# Deep Learning Experiments on Google’s GPU Machines for Free

Update: If you are interested in getting a running start to machine learning and deep learning, I have created a course that I’m offering to my dedicated readers for just \$9.99. Practical Deep Learning with Keras and Python .

So you’ve been working on Machine Learning and Deep Learning and have realized that it’s a slow process that requires a lot of compute power. Power that is not very affordable. Fear not! We have a way of using a playground for running our experiments on Google’s GPU machines for free. In this little how-to, I will share a link with you that you can copy to your Google Drive and use it to run your own experiments.

# Deep Learning for Protein Function Prediction

Protein function prediction is taking information about a protein (such as its amino acid sequence, 2D and 3D structure etc.) and trying to predict which functions it will exhibit. This has implications in several areas of bioinformatics and affects how drugs are created and diseases are studied. This is typically an intensive task requiring inputs from biologists and computer experts alike and annotating newly found proteins requires empirical as well as computational results.

We, here at FAST NU, recently came up with a unique method (dubbed DeepSeq — since it’s based on Deep Learning and works on protein sequences!) for predicting functions of proteins using only the amino acid sequences. This is the information that is the first bit we get when a new protein is found and is thus readily available. (Other pieces require a lot more effort.)

We have successfully applied DeepSeq to predict protein function from sequences alone without requiring any input from domain experts. The paper isn’t peer reviewed yet but we have made the paper available as preprint and our full code on github so you can review it yourself.

We believe DeepSeq is going to be a breakthrough inshaallah in the field of bioinformatics and how function prediction is done. Let’s see if I can come up with an update about this in a year after the paper has been read a few times by domain experts and we have a detailed peer review.

# Writing Better English — Avoid Very

I return with a minor post after another long break. This time, it’s about writing better English. Now, this isn’t humblebragging but I cannot be considered excellent at English writing — at least not by native standards. English is not my first language and I haven’t had much formal English education. I have, however, read a lot. Even if my English is not good, I can still point out some tips shared by experts.

Here’s the first one of those shared by Amanda Patterson on Writers Write. It’s a list of 45 words you can use to put emphasis on words without using the word “very”. I found it refreshingly helpful.

Bear in mind though that you cannot just go ahead and use a word without looking up its usage examples. Some words might have negative connotations even though the dictionary meanings look positive. For example, if you use the word ‘adequate‘ to describe someone’s work, they might be offended even though the dictionary meaning is that of acceptable quality.

p.s. After writing this, I searched for the word “very” and found two instances where I had used the word myself. I replaced it with better alternatives.

# A Basic Naive Bayes classifier in Matlab

Update: If you are interested in getting a running start to machine learning and deep learning, I have created a course that I’m offering to my dedicated readers for just \$9.99. Access it here on Udemy. If you are only here for Matlab, continue reading =]

This is the second in my series of implementing low-level machine learning algorithms in Matlab. We first did linear regression with gradient descent and now we’re working with the more popular naive bayes classifier. As is evident from the name, NB it is a classifier i.e. it sorts data points into classes based on some features. We’ll be writing code for NB using low-level matlab (meaning we won’t use matlab’s implementation of NB). Here’s the example we’ve taken (with a bit of modification) from here.

Consider the following vector:

(likes shortbread, likes lager, eats porridge, watched England play football, nationality)T

A vector $x = (1, 0, 1, 0, 1)^T$ would describe that a person likes shortbread, does not like lager, eats porridge, has not watched England play football and is a national of Scottland. The final point is the class that we want to predict and takes two values: 1 for Scottish, 0 for English.

Here’s the data we’re given:

``` X = [ 0 0 1 1 0 ; 1 0 1 0 0 ; 1 1 0 1 0 ; 1 1 0 0 0 ; 0 1 0 1 0 ; 0 0 1 0 0 ; 1 0 1 1 1 ; 1 1 0 1 1 ; 1 1 1 0 1 ; 1 1 1 0 1 ; 1 1 1 1 1 ; 1 0 1 0 1 ; 1 0 0 0 1 ]; ```

Notice that usually when we represent data, we write features in columns, instances in rows. If this is the case, we need to get the data in proper orientation: features in rows, instances in columns. That’s the convention. Also, we need to separate the class from the feature set:

```Y = X(:,5);
X = X(:,1:4)'; % X in proper format now.
```

Alright. Now, that we have the data, let’s hear some theory. As always, this isn’t a tutorial on statistics. Go read about the theory somewhere else. This is just a refresher:

In order to predict the class from a feature set, we need to find out the probability of Y given X (where

$X = ( x_1, x_2, ldots x_n )$

with n being the number of features. We denote the number of instances given to us as m. In our example, n = 4, m = 13. The probability of Y given X is:

$P(Y=1|X) = P(X|Y=1) * P(Y=1) / P(X)$

Which is called the Bayes rule. Now, we make the NB assumption: All features in the feature set are independant of each other! Strong assumption but usually works. Given this assumption, we need to find $P(X|Y=1), P(Y) and P(X)$.

(The weird braces notation that follows is the indicator notation. $1{ v }$ means use 1 only if condition v holds, 0 otherwise.)

$P(X) = P(X|Y=1) + P(X|Y=0)$

$P(X|Y=1) = prod_j{P(x_i|Y=1)}$

To find $P(X|Y=1)$, you just have to find $P(x_i|Y=1)$ for all features and multiply them together. This is where the assumption comes in. You need the assumption of independence here for this.

$P(x_i|Y=1) = sum_j{1{x_i^j = 1, y^j = 1}} / sum_j{1{y^j = 1}}$

This equation basically means count the number of instances for which both x_i and Y are 1 and divide by the count of Y being 1. That’s the probability of x_i appearing with Y. Fairly straight forward if you think about it.

$P(Y=1) = sum_j{1{y^j = 1 }} / sum_j{1{y^j = 1, y^j = 0 }}$

Same as above. Count the ratio of Y=1 with the total number of Ys. Notice that we need to calculate all these for both Y=0 and Y=1 because we need both in the first equation. Let’s begin from the bottom up. For all of below, consider E as 0 and S as 1 since we consider being Scottish as being in class 1 (positive example).

P(Y):

```pS = sum (Y)/size(Y,1);     % all rows with Y = 1
pE = sum(1 - Y)/size(Y,1);  % all rows with Y = 0
```

P(x_i|Y):

```phiS = X * Y / sum(Y);  % all instances for which attrib phi(i) and Y are both 1
% meaning all Scotts with attribute phi(i)  = 1
phiE = X * (1-Y) / sum(1-Y) ;  % all instances for which attrib phi(i) = 1 and Y =0
% meaning all English with attribute phi(i) = 1
```

PhiS and PhiE are vectors that store the probabilities for all attributes. Now that we have the probabilities, we’re ready to make a prediction. Let’s get a test datapoint:

```x=[1 0 1 0]';  % test point
```

And calculate the probabilities P(X|Y=1) and P(X|Y=0)

```pxS = prod(phiS.^x.*(1-phiS).^(1-x));
pxE = prod(phiE.^x.*(1-phiE).^(1-x));
```

And finally, the probabilities of P(Y=1|X) and P(Y=0|X)

```pxSF = (pxS * pS ) / (pxS + pxE)
pxEF = (pxE * pS ) / (pxS + pxE)
```

They should add upto 1 since there are only two classes. Now you can define a threshold for deciding whether the class should be considered 1 or 0 based on these probabilities. In this case, we can consider this test point to belong to class 1 since the probability pxSF > 0.5.

And there you have it!

# A much better (and useful) eclipse plugin

It’s been four days since I started working with eclipse plugins and I finally have my first useful plugin. It’s useful for my research purposes and hopefully for a small audience interested in my work. It might also be useful for those trying to learn how to write eclipse plugins because I’ll soon be writing a tutorial on how I put this thing together from scratch (inshallah).

For the time being though, enjoy the screenshot.

Update: fixed the plugin with a new ‘View’. It now operates much better with a separate view for the output and controls. Also added is a ‘Hierarchy’ view for viewing the policy in a nice tree structure.

# Flashing Android Dev Phone 1

This tutorial is about flashing your Android Developer Phone 1 with your own custom build. It will provide a concise description of steps involved along with a special portion on how to port Google’s apps on your custom build. I found that particularly troublesome with little help on the Internet. So, that will be a bonus 🙂

First the disclaimer: This is for your Android Dev Phone 1 (ADP1). If you’re using T-Mobile’s SIM/firmware locked phone, stop. This tutorial is not for you. If you’re using ADP1, proceed at your own risk. You may brick your phone if you do something wrong and I shall not be held responsible for it. Finally, you might want to backup your factory-provided image. I don’t think it’s really necessary because you can just flash it again using the HTC provided images.

So, here is how it’s done:

# Getting Started with Android Dev Phone 1

We received our Google Android Dev Phone 1 yesterday and immediately ran into trouble. We don’t have a supported carrier here and we couldn’t get our own carriers to work with Android because we didn’t have the APN information. Android’s distro that comes bundled into the Dev Phone won’t let you in without an APN  though. You get a “SIM not found” message and you can’t do anything other than dial an emergency number. So, after searching for a while, I found some useful tips for getting around the problem.

First, you need to plug in your phone through the provided USB. If you’re running XP, the device will probably not be recognized. (It wasn’t for me.) So, download the Android phone driver here (or here) and install it when XP asks to search for a driver. (Thanks to anddev for this information.) After that, get the Android SDK from here. Go to command prompt and navigate to the tools directory in the SDK. Then execute these commands.``` ```

```adb shell su cd /data/data/com.android.providers.settings/databases sqlite3 settings.db INSERT INTO system (name, value) VALUES ('device_provisioned', 1); .exit reboot ```

Once the device finishes rebooting,``` ```

```adb shell am start -a android.intent.action.MAIN -n com.android.settings/.Settings ```

Many thanks to Android Tricks for writing this tip.

Update 1: Android SDK ships with the latest version of the windows Android phone driver. You can find it in \$ANDROID_SDK_HOME/usb_driver. So, you don’t need to download the driver using the links provided above.

Update 2: To get the Android device to work on Ubuntu 9.04 Jaunty Jackalope, you need to perform the following steps:

1. `sudo nano /etc/udev/rules.d/51-android.rules`
2. Add this line to the file: `SUBSYSTEM=="usb", ATTRS{idVendor}=="0bb4", MODE="0666"`
(You can get the 0bb4 value from lsusb for High Tech Corporation (i.e. HTC) if you work with a different phone)
3. `sudo chmod a+rx /etc/udev/rules.d/51-android.rules`
4. `sudo /etc/init.d/udev restart`
5. `adb devices` (to see the device)