Inspiration

A while ago, I stumbled upon this GIF on Twitter and I thought it was the coolest thing ever.

AK on Twitter: "#StyleGAN2 interps pic.twitter.com/pccZpxHrCo / Twitter"

StyleGAN2 interps pic.twitter.com/pccZpxHrCo

This is a StyleGAN2 interpretation, however, there are a lot of interesting things that can be done with autoencoders.

I am currently taking a neural image processing class at Purdue, and we recently covered generative adversarial networks (GANs). I found generative models really interesting and wanted to learn more about autoencoders.

The gist of an autoencoder

Autoencoders have a bottleneck architecture and are made up of an encoder and decoder.

The goal of the encoder is to reduce the dimensionality of the input and compress it into a latent space representation. The goal of the decoder is to reconstruct the input from the latent vector (encoder’s output).

For example, the encoder would compress this 5x5 matrix into a 3x1 latent vector, and the decoder would try to reconstruct the 5x5 input.

Getting my feet wet with MNIST

At first, I wanted to learn more about how autoencoders worked, so I created a basic implementation for the MNIST dataset.

After training for 30 epochs, here is a bilinear interpolation of the latent space. You can see in the top left, the zeros are very round. As you move through the interpolation, the digits gradually morph into other classes.

The following visualization is a representation of the digits in latent space colored according to each digit. It can be seen that a 0 digit has a vastly different shape than a 1.

Looking for a dataset

On Kaggle, I found the dataset “Face Mask Lite” which has over 23GB of images of faces with and without face masks.

One downside about this dataset is that the faces are overlaid with a clipart face mask instead of real face masks. For a short-term hackathon, this seemed reasonable to train with. However, it would be cool to continue this project with a more robust dataset.

Model

In this model, the input gets reduced down to a latent vector of shape (16, 16, 64). From this latent vector, we attempt to reconstruct the input with the decoder network. I was able to train on a small subset of the dataset (instead of download the full 23GB) for this hackathon.

Model: “model_1” _________________________________________________________________ Layer (type) Output Shape Param # ================================================================= img (InputLayer) [(None, 256, 256, 3)] 0 _________________________________________________________________ conv2d (Conv2D) (None, 256, 256, 16) 448 _________________________________________________________________ max_pooling2d (MaxPooling2D) (None, 128, 128, 16) 0 _________________________________________________________________ conv2d_1 (Conv2D) (None, 63, 63, 32) 4640 _________________________________________________________________ batch_normalization (BatchNo (None, 63, 63, 32) 128 _________________________________________________________________ leaky_re_lu (LeakyReLU) (None, 63, 63, 32) 0 _________________________________________________________________ conv2d_2 (Conv2D) (None, 32, 32, 64) 18496 _________________________________________________________________ max_pooling2d_1 (MaxPooling2 (None, 16, 16, 64) 0 _________________________________________________________________ conv2d_3 (Conv2D) (None, 16, 16, 64) 36928 _________________________________________________________________ batch_normalization_1 (Batch (None, 16, 16, 64) 256 _________________________________________________________________ leaky_re_lu_1 (LeakyReLU) (None, 16, 16, 64) 0 _________________________________________________________________ conv2d_4 (Conv2D) (None, 16, 16, 128) 73856 _________________________________________________________________ batch_normalization_2 (Batch (None, 16, 16, 128) 512 _________________________________________________________________ leaky_re_lu_2 (LeakyReLU) (None, 16, 16, 128) 0 _________________________________________________________________ conv2d_5 (Conv2D) (None, 16, 16, 128) 147584 _________________________________________________________________ conv2d_6 (Conv2D) (None, 16, 16, 256) 295168 _________________________________________________________________ batch_normalization_3 (Batch (None, 16, 16, 256) 1024 _________________________________________________________________ leaky_re_lu_3 (LeakyReLU) (None, 16, 16, 256) 0 _________________________________________________________________ conv2d_7 (Conv2D) (None, 16, 16, 512) 1180160 _________________________________________________________________ conv2d_8 (Conv2D) (None, 16, 16, 512) 2359808 _________________________________________________________________ up_sampling2d (UpSampling2D) (None, 32, 32, 512) 0 _________________________________________________________________ conv2d_9 (Conv2D) (None, 32, 32, 256) 1179904 _________________________________________________________________ dropout (Dropout) (None, 32, 32, 256) 0 _________________________________________________________________ leaky_re_lu_4 (LeakyReLU) (None, 32, 32, 256) 0 _________________________________________________________________ conv2d_10 (Conv2D) (None, 32, 32, 128) 295040 _________________________________________________________________ conv2d_11 (Conv2D) (None, 32, 32, 128) 147584 _________________________________________________________________ dropout_1 (Dropout) (None, 32, 32, 128) 0 _________________________________________________________________ leaky_re_lu_5 (LeakyReLU) (None, 32, 32, 128) 0

Results

We tried running the autoencoder on faces with real face masks, however, the results were not as accurate as the clipart face masks. We even tried using a hacky approach of overlaying a clipart face mask on top of a real face mask, however, that didn’t really work. Although a dataset with real face masks would introduce a lot of variabilities, the results would be very satisfying to see if there were enough training samples.

Source code for Mask Autoencoder: https://colab.research.google.com/drive/1_kNQl1yPUqOWRGQVxO5MK4dq_pEbkxOW

Sources + tutorials

• https://towardsdatascience.com/understanding-latent-space-in-machine-learning-de5a7c687d8d

Floating in latent space was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

We ended up winning 2nd place overall by building a hack that uses bacteria in mud and your saliva to quickly tell you if your blood glucose concentration is abnormally high.

Check out our hack here: https://devpost.com/software/gluclose.

Meet the team

I had just started my first semester in college, and I desperately needed a team for PennApps. I was lucky enough to find some amazing team members from all over North America through the power of Facebook. We all had different skill sets and backgrounds, which made us a very strong, diverse team.

Brainstorming

The hardest part of any hackathon is the idea. We spent almost 8 hours trying to come up with one. We wanted to make something absolutely crazy, something that separates our hack from the rest. At around 12am, we gave up on ideation and decided to go ice skating at the UPenn Ice Rink despite feeling unaccomplished and behind other teams. After that, we hung out with some friendly UPenn students who showed us around and took us to Wawa (East Coast thing apparently).

The Idea

It wasn’t until 3am when we finally sat down together to have another serious brainstorming session. Olivia told us she did some research on exoelectrogens, which are bacteria that can generate electricity by transferring electrons extracellularly. We were curious whether we could harness that electricity for something useful. We finally put two and two together and realized that these exoelectrogens required glucose for energy. What if we could use the sugars from our saliva as an energy source for the bacteria?

And so we came up with an idea: Spit into some mud and a user-friendly app will easily analyze the data collected from your saliva sample, quickly telling you if your blood glucose concentration is abnormally high.

Building

Since we started building our hack at around 4am, we were way behind other teams. When we went downstairs to the MLH Hardware Lab, we were able to get our hands on a voltmeter, but all the Arduinos and bluetooth sensors were checked out. We lucked out when we found out our neighboring hackers brought an extra Arduino from home.

Before we were sold on the idea, we wanted to see what voltage we could get from a small sample of mud. We ran outside the Engineering building in the pouring rain and scooped a cup full of mud. We used two small pieces of pencil lead as electrodes to create an electric potential difference. We were able to get a reading of 20–30 millivolts, which was a major relief. Our next step was to get the data from the voltmeter and somehow send that to an iOS device. Instead of taking apart the voltmeter, we tried to make one ourselves using the Arduino.

For the next 20 hours we were stuck on building a sufficient voltmeter. We had a lot of trouble measuring low voltages using the Arduino. We tried different resistor configurations, attempted to find a voltage sensor, and tried increasing the voltage with an op-amp. At one point, we joked about using an iPhone camera with computer vision to get the reading off the voltmeter. Eventually with the help of our amazing mentor Sami, we resorted to using an op-amp comparator to analyze our voltages.

As we tested our hack more and more, we realized that we needed more lead. We all walked to CVS, bought a 10 pack of Ticonderoga #2 Pencils, borrowed someone’s scissors, and started cutting up some pencils.

Sleep?

“It’s a hackathon… sleep is for the weak!”

I only got around 2–3 of good sleep on the first night. Ice skating, walking around the UPenn campus, and running up and down the stairs a million times made me dead tired. I remember our team attended the workshop on “Self Programming AI by Bloomberg,” and we all fell asleep within 10 minutes.

We continued to work on our project around the clock. The second night I tried a Monster energy drink for the first time and ended up drinking 2 whole cans to stay awake (horrible for me, I know 😬).

Running against the clock

There wasn’t a lot of time left. In 8 hours we had to get our makeshift voltmeter working, send the data from the Arduino to a computer, and then from the computer to an iPhone. After hours of frustration, we finally got the Arduino voltmeter to work. Unfortunately, we hadn’t paid much attention to the clock — we only had 2 hours left until the submission deadline.

We quickly set up the PySerial library to access the Arduino through a serial connection. With an hour remaining until submission, we were scrambling to get ngrok functional, but it wasn’t working. Ian and I were debugging from every possible angle, but we just could not get it to work. An hour before the expo, our whole team completely gave up. We just did not have the energy to get this thing working. When they started kicking us out of our hacking area to set up tables for the expo, I rushed downstairs, grabbed a random table, and tried to use Firebase as an alternate approach. By some miracle, we got our hack working 30 minutes before the expo.

Just as I was starting to feel relieved, my eyelids got heavy. Our whole team took a 30 minute power-nap before the expo.

The Expo

After being through so much, we didn’t even focus on winning. We were just proud to be presenting a working hack by the end of the 36 hours.

I had the amazing opportunity to walk around and see some of the other hacks and demos. As I walked outside the bathroom door, I got a notification on my phone:

With a rush of adrenaline, I sprinted upstairs, and gathered with my team before presenting in front of many more hackers and judges.

When we found out we were one of the Top 10 teams, we freaked out. We never expected to get that far. Without much time to internalize our success, we had prepare for the Top 10 presentation. We ran through the rain from the Engineering building to Meyerson Hall with the hope that our project wouldn’t get ruined.

As we sat in the front row waiting to demo our project, we were all nervous. I was even more nervous considering I had to catch a flight back to Purdue in a couple hours. After we got on stage, I grabbed the microphone and told the audience, “For the past 36 hours, our team was hard at work hacking mud.” We blew everyone’s minds, and gave an phenomenal presentation.

Since we didn’t win any category prizes, we thought we were going home empty handed. As we lost all hope of winning, we heard “…and the 2nd place winner is… the dirty team!” We jumped out of our seats with disbelief and took the stage.

After we were awarded medals and took a ton of pictures, I sprinted to my Uber and drove to the Philadelphia International Airport. I ended up catching my flight on time, and I made a new friend while waiting in line for the security checkpoint.

This hackathon was an emotional rollercoaster. I felt nervous, sleepy, excited, caffeinated, and adrenaline-rushed at the same time. One thing’s for sure: I will 10/10 do this again (hopefully with more sleep 😄).

Building an emergency glucometer using mud and saliva was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Sometimes, I run into a different set of problems. For example, how do I make an accurate neural net if my dataset is too small? Or, how do I train a deep network if I don’t have GPUs?

This is where transfer learning comes in.

Transfer Learning — the improvement of learning in a new task through the transfer of knowledge from a related task that has already been learned. — Handbook of Research on Machine Learning Applications.

With transfer learning, you can take the knowledge that a neural network has learned from one task, and apply that knowledge to a similar, but different task. This approach works very well on Convolutional Neural Networks (CNNs) because they are feature-based networks. The first layers of a CNN detect simple features that can be recognized easy (ex. lines, edges, corners). The last layers of a CNN detect more abstract features in the image (ex. features that differentiate a cat from a dog).

Transfer learning works by pretraining your CNN on a dataset, freezing the feature layers of the CNN, and then training again on a similar dataset so that classification layers are fine-tuned for new dataset.

An Example

“Talk is cheap. Show me the code.” — Linus Torvalds

In the following implementation of transfer learning, I will be making a CNN to classify whether an image is a sunflower or a tulip.

First, I will pretrain the model on a dataset of flowers that contain daisies, dandelions, and roses. Then, I will freeze the feature layers and fine-tune my classification layers by training on my actual dataset of sunflowers and tulips.

Dataset: http://download.tensorflow.org/example_images/flower_photos.tgz.

<a href="https://medium.com/media/735875228401422eb565aeb5e411d32a/href">https://medium.com/media/735875228401422eb565aeb5e411d32a/href</a>

Pretraining on daisy, dandelion, and rose images:

Freezing feature layers and training on sunflower and tulip images:

Notice how the accuracy of the neural net starts at 84% and each epoch only takes 11 seconds to train.

For the full source code, check out the project on Github:

nikhiljay/transfer-learning

The Big Picture

So, where am I going with all of this?

In a realistic setting, image datasets are not going to be very large, so transfer learning is very important. You can pretrain a CNN on a very large dataset such as ImageNet, and then fine-tune the higher-level layers of the CNN by training on your own dataset. Since ImageNet contains 14 million images with 20,000 categories, pretraining on ImageNet would be a nightmare (with or without GPUs). However, people have publicly released their pretrained CNNs, so that we can use them. Some examples of models for image classification with weights trained on ImageNet are: Xception, VGG16, VGG19, ResNet50, InceptionV3, InceptionResNetV2, MobileNet, DenseNet, and NASNet.

Keras has a cool way of implementing these models. For example to use the VGG19 dataset, all you need to do is:

model = applications.VGG19(weights = "imagenet", include_top=False, input_shape = (img_width, img_height, 3))

In other words, there is no need to write a CNN from scratch. Just use transfer learning.

Why I’m interested in transfer learning — it’s humanlike.

One of the reasons why I got engrossed in machine learning is because of the similarities between animal biology and neural nets. For example, CNNs work by using multiple convolutional and fully connected layers. Well, that’s exactly how the human eye works. There are 126 million rods and cones, 6 million bipolar cells, and 2.5 million ganglion cells in the retina of the human eye (each layer has fewer cells than the previous layer).

Similar to CNNs, transfer learning is very human-like. For instance, our brain can categorize certain shaped objects as flowers and analyze the more abstract features of the object to recognize which species the flower is a part of. I think if we want to create better neural networks, we need to thoroughly understand the human sensory system.

Transfer Learning was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

As I experimented with many neural nets on different data sets, I noticed that dealing with matrix multiplication and the dimensions of layers can become very tedious.

One framework that has captured my attention is Keras. Keras is a deep learning API, written in Python that can run on Tensorflow. I like Keras because it’s very easy to use. The framework allows you to go from concept to reality in a more practical way than other complex machine learning frameworks. Additionally, Apple introduced CoreML in WWDC 17 allowing developers to integrate their Keras models into iOS apps. Since I have a background in iOS development, I found this feature appealing.

So far, I have been able to parse data from a variety of datasets, train successful models, and even integrate some of my models into iOS apps.

You can see all my recent work at my Github repo: https://github.com/nikhiljay/ml-projects.

Using Keras, I have been working on the following projects:

1. Boston Housing — Predict the values of houses in Boston based on: per capita crime rate by town, proportion of residential land zoned for lots over 25,000 sq. ft., proportion of non-retail business acres per town, nitric oxides concentration (parts per 10 million), average number of rooms per dwelling, weighted distances to five Boston employment centers, index of accessibility to radial highways, and pupil-teacher ratio by town. Data from CMU Statistics Department (http://lib.stat.cmu.edu/datasets/boston).

2. EEG Eye State — Predict whether your eyes are open/closed based on your brainwaves. Data from UCI ML Repo: EEG Eye State Data Set (https://archive.ics.uci.edu/ml/datasets/EEG+Eye+State).

3. Predicting E. Coli Protein Localization Sites — Predict the localization site of a protein based on accession number for the SWISS-PROT database, McGeoch’s and von Heijne’s method for signal sequence recognition, von Heijne’s Signal Peptidase II consensus sequence score, presence of charge on N-terminus of predicted lipoproteins, score of discriminant analysis of the amino acid content of outer membrane and periplasmic proteins, score of the ALOM membrane spanning region prediction program, and score of ALOM program after excluding putative cleavable signal regions from the sequence. Data from UCI ML Repo: Ecoli Data Set (http://archive.ics.uci.edu/ml/datasets/Ecoli).

4. Cardiovascular Disease — Predict whether a subject has cardiovascular disease based on age, sex, chest pain type (4 values), resting blood pressure, serum cholesterol in mg/dl, fasting blood sugar > 120 mg/dl, resting electrocardiographic results (values 0,1,2), maximum heart rate achieved, exercise induced angina, old-peak = ST depression induced by exercise relative to rest, the slope of the peak exercise ST segment, and number of major vessels (0–3) colored by fluoroscopy. Data from UCI ML Repo: Statlog (Heart) Data Set (http://archive.ics.uci.edu/ml/datasets/statlog+(heart))

5. Iris Flower Classification — Classify the specific species of an Iris flower based on sepal length, sepal width, petal length, and petal width. Data from https://en.wikipedia.org/wiki/Irisflowerdataset .

6. Parkinson’s Disease — Predict whether subject has Parkinson’s disease based on biomedical voice measurements. Data from UCI ML Repo: Parkinson’s Data Set (https://archive.ics.uci.edu/ml/datasets/parkinsons)

7. White Blood Cells — Classify different types of white blood cells based on images taken from a microscope.

ML Update: Keras and CoreML was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Basics

Before learning how to code a neural network, I needed to learn how a neural network works in general. I used MIT 6.034 Artificial Intelligence Lecture 12a on Neural Nets to understand the basic components of a neural network. With this lecture, I was able to understand how forward propagation and backpropagation worked.

Numpy

After learning the theory behind a neural network, I decided to learn numpy which is a library for multidimensional arrays and high-level mathematical functions. This is a great starting point for any beginner learning machine learning because it allows you to truly understand the math behind a neural network. I used iamtrask’s tutorial to learn how to code a two-dimensional and three-dimensional neural network in numpy. You can look at my code here. While other machine learning libraries have predefined methods for softmax/activation functions and backpropagation, numpy forces you to code each component by hand. To fully wrap my head around this complex topic, I used Siraj Raval’s video on Activation Functions and CS231n Lecture #4 from Stanford.

Pytorch

Tensorflow and Pytorch are both libraries for machine learning, and both have their pros and cons (look them up if you care). After trying out both, Pytorch seems a little bit more intuitive and easier to wrap my head around.

I made a two layer neural network, but unlike numpy the neural network had an input layer with 1000 dimensions, a hidden layer 100 dimensions, and an output layer with a one dimension. I also added a batch size since I had so many input dimensions, and a learning rate to optimize my neural network. You can look at my code here.

MNIST

I was very excited after training a basic neural network, but I really wanted to apply what I had learned. What if I could use machine learning to classify handwritten digits?

MNIST is a large database of handwritten digits that is great for trying out different machine learning algorithms. With the help of Justin Francis’s MNIST tutorial I was able to train a neural network to classify handwritten digits in TensorFlow. You can look at my code here.

I came up with a function called predict which took an input of a sample digit and outputted a prediction.

After finishing my neural network, I tried out different sample digits. The network only required a couple hundred epochs for 85% accuracy. For most of the digits, the neural network worked perfectly by predicting the correct number. But then I ran into this:

When I used only 200 epochs to train my model, the prediction came out to be “6”. I can see how the model came up with 6, but the correct answer is 0. Only after 2637 epochs does the model predict the correct answer: 0.

The Future

I have a long way to go in machine learning, and that’s why I will be updating this post with my progress every so often. After I understand more of the fundamentals of neural networks, I will try to learn about Convolutional Neural Networks and Generative Adversarial Networks. I will also try to apply my knowledge to interesting examples because that’s how I like to learn.

My Journey into Machine Learning was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

“Biomedical engineering — the application of engineering principles and design concepts to medicine and biology for healthcare purposes (e.g. diagnostic or therapeutic). This field seeks to close the gap between engineering and medicine, combining the design and problem solving skills of engineering with medical and biological sciences to advance health care treatment, including diagnosis, monitoring, and therapy.”

All technological discoveries in the field of biology fall under this subject. A few ingenious examples of bioengineering in the past include the x-ray, electroencephalogram (I made a project about the EEG), and magnetic resonance imaging (MRI). The first application of biomedical engineering occurred in World War II. This is because there were constant technological developments, and scientists harvested this advanced technology into their medical devices. This new branch of science became famous because biomedical engineering provided more efficient medical treatment.

Although biomedical engineering sounds specific, there are many fields within this subject that are even interesting. Specifically, Bioinformatics, Biomechanics, Tissue Engineering, Genetic Engineering, Neural Engineering, and Pharmaceutical Engineering have had a huge explosion of growth over the past decade.

Why am I even writing this? To get myself interested in pursuing a career. What better way to do research than to write about it?!

What is Biomedical Engineering? was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Several months ago, I decided to take a class that would change my perception on technology and the human brain.

If you did not already see my research proposal, then click here to read it first. After months of hard work and dedication, I finally finished my project. The ability to display brainwave data on a smartphone.

My Initial Plan.

I first created a plan with my mentor Matt Yun. In the AAR program, every student is paired up with their own mentor in that specific field. Lucky for me, my mentor already had previous experience with brainwaves and EEG headsets.

My plan was to find a way to extract the data from the Emotiv headset. After doing so, I would devise a way to send the brainwave data to a server and receive that data on a smartphone. But for any high schooler, this was not an easy task to do with the amount of work given to students in school.

I had an idea to go with my friend to a Treehacks, Stanford University’s official 36-hour hackathon. A hackathon, or an event in which a large number of people meet to make all sorts of creative projects out of code, would be the perfect place to work on this project. Unfortunately due to age restriction, we were not even allowed to attend.

After being kicked out of the hackathon, we got some Jamba Juice, ordered a pizza and started hacking in a Starbucks on the Stanford campus. What only felt like a few minutes, was a few hours. It was 10pm. Starbucks was closing and I still was finding a way to hack the Emotiv headset. I needed to come up with another plan.

Plan B.

After asking my mentor for some advice, I tried playing with the Emotiv headset a few weeks later. My main challenge was to find a way to decrypt the data being sent from the Emotiv. After a few days of playing with Python and bash scripts I finally was successful enough to see the raw stream of data being sent to the computer. I had just conquered my biggest challenge of all.

Completing half of the project gave me the inspiration I needed to finish this project. I texted my friend to meet me at the library and we moved on to the next phase of the project. All I needed to do was find a simple way to send the data from the computer to an iOS device. I used ngrok, or a service that creates secure tunnels to localhost. Then an iOS app was created so that the user could open it and receive the information sent from the computer. After receiving the data on the iOS device, I parsed it into JSON using Alamofire. Lastly, I formatted the data into the app.

The Final Steps.

After finishing the app, I did some more research to give the raw data some purpose. The few months of hard work was worth it. I had done it. I finished it. If you want to read my complete project, click here.

AAR Final Wrap-Up was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

The Feel.

Starting on the outside, this mouse looks amazing at first glance. The raised body feels nice to the hand and I haven’t felt uncomfortable yet. Coming from the Apple Magic Mouse, this mouse has been a game changer. Now it’s not as awkward to left or right click the mouse. This is a perfect mouse for working long hours on the computer and it helps prevent hand pain.

The Outside.

With the MX Master, Logitech added an awesome mosaic pattern on the gesture button. They also have a nice brass-looking finish on the edges of the mouse. Next to the gesture button is a 3-light battery indicator that shows the status of the rechargeable battery. Above the gesture button, are 2 buttons as well as a horizontal scrolling wheel. Using the Logitech Options Free software that they offer, it is possible to customize the buttons and scroll wheel to do different things. On top of the mouse, is a “speed-adaptive” scroll wheel that can shift from a ratchet scroll, to a hyper-fast scroll. Next to the scroll wheel are the left and right click buttons.

The Back.

The back of the mouse is also packed it some awesome features. At the top, there is the on-and-off switch and the connect button. Below that, there is the “Darkfield Laser Sensor” that tracks on any surface including glass. And the last button on the back of the mouse is called the “easy-switch”. The button allows a user to pair the mouse with up to three devices and switch between each device with a press of this button.

Features and Feedback.

The first unique feature on the mouse is the speed-adaptive scroll wheel. The ratchet mode, which is when the wheel moves in a click-to-click scroll, is very useful for scrolling through a document in detail. The hyper-fast scroll, which is when a flick of the finger spins the scroll wheel freely until it loses momentum, is very useful for browsing through long documents or web pages.

Unlike most other mice, the MX Master also features a unique thumb wheel for side-scrolling. However my favorite feature of this mouse is the ability to program all of the buttons of the mouse. Logitech includes software for the Master MX that, in my opinion, goes well with the mouse.

This allows the user to modify all buttons on the mouse including the adaptive scroll wheel and the thumb wheel on the side. Using the Logitech Options software, the buttons can be programmed to perform a variety of actions. And best of all, some of the programmable actions are native to the operating system.

However, before buying this mouse there are some downsides. First of all, this is not a gaming mouse. It does not have the features of a gaming mouse. If you want a gaming mouse, then go for something like the G900 Chaos Spectrum also by Logitech. Another downside is the price. At $100, this is one of the most expensive mice out there.

Overall, I think that the Logitech MX Master is the best mouse from a productivity/work standpoint. It feels great in the hand and definetely helps with my workflow.

Logitech MX Master Review was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

Research Question

How can an application that measures emitted brain waves give feedback based on a person’s behavior?

Background and Significance

Brain waves, or neural oscillations, were first observed by German psychiatrist Hans Berger in 1924. Brainwaves are vibrations emitted from the brain when neurons communicate with each other.

There are several types of brain waves, and each type of wave is used in different circumstances. When we are tired, slow, sleeping, or dreaming, we are emitting oscillations with low frequencies. When we are awake or alert, we emit oscillations with higher frequencies. According to Brainworks Neurotherapy, the different types of brain waves can be categorized according to their frequencies into four groups: delta (.5 to 3 Hz), theta (3 to 8 Hz), alpha (8 to 12 Hz), beta (12 to 38 Hz), and gamma waves (38 to 42 Hz).

According to Brainwave College, these waves are very important because they represent the activity that is occurring in the brain. But the waves don’t only represent brain activity, they can also determine the brain’s behavior. Brain stimulation can alter the patterns of brain waves thus changing the state of mind. An alarm clock is an example of brain stimulation because it is changing delta waves into beta waves (an asleep to awake state). Medications or recreational drugs are the most common methods to alter brain function.

Developed by Hans Berger, Electroencephalography (EEG) is a human-made method to track brain activity without affecting the subject. By placing several electrodes nodes on the scalp of a person, the EEG is able to detect voltage fluctuations of the person’s brain. According to WebMD, this is commonly used by medical professionals to monitor the depth of anesthesia and diagnose epilepsy, catatonia, and seizures. This method of detecting brain waves is extremely uninvasive unlike Electrocorticography, which requires electrodes to be placed on the surface of the brain instead of the scalp.

The device that I will be using to measure brain waves is the Emotiv EPOC. According to Emotive EPOC Product Specifications, the device is a scientific contextual wireless EEG system that has 14 EEG channels and 2 references. Using this device, I will be able to receive brain wave data from the Emotiv which can then be parsed through the server.

Research Methodology

The first step of the research will be to establish a way to attain access to the raw stream data from the Emotiv EPOC headset. For this task, I will be using “EmoKit”, which is an open source driver created by the OpenYou Organization.

Next, I will need to create a mechanism that can parse the data received from the Emotiv to the server. For this I will use Node.js, which is a Javascript Runtime for developing server-side web applications. With Node.js I will be able to receive data on any device, whether it is a browser on a computer, or an app on an iPhone.

After the Emotiv data is synchronized on another device, the data can then be used to construct a graph for the front-end user to visualize. To construct graphs, I will use a third-party library (not decided yet). Another feature that this project will incorporate is to be able to analyze the data and trends, so that the application give feedback based on the data. The app will also have an appealing user interface and experience for the website and iOS app. Finally, this project will be published on a website, uploaded to the App Store, and open-sourced.

Human Subjects

Humans subjects will be used in this project. With human trails, I can get the data of different individuals to test the application. I will be filling out the HASRC application form to ask for permission to use human subjects in my project.

Citations

“What Are Brainwaves?” Brainworks Neurotherapy, n.d. Web. 08 Dec. 2015.

“Why Are Brainwaves Important?” Brainwave College, n.d. Web. 08 Dec. 2015.

“Electroencephalogram (EEG) Test to Measure Brain Electrical Activity.” WebMD, n.d. Web. 08 Dec. 2015.

Brain Computer Interface & Scientific Contextual Eeg. Emotiv EPOC (n.d.): n. pag. Web.

Research Proposal — Measuring Brainwaves and Monitoring Behavior was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.

As Apple’s Annual World-Wide Developer Conference approaches next week, many developers are excited to come together to learn about the future of iOS and OS X. At WWDC, Apple features more than 100 technical sessions, over 1,000 Apple engineers, hands-on labs, and the Apple Design Awards.

How do you attend WWDC? First of all you need to join the Apple Developer Program. To get a ticket for attending this conference you need to apply in a random selection. If you are chosen, then you need to pay $1,600 for a ticket. If you are not chosen, better luck next time!

Luckily, I was able to win a WWDC scholarship. My dream since I was a little kid was to attend an Apple Conference. My passion for designing and coding has pushed my ability to make awesome things. To win a WWDC Scholarship, you need to apply by building an app and sending it to Apple. Apple will then select 350 scholarship winners world-wide if they like your app. The process is pretty simple. Build an app that is creative and talks about yourself. Submit the app to Apple and fill out some information. And finally receive the email!

Receiving this scholarship was not enough for some scholars including myself. A group of about 20 scholars and I decided to build a Scholars of WWDC App that would display the works of the scholarship winners and how you can connect with us. With this app, we are able to showcase our talents and inspire future generation to build things, to make a difference and to change the world. You can check this out at the WWDC Scholars website or download from the App Store.

WWDC is definitely a great place to learn new technologies and take previous knowledge to the next level. I am going to be post another blog post with my experiences soon.

Winning the WWDC Scholarship was originally published in Nikhil's Blog on Medium, where people are continuing the conversation by highlighting and responding to this story.