The Hard Way

Everyone loves to personalize things they own, why should the RaspberryPi be any different ? One of the many ways you can personalize it is to change the name of the device is to change its name. This device name also known as a hostname appears in many places over the device, and usually at the start of every line when you are connected using SSH.

The default is "raspberryPi", so lets change that. For the example of this tutorial i will be renaming my Raspberry Pi to be called "Pumpkin". Step one is to edit the hosts file to ensure that the device knows when it should refer to itself. we can do this by using VIM or VI ("either text editors work"). Enter the following :

sudo vi /etc/hosts OR sudo vim /etc/hosts

This file contains a list of all the system set dns to ip resolutions. What you will need to do is find the line that starts "127.0.1.1", after this is the current name of the device, change this to the device name u want. So mine would look like :

127.0.1.1           Pumpkin

Save this file and then return back to the console. The next file to change is the hostname file so again we open the file in VI or VIM with the following command :

sudo vi /etc/hostname

OR

sudo vim /etc/hostname

This file contains only the current hostname and should be replaced with your desired name, save and close this file.

The final stage is to commit these changes to the system  by entering the following line :

/etc/init.d/hostname.sh

We all have our ways of keeping our lives in order, from a scrap of paper in the pocket, to an overcomplicated digital personal organiser. If you are already using your Raspberry Pi for day to day tasks then the following program might be right down your alley.

Calcurse is a text base calender software offering scheduling, task management and reminders system. Being a lightweight textbased application, it sits right at home on your Raspberry Pi. To install it we simply use Apt-get to grab the latest version, type the following

sudo apt-get install calcurse

As usual you will need to authenticate the sudo command. THanks it, you should now be able to launch Calcurse by typing

Calcurse

You dont need to leave it open to recieve the reminders as the application launches its own background service to ensure that reminders are monitored. Much like many other linux applications, you can also launch it with certain parameters to get information quickly without launching the apllication fully. For example 

Calcurse -a

Will print to the screen a list of all outstanding appointments.

This Blog post will look into how you train a neural network, specifically areas used during my research. There are three types of training that neural networks can undergo depending on both the type of data that is being provided for training, and the required use of the neural network.

The first is supervised training, a method which involves presenting an input to the Neural network and then comparing its output to the desired output. These difference between the two outputs (known as the error delta) would then be used to adjust the weights based on their contribution to the error delta. This is performed for every line within the training set to reduce the overall error of the network. This type of training is commonly used for networks that are being trained for a known problem, environment or scenario, where substantial training data is available. It is also used to train neural networks to simplify existing processing heavy mathematical functions without introducing a significant error.

The second training method is Unsupervised training, a method commonly used for data sets that do not have an error or there is not a function to evaluate its fitness. This method allows the identification of patterns or categorisations within the data as similar data patters will provide a common output. This method of data clustering allows the extraction of features from within the data, along with identifying other fitting data, commonly used for function problems that involve categorisation, as outlined in Unsupervised Learning: Foundations of Neural Computation (Hinton & Sejnowski, 1999).   The third method of training is reinforcement training. This method of training is designed for neural networks that represent the thought patterns of agents within an environment. Commonly used in games it allows for agents to be trained based on their actions without providing the initial inputs. This requires agents to use a balance of existing trained knowledge or initial programmed knowledge along with new explorative knowledge to complete the required task. As the agents either succeeds or fails certain criteria it receives either a positive or negative signal respectively, allowing to improve at the task during its next attempt. Back propagation is a method of minimising the error of a function and is commonly used when training networks that have been in use since 1985 (Rojas, 1996). After the feed forward process is finished the value of the output nodes are compared to the outputs that are expected. The difference in these values is known as the error delta. Each weight that is connected to a node with an error delta is considered responsible for the generation of the error. During the process of back propagation the error is reduced by adjusting all the connected weights by their contribution to the error. This process occurs on all weights leading back to the input nodes where the forward pass began.  

An activation function is the mathematical function that is used to determine the value of a neuron based on its inputs. The activation functions required for a particular shape of network varies depending on the data presented and expected from the Neural Network. There are a number of different activation functions that are commonly seen in neural networks. Binary activation functions provide a two state signal that is based upon the whether the inputs of the node are above or below a set threshold. There are some activation functions that offer an analogue signal based of the sum of the inputs. These activations include a simple linear output, where the sum of the inputs is directly mapped to the output. A number of activations functions adhere to the behaviour of neurons in the brain by providing a form of normalisation to the output signal with the use of a hyperbolic tangent, causing it to have both an upper and a lower bound, commonly ranging from -1 to 1. Below are a number of Activation functions outputs graphed against their inputs. 

This blog post outlines a number of types of neural networks I have worked with during my research

ANNs (Artificial Neural Networks) are an algorithmic way of modelling data categorisation, pattern recognition and other simplified tasks performed by neurons within the brain (Bar-Yam, 2003).

They are commonly used for predicting data in a way that provides better accuracy then simple or even complex fixed algorithms. Neural networks are commonly used in the financial market for a number of different tasks such as risk prediction for insurance companies, price prediction for market traders in stock and currencies, and within the meteorology sector for predicting weather patterns (Bishop, 1995) or risk zones. The ability of neural networks to replicate a reliable model of data rich with noise and anomalies, inferring a function created from its observations of the data, allows the prediction of data with a higher accuracy then that of a fixed model. (Gurney, 1997)

ANNs are made from collection of neurons, interconnected with weights. These collections of neurons are split into layers, with most neural networks containing at least one input and output layer. When a signal or value is presented to the input layer, all the neurons update their values based on the connected neurons and their corresponding weights. (Gurney, 1997) As a result of the neurons updating the output layer of the network will present a different value. The output value can then be used to train the neural network with methods listed in section 2.3.

Further reading into the structures of neural networks and how they function can be found in the bibliography of this document.

RNN (Recursive Neural Networks) are a type of neural network that relies on the use of recurrence to add a form of short term memory to the normal ANN network architecture. This is heavily used for time series prediction is commonly made up of an input, hidden, and output layer. First developed in the 1980’s, recurrent neural networks have been used to learn datasets that are presented in arbitrary sequences, making them ideal for learning and predicting data over a time series (Rojas, 1996). Unlike a normal ANN, RNNs also have at least one context layer that takes input from and provides input to the hidden layer. This recurrent connection causes previously presented inputs to the neural network to affect the values of the nodes when the new inputs are presented, providing a form of context for the neural network (Ulbrich, 1994). Recurrent neural networks have been found to combat tasks such as handwriting and speech pattern recognition where they perform exceptionally well compared to other neural network structures (Graves, et al., 2009).

Echo State Networks are a state of the art form of Recurrent Neural Network and a type of reservoir computing a relatively new direction in the research of neural networks. As a result there is a lot less information about them in the public domain. Much like the recurrent neural networks they have been known to show great results when predicting and reproducing time series data (Jaeger, 2004) . Echo state networks also commonly have three layers. These layers include both and input and output layer, however the centre layer differs from the hidden layer in an RNN. It is known as a reservoir and acts differently than the hidden layer found in RNNs. This reservoir layer is made of a collection of randomly connected nodes based on a connectivity level. The weights of these connections are also randomised using a specified range and unlike the RNN, will not change during the training of the network. The structure of the layer combined with the update during every update causes the layer to present a form of temporal memory commonly seen in the context banks of the RNN. (Jaeger, 2001)

Image of reservoir layer (Jaeger, 2007) 

Jaeger, H., 2007. Echo State Network. 2(9):2330 ed. s.l.:Scholarpedia,.

Jaeger, H., 2001. The “echo state” approach to analysing and training recurrent neural networks. s.l.:German National Research Center for Information Technology.

Jaeger, H., 2004. Harnessing nonlinearity: Predicting chaotic systems and saving energy in wireless communication. s.l.:s.n.

Ulbrich, C., 1994. Multi-Recurrent Netowrks for Traffic Forecasting. Vienna: Austrian Research Institute for Artificial Intelligence.

Rojas, R., 1996. Neural networks: a systematic introduction. s.l.:s.n.

Rojas, R., 1996. The Backpropagation Algorithm. s.l.:Springer-Verlag.

Gurney, K., 1997. An Introduction to Neural Networks. London: s.n.

Bar-Yam, Y., 2003. Dynamics of Complex Systems. s.l.:s.n.

Although this is more about Linux than the Raspberry Pi its self, i thought it would be a nice addition for beginners trying to learn the ropes with their Raspberry PI. Sometimes you need to monitor the CPU usage or the RAM usage on your Raspberry PI to determine how its running. With the standard Debian image(and most other images) comes an application called Top. To run it simply enter the following into the console followed by pressing enter :

top

Neural networks are a form of AI (Artificial Intelligence) that focuses on the ability to identify and classify patterns in data, much like the human brain. They are most commonly used within the financial sector of IT, as they offer a platform to predict changes, such as stock market prices at a stock exchange, or to classify data, such as the risk level of insurance company’s customers (Rojas, 1996). However they are found in many other locations within computing from being utilised as a function simplifier to weather prediction. The ability to identify and predict patterns in the information we perceive, allows humans to classify objects we have not previously encountered, or to predict something based on several factors, this is also true for neural networks.

Recently I have primarily been focusing my researching into Echo State Networks and Recurrent Neural Networks, subtypes of Neural Network. Recurrent Neural Networks are a type of computational model that are widely used to assist in solving problems with learning temporal dependencies (Data that is time dependent). This project will identify the strengths of these subtypes of Neural Networks for predicting temporal time dependencies. The key goal of the project is to identify the requirements of, and produce a working, feature rich Neural Network Framework focusing on recurrent Neural Networks and most specifically Echo State Networks. The engine produced by this project will be based in C#, in a heavily object orientated and configurable context to allow the quick adaptation and customisation of the network to attempt to solve a number of experiments.

Whilst the Raspberry PI has a fairly decent arm chip that the main processing is executed on, it is an old chip and its speed shows its age. Whilst overclocking isn't going to turn your Raspberry PI into a processing demon, it can certainly give you a little more power to work with. Overclocking the Raspberry PI is simple and requires an edit of a text file and a quick reboot. The file we are going to edit is found in the boot folder on the disk, so we switch to that directory with the following line : CD /BOOT The next step is to edit (or create if it doesn't exist) a file called config.txt. To do this we use a text editor called nano and execute the following command : sudo nano config.txt As always you will be likely asked to provide your password to allow the command to be executed with sudo. This command will open Nano with either a new file or the existing file if there was one present. The default speed of the arm processor in the Raspberry PI is 700MHz. By having the following line in the config.txt we are able to get the core to run at a higher 750MHZ. arm_freq=750

One of the key benefits of the linux os on the Raspberry PI is its ability to run with little or no user IO. allowing it to sit happily in hard to access locations, or remote locations, like the other side of a room which requires effort to reach. I personally travel a lot but like playing with my raspberry PI, but rather than taking it with me i leave it at home connected to the big wide web and use ssh to connect to it remotely. SSH comes installed by default in the recommended Debian kernel, however it doesn't launch automatically, and thus if the device resets , it requires you to walk the 10 steps across the room to re-enable it again. This pilgrimage however doesn't need to be made, A simple change to the start-up scripts and you can automate the launching of the SSH daemon : sudo mv /boot/boot_enable_ssh.rc /boot/boot.rc 

So you want to start developing some applications for your Raspberry PI but your knowledge of programming is limited to the modern .net languages such as vb .net and c#. Not a problem!! Welcome to the world of mono, a opensource cross platform and compatible version of the .net framework. By using a development IDE such as Visual studio or even better Mono Develop (http://monodevelop.com/) you can produce exe's that will run on your Raspberry PI. All you need to do is to install the mono run times on the Raspberry PI. To do that we enter the following lines

Hi guys, just thought i'd quickly give you a walk through on installing a small Apache2 with PHP web server and Mysql. This will require the installing of a few pieces of software and a limited configuration. Whether you connected directly to your R-PI or remotely this should be pretty straight forward.  Step One : make sure your apt-get software lists are up to date, this is a simple as running the following sudo apt-get upate