Unit 18 Artificial Intelligence (AI)

A summary of the key specification points (not those relating to graphs) written by ChatGPT-3 can be found here.

Specification

• Show understanding of how graphs can be used to aid Artificial Intelligence (AI)

  • Purpose and structure of a graph

  • Use A* and Dijkstra’s algorithms to perform searches on a graph

    • Candidates will not be required to write algorithms to set up, access, or perform searches on graphs

• Show understanding of how artificial neural networks have helped with machine learning

• Show understanding of Deep Learning, Machine Learning and Reinforcement Learning and the reasons for using these methods.

  • Understand machine learning categories, including supervised learning, unsupervised learning

• Show understanding of back propagation of errors and regression methods in machine learning

If the video does not play above, follow this hyperlink: https://www.youtube.com/watch?v=1fkV5rB13jQ 

Useful Websites

• A* - Wikipedia

• Brilliant - A*

You can find the VB.NET (Python will be done at some point) base program on my GitHub page: GitCoder001 here.  You can then clone this to begin building your A* solving algorithm.  The program is fully documented in the comments.

A* Videos

General Graphing Videos

Here is a video that (after a couple of minutes) looks at neural networks in a very clear way

Machine Learning is a subset of Artificial Intelligence that uses statistical learning algorithms to build systems that have the ability to automatically learn and improve from experiences without being explicitly programmed.

Most of us use machine learning in our day-to-day lives when we use services like recommendation systems on Netflix, YouTube, Spotify; search engines like Google; voice assistants like Google home and Amazon Alexa. In Machine Learning we train the algorithm by providing it with a lot of data and allowing it to learn more about the processed information. 

Machine learning fuels all sorts of automated tasks that span across multiple industries, from data security firms that hunt down malware to finance professionals who want alerts for favourable trades. The AI algorithms are programmed to constantly be learning in a way that simulates as a virtual personal assistant—something that they do quite well.

Machine learning involves a lot of complex math and programming that, at the end of the day, serves a mechanical function the same way a flashlight, a car, or a computer screen does. When we say something is capable of “machine learning”, it means it’s something that performs a function with the data given to it and gets progressively better over time. It's like if you had a flashlight that turned on whenever you said “it's dark,” so it would recognize different phrases containing the word "dark."

ML algorithms (Deep Learning, Linear Regression, Clustering, etc) can be broadly classified into four categories: Supervised, Unsupervised, Reinforcement learning and Semi-Supervised learning.

Supervised Learning

In supervised learning we have input variables (x) and an output variable (Y) and we use an algorithm to learn the mapping from input to output. In other words, a supervised learning algorithm takes a known set of input dataset and its known responses to the data (output) to learn the regression/classification model. A learning algorithm then trains a model to generate a prediction for the response to new data or the test datasets.

Supervised learning requires labelled data, that is data that has been given to it within a known category. E.g., pictures of cats or dogs, etc.  Data labelling is the process of identifying raw data (images, text files, videos, etc.) and adding one or more meaningful and informative labels to provide context so that a machine learning model can learn from it.  For example, labels might indicate whether a photo contains a bird or car, which words were uttered in an audio recording, or if an x-ray contains a tumour. 

Today, most practical machine learning models utilise supervised learning, which applies an algorithm to map one input to one output. For supervised learning to work, you need a labelled set of data that the model can learn from to make correct decisions. Data labelling typically starts by asking humans to make judgments about a given piece of unlabelled data. For example, labellers may be asked to tag all the images in a dataset where “does the photo contain a bird” is true. The tagging can be as rough as a simple yes/no or as granular as identifying the specific pixels in the image associated with the bird. The machine learning model uses human-provided labels to learn the underlying patterns in a process called "model training." The result is a trained model that can be used to make predictions on new data.

In machine learning, a properly labelled dataset that you use as the objective standard to train and assess a given model is often called “ground truth.” The accuracy of your trained model will depend on the accuracy of your ground truth, so spending the time and resources to ensure highly accurate data labelling is essential.

Unsupervised Learning

Unsupervised Learning is used when we do not have labelled data. Its main focus is to learn more about the data by inferring patterns in the dataset without reference to the known outputs. It is called unsupervised because the algorithms are left on their own to group the unsorted information by finding similarities, differences and patterns in the data. Unsupervised learning is mostly performed as a part of exploratory data analysis. It is most commonly used to find clusters of data and for dimensionality reduction.  This type of ML is useful for Big Data applications, where extremely large sets of structured and unstructured data that cannot be handled with traditional methods. Big data analytics can make sense of the data by uncovering trends and patterns. Machine learning can accelerate this process with the help of decision-making algorithms. It can categorize the incoming data, recognize patterns and translate the data into insights helpful for business operations.

Reinforcement Learning

In simple terms, reinforcement learning can be explained as learning by continuously interacting with the environment. It is a type of machine learning algorithm in which an agent learns from an interactive environment in a trial and error way by continuously using feedback from its previous actions and experiences. The reinforcement learning uses rewards and punishments, the agents receive rewards for performing correct actions and penalties for doing it incorrectly. Just as with humans, the algorithms develop traits that favour actions leading to rewards and steer away from those which may have penalties. 

Reinforcement learning differs from the supervised learning in that in supervised learning the training data has the answer key with it so the model is trained with the correct answer itself whereas in reinforcement learning, there is no answer but the reinforcement agent decides what to do to perform the given task. In the absence of a training dataset, it is bound to learn from its experience.

Semi-Supervised (Active) Learning

Semi-supervised learning is an approach to machine learning that combines a small amount of labelled data with a large amount of unlabelled data during training. Semi-supervised learning falls between unsupervised learning (with no labelled training data) and supervised learning (with only labelled training data). It is a special instance of weak supervision.

Unlabelled data, when used in conjunction with a small amount of labelled data, can produce considerable improvement in learning accuracy. The acquisition of labelled data for a learning problem often requires a skilled human agent (e.g. to transcribe an audio segment) or a physical experiment (e.g. determining the 3D structure of a protein or determining whether there is oil at a particular location). The cost associated with the labelling process thus may render large, fully labelled training sets infeasible, whereas acquisition of unlabelled data is relatively inexpensive. In such situations, semi-supervised learning can be of great practical value. Semi-supervised learning is also of theoretical interest in machine learning and as a model for human learning.

What is Deep Learning? Downsides?

Deep learning is a machine learning technique that is inspired by the way a human brain filters information, it is basically learning from examples. It helps a computer model to filter the input data through layers to predict and classify information. Since deep learning processes information in a similar manner as a human brain does, it is mostly used in applications that people generally do. It is the key technology behind driver-less cars, that enables them to recognize a stop sign and to distinguish between a pedestrian and lamp post. Most of the deep learning methods use neural network architectures, so they are often referred to as deep neural networks.

ML refers to an AI system that can self-learn based on the algorithm. Systems that get smarter and smarter over time without human intervention is ML. Deep Learning (DL) is a machine learning (ML) applied to large data sets. Most AI work involves ML because intelligent behaviour requires considerable knowledge

There are lots of confusion and overblown expectations about deep learning and in many cases relatively simpler models more suitable than their deep learning counterparts. Deep learning takes the concept of adaptive AI behaviour a step further from machine learning using unstructured data available. It deals with such concepts as deep neural networks, machine translation, bioinformatics and many more.  

The problems most, especially small, businesses are facing do not really require such complex and sophisticated methods. Deep learning is also inherently slow. For example, it can take several weeks of just training using multiple GPUs.  Deep learning also demands huge sample sizes to train on. Deep neural networks often take hundreds of thousands or even more samples to achieve high performance. Many problems require so-called labelled datasets (each sample is annotated with an expected value). Such labelling is time-consuming and may often require to do it manually.  Deep neural networks are often considered black boxes whose inner operations are not really interpretable. But interpretability is important because it can give new insights into relationships between numerous variables and expected outcomes. It also increases the trust of the people who use it. This ability to explain solutions is inherent to many simpler methods, in particular, linear ones, where the direct relationship between parameters can be analysed.

Deep Learning is basically mimicking the human brain; it can also be defined as a multi-neural network architecture containing a large number of parameters and layers. The three fundamental network architectures are as listed below.

Artificial Neural Networks (ANNs) can be classified in multiple ways. THIS IS BEYOND THE SCOPE OF THE EXAM!

1. Perceptron: It is the simplest type of ANN, which is used for binary classification tasks. It receives weighted inputs, applies an activation function, and produces an output. It can effectively implement logic gates like AND, OR, or NAND.  

2. Feedforward Neural Network (FFNN): A Feedforward Neural Network is a basic neural network with layers of neurons. Data flows from input to output in one direction, making it ideal for tasks like classification, face recognition, and speech recognition.  

3. Multilayer Perceptron (MLP): The Multilayer Perceptron (MLP) is a deeper version of FFNN, consisting of multiple hidden layers between input and output. It can handle more complex tasks such as speech recognition, machine translation, and classification problems.  

4. Convolutional Neural Network (CNN): Convolutional Neural Networks specialize in tasks related to image processing. With convolutional layers that learn to detect features like edges, CNNs are ideal for image and video recognition.  

5. Radial Basis Function Network (RBFN): An RBF Network uses radial basis functions as activation functions to classify data based on its similarity to a center prototype. It is particularly useful for classification tasks in pattern recognition.  

6. Recurrent Neural Network (RNN): Recurrent Neural Networks are designed to handle sequential data, making them perfect for tasks like speech recognition and time-series forecasting. They retain memory of previous inputs, making them ideal for sequential predictions.  

7. Long Short-Term Memory (LSTM): An LSTM is a type of RNN that addresses the long-term dependency problem by using memory cells. This architecture allows the network to remember and use information over longer periods.  

8. Generative Adversarial Network (GAN): Generative Adversarial Networks consist of two networks: a generator and a discriminator.
   The generator creates synthetic data, while the discriminator evaluates it. This setup allows GANs to generate high-quality data, such as images.  

Convolutional Neural Networks (CNN)

Convolutional Neural Network (CNN) is basically an artificial neural network that is most widely used in the field of Computer Vision for analysing and classifying images. It is a deep learning algorithm that takes the input image and assigns weights/biases to various aspects or objects in the image, so that it can differentiate one from the other. The hidden layers of a CNN typically consist of convolutional layers, pooling layers, fully connected layers, and normalization layers. The architecture of a ConvNet is analogous to that of the connectivity pattern of Neurons in the Human Brain and was inspired by the organization of the Visual Cortex.

Recurrent Neural Networks

Recurrent Neural Networks is a type of neural network architecture that is used in sequence prediction problems and is heavily used in the field of Natural Language Processing. RNNs are called recurrent because they perform the same task for every element of a sequence, with the output being dependent on the previous computations. Another way to think about RNNs is that they have a “memory” which captures information about what has been calculated so far.

Recursive Neural Networks

“A recursive neural network is a kind of deep neural network created by applying the same set of weights recursively over a structured input, to produce a structured prediction over variable-size input structures, or a scalar prediction on it, by traversing a given structure in topological order.” 

A recursive neural network is more like a hierarchical network where there is really no time aspect to the input sequence but the input has to be processed hierarchically in a tree fashion. Here is an example of how a recursive neural network looks. It shows the way to learn a parse tree of a sentence by recursively taking the output of the operation performed on a smaller chunk of the text.

How our brains work

Again, this is here for depth and breadth and the important things to know for the exam is an understanding of what an artificial neural network is, how it works and how it learns.

The human brain is a complex network containing around 90 billion cells called neurons that communicate with each other via connections called synapses. The brain communicates by sending signals that travel through the neuron to the synapses. When the message reaches the end of a neuron, called the synapse, it is converted into a chemical signal called a neurotransmitter. The neurotransmitter that travels across the synaptic gap to the neuron on the other side, which turns the neurotransmitter back into a signal and sends it down the line.

The number of these synaptic connections is mind-boggling, and researchers estimate that neurons interconnect at a hundred trillion to one thousand trillion synapses. Neurons have specialized appendages called dendrites and axons. Dendrites bring information to the center of the neuron and axons take information away from the center. The neuron collects multiple incoming signals through several extremely long arms called dendritic trees that form branches.

Every thought, experience, physical sensation and feeling triggers thousands of neurons, which form a neural network. When we repeat an experience, the brain learns to trigger the same neurons each time, by strengthening the synapses, according to Hebb’s axiom.

Hebb’s Axiom

In 1949 Canadian neuropsychologist Donald Hebb suggested that learning occurs by strengthening the synapses, with neurons functioning as merely the computational elements. Dr. Hebb’s axiom has remained the widely held assumption in the field of neuroscience and gave rise to what is known as Hebb’s axiom,

“neurons that fire together wire together.”

A phrase that Donald Hebb coined in 1949 and had remained a central tenet of neuroscience.

Wiring neurons together to create a neural network is beneficial in that it helps us to store and recall information – and thereby learn – in an efficient way. For example, a neural network helps you to remember the name of a new acquaintance by creating connections. However, the neural network can be stubborn when we try to rewire neurons to respond to a familiar situation in a different way, giving rise to expressions like ‘you can’t teach an old dog new tricks.’. Further research has pointed to learning developing more in the synapses but this is irrelevant for this topic.

What is an Artificial Neural Network?

We talked above about different types of neural network.  Artificial neural networks are key to deep learning (supervised learning) and mimic the structure of the human brain.  The extract below is from an excellent article published by Bernard Marr and you can read the full article here.

Artificial neural networks (ANN) give machines the ability to process data similar to the human brain and make decisions or take actions based on the data. While there’s still more to develop before machines have similar imaginations and reasoning power as humans, ANNs help machines complete and learn from the tasks they perform.

What else can artificial neural networks do?

Artificial neural networks are a main component of machine learning and they are designed to spot patterns in data. This makes ANNs an optimal solution for classifying (sorting data into predetermined categories), clustering (finding like characteristics among data and pulling that data together into categories) and making predictions from data (such as helping determine infection rates for COVID, the next catastrophic weather event or box-office smash). In everyday life, ANNs are powering the “watch next” feature of YouTube videos, creating realistic CGI faces, helping detect fraud, giving us the ability to chat with chatbots and more. In fact, there are probably not many tasks an artificial neural network can’t do as long as it’s trained to do it.

How do artificial neural networks work?

Ultimately, ANNs try to replicate how our human brains process information and make decisions. While ANNs are based on mathematical theory created in the 1940s, it wasn’t until the last couple of decades that it became a focus for artificial intelligence. When back propagation was developed to help these networks learn and adjust actions based on outcomes its development and adoption really began to accelerate.

When a human brain receives an input, it processes it through a series of neurons. Different neurons of the human brain are responsible for processing different aspects of input in a hierarchical fashion. ANNs try to replicate this through artificial neurons called units that are arranged in layers and connected to each other to create a web-like structure.

ANNs have an input layer and output layer. Between these two layers there are other hidden layers that perform the mathematical computations that help determine the decision or action the machine should take. Ultimately, these hidden layers are in place to transform the input data into something the output unit can use.

The data is processed by each hidden layer and then moves on to the next based on connections that are weighted. Think of this process as an assembly line in a factory—raw materials as the input and different stops on the conveyor belt to add an element to the product equate to the hidden layers of an ANN that processes the data until you get to the output. Based on what the machine learns about the data when processed by one layer, it determines how to move it through to the next, more senior layer based on the value it receives when evaluated. Based on the complexity of the issue at hand, it can continue to process through more senior units until delivered to the output layer.

Before an ANN can be fully deployed, it must be trained. This training involves comparing an outcome a machine gets with the human-provided description of what outcome is expected. If these don’t match, the machine uses this feedback and goes back to adjust the weights of the layers (called backpropagation). These new learning rules are applied and help guide the neural networks on future processing.

To illustrate how this works for the human brain, consider how humans might learn how to shoot a basketball so they score more baskets. Over time and with experience, different techniques are tried to improve the odds the shot will make it in the basket—bending legs less or more, adjusting the hand position, shooting force, the angle of the shot, use of backboard, etc. When a shot doesn’t make it in, the brain adjusts based on this feedback and tries something else. Over time, there is enough learning to improve the outcome so that more balls make it through the net than get rejected.

Types of artificial neural networks

There are several types of artificial neural networks including the feedforward neural network, recurrent neural network and a variety of others. The network you use is based on the data set you have to train it with as well as the task you want to accomplish.

A feedforward neural network, the most basic type of neural network, can only process data from input to output in one direction. This is what is used for supervised machine learning when you already know what outcome you want the network to achieve. It’s the basis for many commercial applications such as machine vision. A recurrent neural network has data flow in multiple directions and is widely used for more complex tasks. Use cases for recurrent neural networks include document generation and real-time language translation.

Future of ANNs

While ANNs can tackle most tasks if they are allowed to train for it, the biggest obstacle to overcome is the amount of time it takes to train ANNs and the computing power required for a complex task. In addition, it’s impossible for humans to fully understand what happens in the hidden layers of an artificial neural network. Although researchers are actively working on this, there is still a lot to learn even though we’ve come so far in helping machines think and act like a human.

Summary

Hidden layers are where the processing is done. The front and end take the input and final outputs, but the processing is carried out on the layers. For very complex problems, more layers are required. More layers also improves the accuracy of the result and enables the neural network to learn and make decisions on its own.

How does an Artificial Neural Network operate?

Neurons are the fundamental units of the brain which contains Dendrites, Axons, Synapses, etc. Dendrites act as a receiver whereas axon acts as a transmitter of signals to and from other neurons. In case of a human brain, the input comes from our senses like ear, nose, eyes etc., and gets processed by our brain.

In an artificial neural network (ANN), inputs (I1, I2, I3..In) are independent variables whereas the output(s) are dependent variables. The synapses are the weights assigned to each input neurons.  An independent variable is one that can represent any value and do not have to be related to other inputs.  A dependent variable is one whose value is calculated based on the inputs and a given function.

By changing the weights, neural networks learn which signal is important and these are the things that get adjusted during the process of learning.

Step 1: Once the input neurons(independent variables) are chosen its weights are assigned.

Step 2: Now the activation function is applied to the summation of inputs and weights . Depending on the activation function, neurons learn which signal needs to be passed. 

Step 3: Output (y) is generated and the cost function is calculated for the actual and predicted output. Based on the feedback from this cost function, all the weights get reassigned and again the whole process repeats until we get better feedback. This whole process is known as model training.

Note: There may be many layers between input and output neurons and these intermediate layers are known as hidden layers. Neural networks depend on the number of hidden layers, so be very calculative and selective while selecting hidden layers.

Deeper Look at an ANN Through Example

The following section goes into detail surrounding the function and mathematics of an artificial neural network.  This is beyond what you need to know for the exam but does help to clarify how ANNs operate.

Key Points related to the architecture:

1. The network architecture has an input layer, hidden layer (there can be more than one) and the output layer. It is also called MLP (Multi Layer Perceptron) because of the multiple layers.

2. The hidden layer can be seen as a “distillation layer” that distils some of the important patterns from the inputs and passes it onto the next layer to see. It makes the network faster and efficient by identifying only the important information from the inputs leaving out the redundant information

3. The activation function serves two notable purposes:

   • It captures non-linear relationship between the inputs

   • It helps convert the input into a more useful output.

   • In the above example, the activation function used is sigmoid:

O1 = 1 / (1+exp(-F))

Where F = W1*X1 + W2*X2 + W3*X3

Sigmoid activation function creates an output with values between 0 and 1. There can be other activation functions like Tanh,   softmax and RELU.

4. Similarly, the hidden layer leads to the final prediction at the output layer:

O3 = 1 / (1+exp(-F 1))

Where F 1= W7*H1 + W8*H2

Here, the output value (O3) is between 0 and 1. A value closer to 1 (e.g. 0.75) indicates that there is a higher indication of customer defaulting.

5. The weights W are the importance associated with the inputs. If W1 is 0.56 and W2 is 0.92, then there is higher importance attached to X2: Debt Ratio than X1: Age, in predicting H1.

6. The above network architecture is called “feed-forward network”, as you can see that input signals are flowing in only one direction (from inputs to outputs). We can also create “feedback networks where signals flow in both directions.

7. A good model with high accuracy gives predictions that are very close to the actual values. So, in the table above, Column X values should be very close to Column W values. The error in prediction is the difference between column W and column X:

8. The key to get a good model with accurate predictions is to find “optimal values of W — weights” that minimizes the prediction error. This is achieved by backpropagation and this makes ANN a learning algorithm because by learning from the errors, the model is improved.

9. The most common method of optimisation algorithm is called gradient descent, where, iteratively different values of W are used and prediction errors assessed. So, to get the optimal W, the values of W are changed in small amounts and the impact on prediction errors assessed. Finally, those values of W are chosen as optimal, where with further changes in W, errors are not reducing further. 

Backpropagation

Most deep neural networks are feed-forward, meaning they flow in one direction only from input to output. However, you can also train your model through backpropagation; that is, move in opposite direction from output to input. Backpropagation allows us to calculate and attribute the error associated with each neuron, allowing us to adjust and fit the algorithm appropriately.

The Watson book and PowerPoint slides (L8) go into better depth on this, but the overall aim of backpropagation is to reduce the error rate.  That is, we compare the outputs of the model against the expected output.  We then use this knowledge to tweak the network's weightings to reduce the errors to an acceptable limit.  In complex systems, it is not always possible to reduce this to zero.

Derivatives (calculus) are used to determine the appropriate weights for each neuron.  The 'back' part is in reference to feeding backwards along the path from outputs to improve the model.  Backpropagation, is essentially the chain rule of calculus applied to computational graphs.

There are two types of backpropagation networks.

• Static backpropagation

• Recurrent backpropagation

Static backpropagation

In this network, mapping of a static input generates static output. Static classification problems like optical character recognition will be a suitable domain for static backpropagation.

Recurrent backpropagation

Recurrent backpropagation is conducted until a certain threshold is met.  After the threshold, the error is calculated and propagated backward.

The difference between these two approaches is that static backpropagation is as fast as the mapping is static. 

In conclusion, Neural network is a collection of connected units with input and output mechanism, each of the connections has an associated weight. Backpropagation is the “backward propagation of errors” and is useful to train neural networks. It is fast, easy to implement and simple. Backpropagation is very beneficial for deep neural networks working over error prone projects like speech or image recognition.

Regression

Some prediction problems require predicting both numeric values and a class label for the same input.

A simple approach is to develop both regression and classification predictive models on the same data and use the models sequentially.

An alternative and often more effective approach is to develop a single neural network model that can predict both a numeric and class label value from the same input. This is called a multi-output model and can be relatively easy to develop and evaluate using modern deep learning libraries such as Keras and TensorFlow.

It is common to develop a deep learning neural network model for a regression or classification problem, but on some predictive modelling tasks, we may want to develop a single model that can make both regression and classification predictions.

• Regression refers to predictive modelling problems that involve predicting a numeric value given an input (e.g. selecting 

• Classification refers to predictive modelling problems that involve predicting a class label or probability of class labels for a given input.

Difference Between Classification and Regression in Machine Learning

There may be some problems where we want to predict both a numerical value and a classification value.

One approach to solving this problem is to develop a separate model for each prediction that is required.

The problem with this approach is that the predictions made by the separate models may diverge.

An alternate approach that can be used when using neural network models is to develop a single model capable of making separate predictions for a numeric and class output for the same input.

Regression is method dealing with linear dependencies, neural networks can deal with nonlinearities. So if your data will have some nonlinear dependencies, neural networks should perform better than regression.

Further information

The following web links give more information on the topics covered in this section and often go far beyond the specification

• Machine Learning Glossary (Google)

• ML classifications (supervised, unsupervised, semi-supervised and reinforcement) (Towards Data Science)

• Supervised Learning (Tech Target)

• Reinforcement learning (Geeks for Geeks) 

• Machine Learning types

• How does a neural network work? (Towards Data Science)

• ML Vs Deep Learning vs Neural Networks (IBM)

• Backpropagation

• AI, ML and DL - How not to get them confused (Towards Data Science)

• How our brains think (Longevity Facts)

• How do memories work

• Backpropagation and logistical regression in Python (advanced)

• What alternatives exist to ANNs? (Quora)

Large Language Models (ChatGPT, PALM, CALM, LaMDA, etc)

Large language models are a type of artificial intelligence that are trained to generate human-like text. They are trained on vast amounts of data, such as books, articles, and websites, in order to learn the patterns and structures of natural language.

Once trained, the model can then generate new text that is similar to the text it was trained on. For example, if a large language model was trained on a dataset of news articles, it might be able to generate a new article on a similar topic that reads like it was written by a human.

To generate text, the model processes an input prompt and predicts the next word in the sequence based on the words that came before it. The model uses this process to generate a sequence of words that form a coherent piece of text.

Large language models are able to generate human-like text because they have learned the patterns and structures of language from a large amount of data. They can also understand and respond to context and generate text that is appropriate for the given situation.

Videos