IB Digital Society 3.6C Uses of Artificial Neural Networks Notes

3.6C Uses of Artificial Neural Networks

What an artificial neural network is

1. An artificial neural network is the structure behind deep learning. Loosely inspired by the brain, it is made of simple units called neurons arranged in connected layers that pass signals forward.
2. It has three kinds of layer: an input layer that receives data, one or more hidden layers that process, and an output layer that gives the answer. Information spreads across many small connections, not one rule.
3. This lets networks handle messy, high-dimensional inputs (images, sound, language) no human could hand-code. The cost is interpretability. Intelligence spread across millions of connections is hard to trace.

Inside the network: neurons and weights

1. A single neuron takes several numbers as inputs, multiplies each by a weight, adds them, and passes on a signal only if the total is high enough. The weight decides how much each input matters.
2. Weights are the whole story of what a network knows: a large weight strongly influences the next neuron, a small one barely matters. A network can hold millions of weights that together encode the patterns it picked up.
3. No human sets the weights; they start random and adjust automatically during training. The knowledge is a grid of numbers, not readable instructions, which is a major reason networks are a black box: even builders cannot say exactly why it decided as it did.

How a neural network learns

1. A network learns by practising: it makes a guess, compares it with the correct answer to measure the error, nudges its weights to shrink that error, and repeats millions of times.
   1. This process is called training and it is expensive: reaching high accuracy can take millions of examples on banks of specialised chips for days or weeks, so only a handful of well-resourced organisations can train the largest networks.
2. Because learning comes entirely from the training data, the network absorbs its gaps and biases. Whoever assembles the data effectively decides who it serves well; a group missing from the examples is one the network never learns to handle.

Neural networks for images

1. In computer vision the network learns a hierarchy of features: early layers detect edges, middle layers combine them into shapes, deeper layers recognise whole objects such as a face or a tumour.
   1. This layered feature-building powers image recognition: photo tagging, self-driving car perception, factory quality checks, and medical imaging that can flag cancer earlier than a rushed human eye, all learned from labelled images.
2. Risks land unequally. A vision system trained on few dark-skinned faces or few night-time scenes tends to fail exactly where the cost is highest, and tiny deliberate changes can fool it. Whether it is safe to rely on such a system depends on more than its accuracy score; it also depends on who bears the cost when it fails.

Neural networks for language and sound

1. In speech recognition a network turns a voice waveform into text; in natural language processing networks translate, power chatbots and search, and predict the next words in a sentence.
   1. A voice assistant shows the chain: a speech network transcribes your request, a language model finds a likely response, another network speaks it back, in about a second. The same NLP powers live translation.
2. A network predicts likely words from patterns, so fluency is not understanding. It can be confidently wrong, and it mishears the voices and languages least present in its data. These tools work best in widely spoken languages and standard accents, so speakers of less common languages and non-standard accents are served worst.

Uses, strengths, and the black-box problem

1. Across images, language, sound, and games, networks are extraordinarily capable, often matching or beating human experts on narrow tasks, which is why they are deployed in high-stakes settings from hospitals to courts to hiring.
2. The same feature that makes a network powerful also makes it hard to trust. A decision spread across millions of weights turns it into a black box: accurate, yet unexplainable even to its developers. When it denies a loan or flags a disease, the affected person may get no real explanation.
3. The debate goes beyond the technical. Accuracy, accountability, and the right to an explanation can pull in different directions, and different communities carry the risks and benefits unequally.