AI info by David Papkin

This David Papkin page has  info on AI (Artificial Intelligence).

Attention -the mechanism inside the transformer that does this comparing. For each word, it works out how much every other word should affect its meaning, and gives each one a weight.

Take the sentence “The bank approved the loan because it had strong collateral.” Attention is how the model works out that “it” means the loan and not the bank. It gives a strong weight to “loan” and “collateral” when interpreting “it.”

Bias testing means checking whether an AI system treats different groups of people unfairly or produces systematically skewed results.

For example, suppose an AI system is used to screen job applicants. Bias testing might compare its recommendations across gender, age, ethnicity, disability status, or other relevant groups to see whether one group is rejected much more often without a valid job-related reason.

It can involve checking:

• whether the training data under-represents certain groups,
• whether accuracy differs between groups,
• whether the model produces stereotyped or discriminatory outputs,
• and whether corrective measures reduce those differences.

Bias testing asks: “Does this AI work equally well and fairly for different groups of people?”

Deep learning is machine learning that uses multi-layer neural networks to automatically learn complex patterns from data. It is called “deep” because the neural network has many processing layers between the input and output. Input, hidden layers, output

For example, with image recognition:

• Early layers may detect edges and colors
• Middle layers may detect shapes and textures
• Deeper layers may recognize eyes, faces, cars, animals, etc.

Deep learning is used in things such as ChatGPT and other LLMs, speech recognition, image recognition, translation, autonomous driving, and medical image analysis.

Embedding – a model can convert text into an embedding, which is basically a long list of numbers representing its meaning.

Foundation model -is a large, general-purpose model trained once on a huge, broad dataset of text, code and sometimes images. It is then adapted to many different tasks instead of a new model being built for each one. Examples are GPT-4, Claude, Gemini and Llama.

The old approach was one model per job: one for fraud, one for document classification, one for chatbots, each trained from scratch. A foundation model reverses that. You build one general model and adapt it in one of three ways:

• Prompting: tell it what to do.
• RAG: give it your documents at the moment you ask.
• Fine-tuning: retrain it slightly on your own examples.

Grounded answer is an AI response supported by trusted source data or retrieved documents.

Inference is the stage where a trained AI model is actually used to make a prediction or generate an answer.

Neural Network – a machine-learning model inspired loosely by the way neurons in the brain process signals. It is made of connected units called artificial neurons, usually arranged in layers:

Input layer → Hidden layer(s) → Output layer

For example, for recognizing a cat:

• Input layer: receives image pixels
• Hidden layers: learn patterns such as edges, ears, eyes, fur
• Output layer: predicts “cat” or “not cat”

Each connection has a weight. During training, the network adjusts those weights to reduce errors and improve its predictions.

Goal – recognize Circle

Wrong prediction, needs to be trained

Neurons that learn patterns by adjusting weights during training.

And the relationship is:

AI → Machine Learning → Neural Networks → Deep Learning

Deep learning simply means using neural networks with many layers.

RAG (Retrieval-Augmented Generation) – an AI technique that retrieves relevant external information and gives it to a language model so it can produce more accurate, grounded answers instead of relying only on what the model learned during training.

parameter – a number inside the neural network that controls how strongly one thing influences another.

Transformer –architecture behind every modern llm(large language model) and its defining trait is reading an entire input at once rather than moving strictly left to right the way earlier language-processing methods did. 

 

Vector database (vector DB) is a database designed to store and search vectors—numerical representations of things like text, images, audio, or documents.

In AI, a model can convert text into an embedding, which is basically a long list of numbers representing its meaning.

For example:

“How do I reset my password?”

might become something like:

[0.12, -0.44, 0.81, 0.07, ...]

A vector database stores these embeddings and can quickly find other vectors that are similar in meaning, not just exact keyword matches.

For example, if your database contains:

• “Change your account password”
• “Reset forgotten login credentials”
• “Configure a Cisco router”

and you search for “I forgot my password”, the vector DB will likely retrieve the first two because they are semantically similar.

This is especially important in RAG — Retrieval-Augmented Generation:

Documents → embeddings → vector database → similarity search → relevant chunks → LLM → answer

So if you upload 1,000 company manuals, the AI does not normally send all 1,000 manuals to the LLM. It searches the vector DB for the most relevant sections and sends only those sections to the model.

Common vector databases include Pinecone, Milvus, Weaviate, Qdrant, Chroma, and vector-search features in databases such as PostgreSQL with pgvector and cloud AI search services.

Update cadence -AI models can change over time, so organisations need to know how often the model is updated, whether behaviour may change after an update, and how customers are notified.

AI links

Attention is all you need – is a 2017 research paper on machine learning authored by eight scientists and engineers working at Google. The paper introduced a new deep learning architecture known as the transformer, based on the attention mechanism proposed in 2014 by Bahdanau et al.^[2] 

AI videos

 

End of David Mark Papkin AI page.

http://davidpapkin.org/

David Papkin favorite movies

Robert Deniro in GoodFellas