you're doing right now and read their post.

Good, you're back.

As you just learned, Markov chains are popular modeling tools in a variety of industries where people need to model impossibly-large real world systems – finance, environmental science, computer science. Powell and Lehe point out that Google's big bad PageRank algorithm is a form of Markov chain. So they're, like, a big deal.

But Markov chains have also found a nice sweet spot in the text generation realm of natural language processing (NLP). Or, in other words, they're perfect for creating Captain Picard Twitter bots.

Generating the best speech ever

In our case, we want to use a Markov chain to generate random sentences based on a corpus of famous commencement speeches.

Luckily, there's a simple Python library for that first part. It's called markovify. I'll show you how to use it in just a second.

First, we need to get some speech transcripts. Ah, data — the cause of and solution to all of deep learning's problems.

NPR's The Best Commencement Speeches, Ever, a site I frequent often, was a great starting point. From there, some BeautifulSoup scraping – along with a shocking amount of manual spam removal from the transcripts found on popular commencement speech sites (ugh, don't ask) – led to the assemblage of a dataset containing ~300 plaintext commencement speech transcripts.

The commencement speech dataset is publicly available on FloydHub so that you can use it your own projects. I've also put together a simple Gatsby.js static site if you'd like to casually read the speeches at your leisure.

Okay, back to business. As mentioned, the markovify library is insanely easy to use. Let me remind you again to just click the "Run on FloydHub" button above to follow along in the speech_maker notebook. Actually, here it is again:

To generate our sentences, we're going to iterate through all the speeches in our dataset (available at the /floyd/input/speeches path) and create a Markov model for each speech.

import os
import markovify

SPEECH_PATH = '/floyd/input/speeches/'

speech_dict = {}
for speech_file in os.listdir(SPEECH_PATH):
    with open(f'{SPEECH_PATH}{speech_file}') as speech:
        contents = speech.read()

        # Create a Markov model for each speech in our dataset
        model = markovify.Text(contents)
        speech_dict[speech_file] = model

Then, we'll use a markovify's combine method to combine them into one large Markov chain.

models = list(speech_dict.values())

# Combine the Markov models
model_combination = markovify.combine(models)

Finally, we'll generate our random sentence:

print(model_combination.make_sentence())

Certainly I could do the most elegant and extraordinary products in federally funded highway projects.

You may have noticed that I've organized the individual speech models into a dictionary with the speech's filename as keys. Why did you do that, dummy? Well, if you keep following along with the speech_maker notebook, you'll see that this helps you more easily filter the speeches as you keep experimenting.

For example, maybe you only want to generate sentences from speeches delivered at Stanford University? Or at MIT? Or only from speeches delivered in the 1980s?

The workspace contains a CSV with metadata for each speech (speaker name, school, and year). This is your chance to make a dent in the commencement speech universe. Oh, here's an idea: why don't you brush off your Simpson's TV script Recurrent Neural Network (RNN) code and give it a whirl on this dataset?

Add 'Run on FloydHub' to your own projects

It can be a real pain in the you-know-what organizing your own machine learning experiments, let alone trying to share your work with others. Many folks wiser than I (or is it me?) have acknowledged the reproducibility crisis in data science.

The "Run on FloydHub" button is here to help. With this button, we're making it just a little bit easier to reproduce and share your data science projects.

Now, you can simply add this button to your repos on GitHub and anyone will be able to spin up a Workspace on FloydHub along with your code, datasets, deep learning framework, and any other environment configs.

Here's what you need to do:

Create a floyd.yml config file in your repo

machine: gpu
env: pytorch-1.4

Add this snippet to your README

<a href="https://floydhub.com/run">
    <img src="https://static.floydhub.com/button/button.svg" alt="Run">
</a>

You're done!
Seriously. Try it out right now on with our Object Classification repo. Or even this post's repo.

The real magic is when you also include the required datasets in your config file. For example, the floyd.yml config file for the Sentiment Analysis project looks like this:

env: tensorflow-1.7
machine: cpu
data:
  - source: floydhub/datasets/imdb-preprocessed/1
    destination: imdb

This will spin up your code in a CPU-powered Workspace using Tensorflow 1.7 as well as attach the IMDB Dataset. This makes it insanely easy for anyone to reproduce your experiments, right from your GitHub README.

You can read more about the Run on FloydHub button in our docs. We're looking forward to seeing what you share with the world! Good luck, graduates!

Teaching my robot with TensorFlow

Charlie Harrington — Fri, 21 Sep 2018 18:06:14 +0000

If you're like me, then you'd do pretty much anything to have your own R2-D2 or BB-8 robotic buddy. Just imagine the adorable adventures you'd have together!

I'm delighted to report that the Anki Cozmo is the droid you've been looking for.

Cozmo is big personality packed into a itty-bitty living space. You don't need to know how to code to play with Cozmo - but if you do - then Cozmo has even more phenominal cosmic power.

In this post, I'm going to show you how you can teach your own Cozmo to recognize everyday objects using transfer learning with TensorFlow on FloydHub.

The setup

Install the Cozmo Python SDK, create a new virtualenv, and clone the cozmo-tensorflow project to your local machine.

virtualenv ~/.env/cozmo -p python3
source ~/.env/cozmo/bin/activate
git clone https://www.github.com/whatrocks/cozmo-tensorflow
cd cozmo-tensorflow
pip install -r requirements.txt

Next up - login to the FloydHub CLI (sign up for a free account here). If you need to install the FloydHub CLI, just check out this guide in our documentation.

floyd login

1. Use Cozmo to generate training data

Getting enough training data for a deep learning project can be a pain. But thankfully we have a robot who loves to run around and take photos with his camera, so let's just ask Cozmo to take pictures of things we want our robot to learn.

Let's start with a can of delicious overpriced seltzer. Place Cozmo directly in front of a can of seltzer. Make sure that your robot has enough space to rotate around the can while it is taking pictures. Be sure to enter the name of the object that Cozmo is photographing when you run the cozmo-paparazzi script.

python3 cozmo-paparazzi.py seltzer

Repeat this step for as many objects (labels) as you want Cozmo to learn! You should now see all your image labels as subdirectories within the /data folder of your local directory.

Uploading dataset to FloydHub

Next up - let's upload our images to FloydHub as a FloydHub Dataset. This will allow us to mount these images during our upcoming model training and model serving jobs on FloydHub. Datasets on FloydHub are an easy way for your training jobs to reference a version-controlled dataset.

cd data
floyd data init cozmo-images
floyd data upload

In our case, I've named this image dataset cozmo-images. I've made it a public dataset, so feel free to use it in your own Cozmo projects!

2. Training our model on FloydHub

And now the fun begins. First, Make sure you are in the project's root directory, and then initialize a FloydHub project so that we can train our model on one of FloydHub's fully-configured TensorFlow cloud GPU machines.

Side note - if that last sentence sounded like a handful, then just know that FloydHub takes care of configuring and optimizing everything on your cloud machine so that it's ready for your GPU-powered deep learning experiments. You can specify the exact deep learning framework you'd like to use - whether that's TensorFlow 1.4 or PyTorch 0.3 or more - and FloydHub will make sure your machine has everything you need to start training immediately.

Okay, back to business, let's initialize our project:

floyd init cozmo-tensorflow

Now we're ready to kick off a deep learning training job on FloydHub.

A few things to note:

We'll be doing some simple transfer learning with the Inception v3 model provided by Google. Instead of training a model from scratch, we can start with this pre-trained model, and then just swap out its final layer so that we can teach it to recognize the objects we want Cozmo to learn. Transfer learning is a very useful technique, and you can read more about it on TensorFlow's website.
We're going to be mounting the images dataset that Cozmo created with the --data flag at the /data directory on our FloydHub machine.
I'm enabling Tensorboard for this job with the --tensorboard flag so that I can visually monitor my job's training process
I've edited the retrain.py script (initially provided by the TensorFlow team) to write its output to the /output directory. This is super important when you're using FloydHub, because FloydHub jobs always store their outputs in the /output directory). In our case, we'll be saving our retrained ImageNet model and its associated training labels to the job's /output folder.

floyd run \
  --gpu \
  --data whatrocks/datasets/cozmo-images:data \
  --tensorboard \
  'python retrain.py --image_dir /data'

That's it! There's no need to configure anything on AWS or install TensorFlow or deal with GPU drivers or anything like that. (If you're paying close attention, I didn't include the --env flag in my job command - that's because FloydHub's default environment includes TensorFlow 1.1.0 and Keras 2.0.6, and that's all I need for my training 😎).

Once your job is complete, you'll be able to see your newly retrained model in your job's output directory.

I recommend converting your job's output into a standalone FloydHub Dataset to make it easier for you to mount our retrained model in future jobs (which we're going to be doing in the next step). You can do this by clicking the 'Create Dataset' button on the job's output page. Check out the Dataset called cozmo-imagenet to see my retrained model and labels.

3. Connecting Cozmo to our retrained model

We can test our newly retrained model by running another job on FloydHub that:

Mounts our retrained model and labels
Sets up a public REST endpoint for model serving

Model-serving is an experimental feature on FloydHub - we'd love to hear your feedback on Twitter! In order for this feature to work, you'll need to include a simple Flask app called app.py in your project's code.

For our current project, I've created a simple Flask app that will receive an image from Cozmo in a POST request, evaluate it using the model we trained in our last step, and then respond with the model's results. Cozmo can then use the results to determine whether or not it's looking at a specific object.

floyd run \
  --data whatrocks/datasets/cozmo-imagenet:model \
  --mode serve

Finally, let's run our cozmo-detective.py script to ask Cozmo to move around the office to find a specific object.

python3 cozmo-detective.py toothpaste

Every time that Cozmo moves, the robot will send an black and white image of whatever it's seeing to the model endpoint on FloydHub - and FloydHub will run the model against this image, returning the following payload with "Cozmo's guesses" and how long it took to compute the guesses.

{
  'answer': 
    {
      'plant': 0.022327899932861328, 
      'seltzer': 0.9057837128639221, 
      'toothpaste': 0.07188836485147476
    }, 
  'seconds': 0.947
}

If Cozmo is at least 80% confident that it is looking at the object in question, then the robot will run towards it victoriously!

Once Cozmo's found all your missing objects, don't forget to shut down your serving job on FloydHub.

A new hope

It's a magical world, Cozmo, ol' buddy... let's go exploring!

I'm eager to see what you and your Cozmo can find together, along with a little help from your friends at FloydHub. Share your discoveries with us on Twitter!

References

This project is an extension of @nheidloff 's Cozmo visual recognition project and the Google Code Labs TensorFlow for Poets project. I also wrote about this project on my personal site - except with a lot more references to Short Circuit and The Legend of Zelda.