Generative AI¶

Introduction¶

Generative AI is a combination of neural network models to create new content (text, image, music, videos..) from a requesting query. Models are pre-trained on vast amounts of unlabeled data, using from 7B up to 500B of parameters. Current Gen AI models are based on the Transformer architecture.

Gen AI applies well to different category of use cases: improve customer experiences, improve employee's productivity, help around creativity, and help optimizing business process (See also the Use Case section).

Transformer Architecture¶

Transformer is a neural network used to generate the next word in the sentence using the best probability. This is what most chat application use to propose the next word sugestion we can select. If we just select the most likely words we got paragraph with no real meaning. The context of the text is lost at each word. Adding the self-attention mechanism to the transformer, it helps to weight the significance of different words by taking into account the previously seen context.

The attention mechanism computes the similarity between tokens (from the embeddings of words) in a sequence. That way, the model builds an intuition of what the text is saying. The closer two words are in a vector space, the higher the attention scores they will obtain and the higher the attention they will give to each other (Recall the example of "bank of the river" vs "money in the bank").

The transformer architecture looks like:

• The tokenization step takes every word, prefix, suffix, and punctuation signs and assign a matching token
• Embedding to transform token to numerical vector
• Positional encoding consists of adding a sequence of predefined vectors to the embedding vectors of the words. This ensures we get a unique vector for every sentence, and sentences with the same words in different order will be assigned different vectors
• The attention component is added at every block of the feedforward network. It uses multi-head attention where several different embeddings are used to modify the vectors and add context to them.
• There is a large number of transformer blocks in the network. The architecture has the layers of transformers stacked on top of each other. Within each layer, there are feed-forward neural networks to process the data.
• The last step of a transformer is a softmax layer, which turns these scores into probabilities (that add to 1). The model returns result tokens which are then turned back into readable text.

The models are trained on vast amounts (Terabytes) of text data like books, articles, websites etc. This helps the model learn grammar, facts, reasoning abilities and even some level of common sense from the content.

GPT-3 (Generative Pre-trained Transformer 3) breaks the NLP boundaries with training on 175B parameters.

The training has two stages: Pre-training where the model attempts to predict the next word in a sentence using its own corpus, and fine tuning where the model can be tuned for specific tasks or content. During the pre-training process, the model automatically takes context into account from all the training data, and tracks relationships in sequential data, like the words in a sentence, to develop some understanding of the real world.

The models are commonly referred to as foundation models (FMs).

The unlabeled data used for pre-training is usually obtained by crawling the web and public sources.

At inference time, the input text is tokenized into individual tokens which are fed into the model.

The largest pre-trained model in 2019 (BERT) was 330M parameters while the state-of-the-art LLM in 2023 is 540B parameters.

A transformer-based model has an encoder component that converts the input text into embeddings, and a decoder component that consumes these embeddings to emit some output text. Transformers process the entire input all at once, during the learning cycle, and therefore can be parallelized.

Three types of transformer:

1. Encoded only: generate no human readable content, used when applications need to efficiently query content to find similar items.
2. Encoder-decoder model is trained to treat every natural language processing (NLP) problem (e.g., translate an input string in one language to another) as a text-to-text conversion problem.
3. Decoder-only model is for text generation.

Models with encoder-decoder and decoder-only architectures are generative models.

The process is text -> tokens (a token may be less than a word, and on average a 5 chars) -> vector. Vectors of similar word are close in the multi-dimensional space. A vector, in NLP, has a lot of dimensions, representing its characteristics in the world of meaning. The best tokenization method for a given dataset and task is not always clear, and different methods have their own strengths and weaknesses. Sub-word tokenization combines the benefits of character and word tokenization by breaking down rare words into smaller units while keeping frequent words as unique entities.

HuggingFace Transformer¶

HuggingFace Transformer provides thousands of pre-trained models to perform tasks on text, images and audio.

Pre-training process¶

The goal of pre-training is to teach the model the structure, patterns and semantics of the human language. The pre-training process for GPT-3 involves collecting and preprocessing vast amounts of diverse text data, training a Transformer-based model to predict the next token in a sequence, and optimizing the model using powerful computational resources.

Corpus = a collection of texts, and a vocabulary is the set of unique tokens found within the corpus. Corpus needs to be large and with high quality data.

The process looks like in the figure below:

For the data collection, it is import to get diverse source of data, including web sites, books, curated datasets to address wide range of topics, writing styles and linguistic nuances. Data preparation is still key, but complex as to remove low-quality text, harmful content... As part of this preparation, text can be converted to lowercase to reduce variability. Tokenization helps to handle rare words and different languages.

The primary objective during pre-training is to predict the next token in a sequence. This is a form of unsupervised learning where the model learns from the context provided by preceding tokens.

The training phase includes the forward pass where input tokens go through the transformer layers. The loss calculation is computing the difference between predicted token and actual next token. Finally the backward pass applies gradients computation to minimize the loss, and tune the model parameters.

The entire dataset is split into batches, and the model is trained over multiple epochs.

The optimization phase includes tuning hyper parameters like learning rate and batch size. To be able to scale we need to run training on distributed computers.

A portion of the data is set aside as a validation set to monitor the model's performance and prevent overfitting.

Perplexity is a common metric used to evaluate language models, measuring how well the model predicts a sample

Generic development approach¶

Some ways to use Generative AI in business applications:

• Build foundation model from scratch: very expensive and time consuming, with highly skilled ML scientists.
• Reuse existing foundation models available as open-source (hundred of model on Hugging Face hub), then add own corpus on top of it, to fine tune the model for better accuracy.
• Use generative AI services or APIs offered by foundation model vendors. There is no control over the data, cost and customization. Use prompt engineering or RAG techniques to get better answers, .

Model fine-tuning¶

The Huggingface LLM leader board is a good source of information for model quality assessments relative to certain use cases.

See this detailed article on developing a LLM and this one on fine tuning.

Use cases¶

We can group the Generative AI use cases in different categories:

• Uber has conducted Hackathon using genAI and identified that all Software Development Life Cycle phases are impacted by Generative AI usages, including: capturing complete specifications faster, explain existing code, generating UI code, automate code refactoring, unit test generation or e2e testing for mobile app, review code, code relationship map automatically created from call stacks, Pull Request and code review automation, code documentation generation based on PRs. But LLMs may generate buggy code, as well as spreading error-prone code pattern.

• Generative Adversarial Networks are used to limit the risk of adversarial manipulation in deep learning image recognition. It attempts to generate fake data that looks real by learning the features from the real data.

It would be difficult to find any business use-case where a base FM can be used effectively. Added techniques are needed to be useful in enterprise, like RAG, fine tuning, new training, knowledge graph and neuro-symbolic AI solutions.

Industries¶

These following industry-specific use cases present the potential applications of Generative AI:

1. Supply Chain Management:

   • Improve visibility into multi-tier supplier performance concerns
   • Identify potential risk areas within the supply chain

2. Quality Control and Nonconformance Management:

   • Identify the root cause of nonconformance issues
   • Prescribe resolutions to address quality concerns

3. Engineering Cost Optimization:

   • Promote the reuse of common parts across different platforms to reduce costs

4. Cross industry:

   • Improve chatbot user's experience, with open responses more empathic to the user.
   • Sentiment analysis: Gauge customer sentiment towards products, services, or brands
   • Assist with proofreading tasks
   • Update and maintain databases
   • Analyze customer reviews
   • Monitor social media platforms

5. Education and Universities:

   • Moderate and develop educational content
   • Help students find the most effective pathways to graduation

6. Safety and Risk Management:

   • Identify potential safety risks, such as gas leaks
   • Generate recommendations for remedial work to mitigate risks

7. Travel Industry:

   • Enhance trip planning with personalized recommendations, services, and offers

8. Product Review Summarization:

   • Offload the task of summarizing product reviews from humans to LLMs
   • Add unstructured reviews as a new corpus for search functionality
   • Separate reviews based on user-provided ratings
   • Task an LLM to extract different sets of information from each high-level category of reviews

Classical concerns and challenges¶

LLM's are amazing tools for doing natural language processing. But they come with challenges due to the underlying training and inference technology, due to the fact that they are trained only occasionally and are thus always out of date, and also due to the fact that natural language generation is not grounded in any model of reality or reasoning but instead uses probabilistic techniques based on correlations of a huge number of strings of tokens (words). Which means hallucination and approximate retrieval are core of their architecture: the completion they are generating is in the same distribution as the text they have been trained on. Prompt engineering does not change hallucination as the decision to assess the response is a factual completion depends of the knowledge of the prompter and requires to continuously assess all the responses.

• Accuracy: The accuracy of LLM's is not acceptable to any enterprise that must follow regulations and policies and respect contractual agreements with suppliers and customers. Because they cannot truly reason or take into account regulations and policies precisely, models often produce incorrect and contradictory answers when asked for decisions or actions to undertake. A single large language model is unlikely to solve every business problem effectively. With classical ML, probabilistic output is expected. Symbolic approaches like business rules that precisely express policies produce reliable results at the cost of coding the policies mostly manually.

• Specificity: A single large model is unlikely to solve every business problem effectively because it is trained on generally-available information rather than enterprise-specific information. To differentiate their generative AI applications and achieve optimal performance, enterprises should rely on their own data sets tailored to their unique use case. Even then, enterprise data changes constantly, so techniques such as RAG and tool calling are needed to leverage the most up-to-date and relevant information for a specific query.

• Cost and Risk of training and inference, as well as privacy and intellectual property are top concerns. LLM's can be "fine-tuned" for a specific task by using a small number of labeled examples specific to the company's industry or use case. Fine-tuned models can deliver more accurate and relevant outputs. But training and retraining models, hosting them, and doing inference with them are expensive. Cloud providers see this opportunity to sell more virtual servers with GPU's at a higher price.

• Skills: developing a new LLM may not make sense today, but fine tuning an existing model may in some circumstances. There are relatively few developers with expertise in model tuning, understanding their architecture and limitations, integrating them in applications, and in tuning their hyper parameters. Reinforcement learning to fine-tune existing LLM requires a huge number of trials, and data quality is still a very difficult and poorly-mastered topic.

• Reliability and reasoning: Generative AI models do not reason and do not plan accurately. New versions of LLMs attempt to improve in this domain, but by design the transformer architecture is probabilistic and greedy for text generation and does not inherently do any kind of structured symbolic reasoning or manage ontologies of concepts (knowledge graphs). LLM is a very big system-1 with their knowledge based from digital representation of humanity created content.

• For generative AI, the input is very ambiguous, but also the output: there is no determinist output. Models produce incorrect and contradictory answers. With classical ML, output is well expected. Trained sentiment analysis algorithms on labelled data will perform better than any LLM for that task. Always try to assess when to use ML versus using an existing LLM.

• There are a lot of models available today, each with unique strengths and characteristics. How to get the ones best suited for business needs? The size of the model is linked to the compute power developers have and how much they can pay (availability of the hardware is also an issue). Effectice mechanisms need to be set up to assess the LLM's responses for each different use cases.
• There is difficulty to determine which source documents are leading to the given answers. Model that links results to source, via citations help to assess for any hallucinations.
• Developers need to optimize for training and inference cost, then assess how to amortize the training cost to better evaluate what may be billed to end-users.
• Current models comparison is based on Multi-task Language Understanding on MMLU benchmark. But CIOs care less about standard results, they want models that work well on their data.
• For large enterprise, adopting LLM at scale means running hundreds, even thousands, of models at any time. A high frequency of innovation, leads customers to replace their models quicker than expected, reinforcing the need to train and deploy new models in production quickly and seamlessly.
• Cost being a major issue in short term, model may become smaller and access to powerful distributed smaller hardware will help to do inference locally (smartphone with TPU).

Some fallacies¶

• LLMs can't do planning in autonomous modes. They may support planning activities done by planner software and human to translate formats, or to elaborate the problem specifications.
• Chain of Though, ReAct, Fine tuning do not help for planning as they do not generalize well
• There is no self-verification as LLM has no mean to do self verification

Interesting legal considerations¶

• Think not created by a human could not be copyrighted.
• Model deployed will not use data sent to improve itself.
• The right to use an image/photo to train a model is a problem.
• Protect Intellectual Property: never pass confidential information to Gen AI SaaS based API.
• Protect the brand, avoid bias, discrimination, aligned to company values: any business decision should not be done by uncontrolled Gen AI.

Discovery assessment¶

Classical questions to address before starting a Gen AI solutions:

• How to leverage open model like Mistral, LLama, DBRX?
• How to verify accuracy on unstructured data such as queries and document shunks?
• AI Models are becoming stronger when connected to data, enterprise data, and fine tuning. It is important to adopt a Gen AI strategy that is linked to the data strategy too.
• When transitioning to more stateful Gen AI-based solutions to make the model more specialized or better tuned for dedicated use cases, the model will require self-improvement capabilities. To support this, the solution will need to maintain states, which will be persisted using Lake House technology
• Current main stakeholder for Gen AI, is the developer, moving to business users will be challenging as some of their jobs are at risk, they may reject the technology at a all.
• How to stop doing a lot of prompt engineering and start doing model fine tuning? Always address what we should evaluate the solution on. Gen AI can help to build prompt with the use of meta prompting techniques.
• How to measure hallucination where the models make up inaccurate responses that are not consistent with the training data.
• How to integrate AI capability in the enterprise business processes and decisions? How does human in the loop step can be added to the process?
• Whenever we want to teach an LLM to use a tool, we need enough annotated tool calls to fine tune the LLM. We can use in-context learning to create a model that annotates tool calls for the input query. Incorrect calls can be filtered by executing the tools and filtering the outputs based on the ground truth answer.

Concepts¶

A LLM is part of the evolution of NLP as it is a trained deep learning model that understands and generates text in a human like fashion.

NLP processing¶

To process an input text with a transformer model, the text is tokenized into a sequence of words or part of words. These tokens are then encoded as numbers and converted into embeddings, which are vector-space representations of the tokens that preserve their meaning: for example a word dog will have 512 potential numerical attributes used to describe what is a dog. Below is a simple representation of the embedding in the 3 dimension space:

See the web site projector.tensorflow.org/

The encoder, in the transformer, transforms the embeddings of all the tokens into a context vector. Using this vector, the transformer decoder generates output based on clues. The decoder can produce the subsequent word. We can reuse the same decoder, but this time the clue will be the previously produced next-word. This process can be repeated to create an entire paragraph. This process is called auto-regressive generation.

When processing text, the AI looks at a few tokens around each word to help understand the context. This surrounding group of tokens is called the context window. It is the sliding group of tokens around a word that provides contextual information to help the AI understand and generate natural language.

Transformers do not need to code the grammar rules, they acquire them implicitly from big corpus.

During the training process, the model learns the statistical relationships between words, phrases, and sentences, allowing it to generate coherent and contextually relevant responses when given a prompt or query.

The techniques to customize LLM applications from simplest to more complex are:

• Zero-shot inference: allows a pre-trained LLM to generate responses to tasks that it hasn’t been specifically trained for. In this technique, the model is provided with an input text and a prompt that describes the expected output from the model in natural language.
• Prompt engineering with zero-shot inference.
• Prompt engineering with few-shot inference: Few-shot learning involves training a model to perform new tasks by providing few examples. This is useful where limited labeled data is available for training.
• Retrieval augmented generation (more complex).
• Fine tune an existing foundation model.
• Pre-train an existing foundation model: example is domain specific model, like the Bloomberg's LLM.
• Build a foundation model from scratch.
• Support human in the loop to create high quality data sets.

Important Terms¶

Summarization¶

Text summarization is a Natural Language Processing (NLP) technique that involves extracting the most relevant information from a text document and presenting it in a concise and coherent format.

Summarization works by sending a prompt instruction to the model, asking the model to summarize our text.

See hands-on notes on LangChain.

Retrieval augmented generation (RAG)¶

See separate chapter.

Common LLM inference parameters¶

Randomness and Diversity¶

Foundation models support the following parameters to control randomness and diversity in the response:

Temperature – Large language models use probability to construct the words in a sequence. For any given next word, there is a probability distribution of options for the next word in the sequence. When we set the temperature closer to zero, the model tends to select the higher-probability words. When we set the temperature further away from zero, the model may select a lower-probability word which leads to creative output.

In technical terms, the temperature modulates the probability density function for the next tokens, implementing the temperature sampling technique. This parameter can deepen or flatten the density function curve. A lower value results in a steeper curve with more deterministic responses, and a higher value results in a flatter curve with more random responses.

Top K – Top K defines the cut off where the model no longer selects the words. For example, if K=50, the model selects from 50 of the most probable words that could be next in a given sequence. This reduces the probability that an unusual word gets selected next in a sequence.

In technical terms, Top K is the number of the highest-probability vocabulary tokens to keep for Top-K-filtering - This limits the distribution of probable tokens, so the model chooses one of the highest-probability tokens.

Top P – Top P defines a cut off based on the sum of probabilities of the potential choices. If we set Top P below 1.0, the model considers the most probable options and ignores less probable ones. Top P is similar to Top K, but instead of capping the number of choices, it caps choices based on the sum of their probabilities. For the example prompt "I hear the hoof beats of ," we may want the model to provide "horses," "zebras" or "unicorns" as the next word. If we set the temperature to its maximum, without capping Top K or Top P, we increase the probability of getting unusual results such as "unicorns." If we set the temperature to 0, we increase the probability of "horses." If we set a high temperature and set Top K or Top P to the maximum, we increase the probability of "horses" or "zebras," and decrease the probability of "unicorns."

Lengths¶

The following parameters control the length of the generated response.

Response length – Configures the minimum and maximum number of tokens to use in the generated response.

Length penalty – Length penalty optimizes the model to be more concise in its output by penalizing longer responses. Length penalty differs from response length as the response length is a hard cut off for the minimum or maximum response length.

In technical terms, the length penalty penalizes the model exponentially for lengthy responses. 0.0 means no penalty. Set a value less than 0.0 for the model to generate longer sequences, or set a value greater than 0.0 for the model to produce shorter sequences.

Repetitions¶

The following parameters help control repetition in the generated response.

Repetition penalty (presence penalty) – Prevents repetitions of the same words (tokens) in responses. 1.0 means no penalty. Greater than 1.0 decreases repetition.

Vector Database¶

A vector database is optimized for storing and querying large vector arrays using machine learning techniques. It's highly scalable and fast at performing operations like similarity searches across vectors.

Similarity search helps to identify items (vectors) that share similar characteristics or properties with the query item (a new vector).

Queries return results based on vector similarity scores, revealing hidden semantic connections in data.

FAISS from Facebook is a library for efficient similarity search and clustering of dense vectors. Faiss can compute vector Euclidien distance using GPU or CPU.

ChromaDB is an open source embedding database which supports Queries, filtering, density estimation and similarity search. It can persist on local disk or use a server deployment. It uses collection for storing the documents, metadatas, embeddings, and ids. Chroma DB by default uses a sentence transformer model to calculate embeddings.

(Code using ChromaDB end to end solution with qa-retrieval) or code langchain/rag folder specially build_agent_domain_rag.pyhttps://github.com/jbcodeforce/ML-studies/blob/master/llm-langchain/rag/build_agent_domain_rag.py

Docker compose to start chromadb

  chroma:
    image: ghcr.io/chroma-core/chroma:latest
    volumes:
      - ./chromadb/.chroma/index
    ports:
      - 8005:8000

Traditional open source index or database such as OpenSearch, Postgresql support now vector store and similarity search.

Current Technology Landscape¶

OpenAI - ChatGPT¶

OpenAI is an AI research and deployment company. Their vision: intelligence—AI systems are generally smarter than humans:

1. With broad general knowledge and domain expertise, GPT-4 can follow complex instructions in natural language and solve difficult problems with accuracy.
2. DALL·E 2 can create original, realistic images and art from a text description. It can combine concepts, attributes, and styles.
3. Whisper can transcribe speech into text and translate many languages into English.

Chat Generative pre-trained Transformer is a proprietary instruction-following model, which was released in November 2022. It is a system of models designed to create human like conversations and generating text by using statistics. It is a Causal Language Model (CLM) trained to predict the next token.

The model was trained on trillions of words from the web, requiring massive numbers of GPUs to develop. The model was trained using Reinforcement Learning from Human Feedback (RLHF), using the same methods as InstructGPT, but with different data collection setup.

• Build an AI that can answer questions about your website: crawl, use embeddings, and a search function. It is a good starting point for knowledge based app.

Meta with LLama 2 & 3¶

A foundational, 65-billion-parameter large language model created by Facebook which has been open sourced for academic use. Many models have been released based on Llama2, but they also inherit the license requirement for non-commercial use.

It is possible to run LLama2 on local machine with ollama, and a simple LangChain

• Anakin is a platform to access different models
• Perplexity Labs, a part of Perplexity AI, provides a user-friendly platform for developers to explore and experiment with large language models, including Llama 3
• https://huggingface.co/chat/
• Replicate to run, fine-tune open-source models, and expose them as APIs. They also lead Cog an open-source tool for packaging machine learning models.
• Vercel serverless, hosting platform for web app and AI app. The led vite for next.js project. See a langchain nextjs template to easily deploy to Vercel.

Mistral - Mixture of Experts¶

A french company who has developed the Mixtral 8x7B model, a high-quality sparse mixture of experts model (SMoE) with open weights.

Google Gemini¶

Is the public Generative multimodal AI from Google DeepMind team with the support of 3 different sizes, the smallest being able to run on Mobile. Its reasoning capabilities can help make sense of complex written and visual information.

Amazon SageMaker¶

SageMaker Jumpstart provides pre-trained, open-source models for a wide range of problem types to get started on ML.

It supports training on LLMs not in Bedrock, like OpenLLama, RedPajama, Mosaic Pre-trained Transformer-7B, Flan-T5/UL2, GPT-J-6B, NEOX-20B and Bloom/BloomZ, with a gain of up to 40% faster.

Some useful articles:

• AWS- Quickly build high-accuracy Generative AI applications on enterprise data using Amazon Kendra, LangChain, and large language models.
• SageMaker my own personal study..

Deeper dive¶

• Vulnerabilities of LLM.
• GANs for Synthetic Data Generation.
• Artificial Intelligence and the Future of Teaching and Learning.
• Fine-tune a pre-trained model HuggingFace tutorial.
• Prompt engineering is the new feature engineering.
• Amazon-sponsored workshop advances deep learning for code.
• RAG with OpenSearch Service.
• Running LLM on local laptop using llama.cpp
• BertNet knowledge graphs from llm
• Git repo explaining the transformer from Alammar.
• Attention Is All You Need - research paper
• On the role of LLM in planning - tutorial

>>> Prompt Engineering