Generative AI Langchain Chapter 1 by Ben Auffarth
What two things created the major breakthroughs in performance between GPT-3 and GPT-4?
1. Parameters (4 has over a trillion) 2. post-training fine-tuning with human instruction
Dimensionality reduction with no reconstruction loss often comes with a price. What are those prices?
1. lack of regularity (no exploitable or intepretable structures in the latent space) 2. When reducing it is also important to reduce the number of dimensions and keep the information LARGELY the same. (accuracy for speed)
What is significant about representation learning?
A model learns the internal representations of raw data in order to perform a task rather than engineered feature extraction.
What do transfer learning techniques allow?
A pre-trained model trained on one task to be fine-tuned on another similar task.
What is a Variational Autoencoder?
A process to match inputs and expected outputs better
What is a language model?
A statistical model used to predict words in a sequence of natural language. Some models may use deep learning with massive sets of data, making them large language models, or LLMs.
What is question answering in the context of language models?
AI Chatbots and virtual assistants can provide personalized and efficient help reducing response times
What is a "token"?
AN instance of a sequence of characters, typically forming a world, punctuation mark, or number. They're base elements of constructing sequences of text.
In transformer architecture, what is multi-head attention?
Attention is focused multiple times in parallel improving the model's ability to focus on different types of information.
At what number of parameters do emergent capabilities come about?
Between 2 and 7 billion parameters.
What is the third challenge that impacts business operations use of AI?
Bias in Data
What is a common example of a tokenizer?
Byte-pair encoding
What is the second essential concept for regularity within a latent space?
Completeness (a sample from the latent space should have meaning when decoded)
What is the second challenge that impacts business operations use of AI?
Compute Requirements
What is automatic summarization in the context of language models?
Concise summaries of articles, research papers, and more. This enables users to consume and understand information more rapidly.
What is the first essential concept for regularity within a latent space?
Continuity (two close points in. the space should be similar)
What is the first challenge that impacts business operations use of AI?
Data availability
What are some reasons that one might want to utilize dimensionality reduction?
Data visualization, data storage, heavy computation
What breakthrough enabled LMs such as BERT or GPT?
Deep neural network architecture
What tool, made by NVIDIA, created the "Big Bang of DL"?
Direct to GPU programming utilizing CUDA
What is the fourth challenge that impacts business operations use of AI?
Evaluation Difficulties
What concept illuminates the creative potential of LLMs, but on the flipside also illustrates the potential for fabricated results?
Hallucinations
Summarize the backpropagation algorithm made by Hinton, and Williams in 1980s
Hinton and Williams created a method that allows arbitrarily connected neural networks to develop internal structures that are appropriate for a particular task domain.
What is meant by Artificial Intelligence (AI)?
It is a broad field of study within computer science focused on developing agents who can reason, learn, and act autonomously.
In Transformer architecture, what is positional encoding?
It is essentially a form of meta data tied to the input embedding to allow the transformer to remember the order of words (as they process all words simultaneously meaning order is lost.
T/F why, is it fair to say that a decoder could be interpreted as uncompressing data?
It is fair to say so, so true, but this can sometimes be lossy meaning that some data is not recovered from the decompression.
What does a tokenizer do?
It splits on whitespace and punctuation to break text into individual tokens.
What is the value of Neural Machine Translation (NMT) in machine translation?
It uses DL to capture long-range dependencies in a sentence.
What is a generative model?
It's a type of ML model which can make new data based on patterns from existing input data.
What is topic modeling in the context of language models?
LLMs can find abstract topics and themes across many documents. This uses word clusters and latent semantic structuring.
What is machine translation in the context of language models?
Language models translate texts from one language to another. This supports businesses with their global expansion efforts. New models often performs on par with commercial products.
What is the third thing that has contributed to the rise of generative models?
Large labeled datasets
What sort of calculations are used to train deep learning neural nets?
Matrix/vector computations
True/False, why. If the internet had never been invented, would G.AI models be possible?
Maybe, but likely False. This is because the internet produces so much data (text, images, videos, etc.) which can fuel these ML models.
What attention mechanism does ChatGPT use?
Multi-query attention (MQA)
What does NLP stand for in the context of machine learning?
Natural language processing
In dimensionality reduction, what is selection?
Only some features are conserved
What is sentiment analysis in the context of language models?
Opinions and emotions and text can be used for businesses to understand customer feedback and opinions more efficiently.
What is the fifth challenge that impacts business operations use of AI?
Potential Misuse
What is the difference in types of models?
Predictive, explanatory or generative
T/F Is a variational autoencoder probabilistic or deterministic?
Probabilistic!
Generally speaking language models have applications in what areas?
Question answering, automatic summarization, sentiment analysis, topic modeling, semantic search, machine translation
What is the acronym which talks about the challenge of impact for AI in business operations?
Rab, Decamped from the mountains in Bali.
What is dimensionality reduction?
Reducing the number of features which describe data
What is the first thing that has contributed to the rise of generative models?
Software libraries
What is the second thing that has contributed to the rise of generative models?
Software tools
What is "Tokenization"?
The process of splitting txt into tokens.
In dimensionality reduction, what is a decoder?
The process whereby "Old features" are produced from "new features".
What is meant by Deep Learning (DL)?
This is a subsection of AI which develops algorithms that learn from data.
What is meant by "latent space" of input data?
This is the "area" that information is encoding "upon" to manipulate it before decoding it or observing it.
True or false, language modeling serves as a way of encoding the rules and structures of a language in a way that machines can understand?
True
T/F why, is it fair to say that dimensionality reduction could be interpreted as data compression?
True, it can be thought of data compression as the encoder "compresses" the data from the initial space, to the encoded or latent space.
T/F why, when prompting LLM models, to get a ~20% increase in the effectiveness of the model, one should provide 5 question-answer pairs.
True, this enables the model to learn expected answers.
T/F, explain: An autoencoder's sole function is to encode and decode with as few losses as possible, no matter the latent space's organization.
True. This is important because as such there is much "overfitting", even with a perfect encoder. This results in meaningless decoding information or noise.
What is semantic search in the context of language models?
Understanding the meaning of themes within individual topics utilizing NLP to interpret words and concepts for improved relevance
What are the two phases of training a GPT like "agent"?
Unsupervised pre-training and discriminative task-specific fine-tuning.
Given no restrictions on an autoencoder, it will take advantage of overfitting to achieve the task of perfect encode/decode. As such, what is a method to solve this?
Utilization of regularization
When is an encoding-decoding combination considered lossless?
When the initial data, x, is = to d(e(x)). This is to say under transformation e, which is the encoder, and transformation d, which is the decoder, it is still the same.
In transformer architecture, what is Layer normalization?
a technique to stabalize network learning by "normalizing" the model's inputs across the features dimension.
What is the main difference between an autoencoder and a variational autoencoder?
a variational autoencoder has generative capabilities derived from its utilization of good latest space properties (regularized)
What does perplexity measure?
how well a model can approximate the training dataset
What is a context vector's use?
is used to selectively attend to relevant parts of the input during decoding at subsequent model layers.
What is an example of representation learning in ML?
learning to represent visual features such as edges, shapes, and textures. It learns representations of raw pixel data to help make predictions.
In dimensionality reduction, what is extraction?
reduced new features based on the old features
What is a context vector?
take a weighted sum of the input sequences values based on similarity between current and other positions.
How does a Variational Autoencoder work?
the encoder-decoder NN architecture is fed data. The encoded-decoded output with the initial data backpropagate the error through the architecture to update the weights. Thus reducing the error.
In dimensionality reduction, what is an encoder?
the processes whereby "new features" are produced from the "old features"
