Artificial Neural Networking
- 1. Date : 14ᵗʰ December 2023 Kathmandu Model Secondary School Artificial Neural Networking By: Rij Amatya Pragyan Shrestha Kritika Silwal
- 2. Branch of Machine learning models Was proposed in 1944 Information processing model A computational model inspired by the human brain's structure and function. Follows principles of neuronal organization Makes prediction based on what they’ve learned Applied in the concept of AI Learns by examples Artificial Neural Networking(ANN) INTRODUCTION
- 3. Artificial Neural Networks Core of deep learning Image recognition, NLP, Robotics, etc. 3 different layers Input Layer 1. Hidden Layer 2. Output layer 3. Sometimes referred to as the MLP (Multi-Layer Perceptron).
- 4. The amount of data produced has increased, and big data has become a buzzword. Big data made it easy to train ANNs. While classical machine learning algorithms fell short of analyzing big data, artificial neural networks performed well on big data.
- 5. Biological Neural Networks A neural network consists of the cell body containing the nucleus, many-branched dendrites, and a long axon. As you can see in the image above, the length of the axon is much longer than the cell body. The axons are divided into many branches that combine with the dendrites or cell bodies of other neurons. Biological neurons generate electrical signals that travel along axons. If a neuron gets enough stimulation, it fires. In general, biological neurons work this way. This work may seem simple, but billions of neural networks can be managed by connecting neurons.
- 6. Artificial Neurons – Associated with the weights which contain information about the input signal Compromises interconnected nodes (neurons) organized in layers Interconnected to each other by weighted links
- 7. As you can see, the inputs and output are numbers and each input has a weight in this architecture. These weighted inputs are summed first, and then a bias is added. This sum is passed through a step function. This function can be, for example, a sign function. Perceptron A perceptron is the smallest unit of a neural network. This architecture was developed by Frank Rosenblatt in 1957. This is a simple model of biological neuron in Artificial Neural Network
- 8. An algorithm for binary classification can be used. A simple threshold logic unit (TLU) is a single linear function. This simple approach has paved the way for developing AI tools like ChatGPT. The image shows a perceptron with two inputs and three outputs, connected via a dense layer. It is used for multi- label classification.
- 9. How to train a perception? First, assign a random weight to each input. The sum of the weighted inputs is found, and then a bias is added to this sum. Note that, when one neuron triggers another neuron frequently, the connection between them becomes stronger. Inputs passing through neurons produce an output. This output is a prediction. The actual value is compared with the prediction and the error is calculated. Weights are updated to make predictions with fewer errors. Perceptron was a nice approach, but failed to solve some simple problems like XOR. To overcome the limitations of the perceptrons, the multilayer perceptron was developed. Let’s dive into multilayer perceptrons. If you learn how a perceptron is trained, you will have a better understanding of how ANNs work. Let’s now discuss how to train a perceptron.
- 10. Multi-layer Perceptron Multilayer perceptron consist of an input layer, a hidden layer, and an output layer. As you can see in the image above, there is a hidden layer. If there is more than one hidden layer, it is called a deep neural network. This is where deep learning comes into play. Deep learning became popular with the development of modern AI architectures.
- 11. In short, the inputs go through neural networks and a prediction is made. But, how to improve the prediction of a neural network? This is where the backpropagation algorithm comes in. This algorithm takes a neural network’s output error and then propagates that error backward through the network. So, the weights are updated. And this cycle continues until the best prediction is obtained.
- 12. A P P L I C A T I O N ANNs power language translation, sentiment analysis, and chatbots in NLP applications. Natural Language Processing (NLP): ANNs analyze medical images for disease detection, assisting in diagnostics using techniques like computer- aided diagnosis Medical Diagnosis: ANNs predict stock prices, market trends, and assess financial risk by analyzing historical data. Financial Forecasting: ANNs contribute to the development of self-driving cars by enabling object detection, lane keeping, and decision- making. Autonomous Vehicles: ANNs are used for image and speech recognition, enabling applications like facial recognition and voice assistants. Image and Speech Recognition: 12
- 13. ANNs often require large labeled datasets for training, and obtaining such data can be challenging. Training complex neural networks may demand significant computational power. Neural networks can overfit to training data, capturing noise and hindering generalization to new data. Issues related to bias, fairness, and accountability arise, especially in sensitive applications or high- stakes decision-making. CHALLENGES ANNs require large amounts of labeled data for effective training, and performance may suffer with insufficient or biased datasets. Training deep networks can be computationally demanding Models memorize training data noise rather than learning general patterns. Neural networks are often perceived as black- box models, making it challenging to interpret their decision-making processes. LIMITATIONS C H A L L E N G E S A N D L I M I T A T I O N S
- 14. 14