Evolution of AI and Key Concepts of Gen-AI

Collins Dictionary recently declared AI as the Word of the Year 2023. No one should be surprised by that. But with GenAI becoming the most popular theme right now, is that all of AI? And what happened to other terms that we learned over last couple of years? Let me try and decode the relation between AI, ML, Deep Learning, LLMs, GenAI, etc. and also summarize few key concepts of Gen-AI terminology that we would keep hearing about.

1.????? Artificial intelligence: AI is a discipline. This refers to programs that sense, reason, adapt and act. It is an overall framework which encompasses everything in theory and development of computer systems to be able to perform tasks normally require human cognition and intelligence.

Below figure summarizes the relationship between AI, ML, DL and Gen-AI.

2.????? Machine learning: ML is a subset of AI. It uses algorithms whose performance improves as they are exposed to more data over time. ML gives computers the ability to learn without being explicitly programmed. Typically, machine learning involves the following steps:

• Build or use an existing algorithm.
• Feed the data into an algorithm.
• Use this data to train a model.
• Test and deploy the model.
• Consume or use the model for the intended purpose (in an application).

Most of the machine learning algorithms can be categorized into 3 buckets:

2.A.????? Supervised Learning: these algorithms make predictions based on a set of labeled examples that you provide. This technique is useful when you know what the outcome should look like, typically to make predictions. E.g., you have rainfall data of a city for last 100 years across the year and want to make prediction purely based on that that data for next year. For this, the outcome uses labels which are already in the data set: city, rainfall, month, etc.

2.B.????? Unsupervised Learning: these algorithms use the data sets that aren’t labeled—the algorithm labels them for you by organizing the data or describing its structure. This technique is useful when you don’t know what the outcome should look like, typically when you want to build or identify patterns or relationships in the data set, like anomaly detection or categorization. E.g. if you want the algorithm to look at large data and categorize it without any human intervention or identify anomaly in the data. The labels in this care are automatically generated based on the patterns discovered in the data. Typical use cases are to identify consumer behavior or affinity, anomaly or defect detection, etc.

2.C.???? Reinforced Learning: algorithms that learn from outcomes and decide which action to take next. After each action, the algorithm receives feedback that helps it determine whether the choice it made was correct, neutral, or incorrect. It’s a good technique to use for automated systems that have to make a lot of small decisions without human guidance. Typically used in Autonomous cars, NLP, etc.

?

3.????? Deep Learning: It is a subset of Machine Learning in which Multilayered Deep Neural Networks learn from vast amounts of data. In machine learning, the algorithm needs to be told how to make an accurate prediction by consuming more information (for example, by performing feature extraction). In deep learning, the algorithm can learn how to make an accurate prediction through its own data processing, thanks to the artificial neural network structure. Deep Learning models are designed to automatically and adaptively learn from experience using neural networks with multiple layers. These deep networks can model complex patterns in large datasets, a key capability that has empowered the development of both Large Language Models and Generative AI. Within the realm of Deep Learning, there are two fundamental types of models that are commonly used: discriminative models and generative models. These models represent different approaches to understanding data, each with its unique strengths and use cases.

?

3.A.????? Discriminative Models: these models differentiate or discriminate between different types of outcomes. Given an input, a discriminative model is trained to make a decision—usually, this decision involves classifying the input into one of several pre-defined categories. For example, consider a machine learning task to classify emails as either "spam" or "not spam". A discriminative model would be trained on a labeled dataset with these two categories and would learn to recognize the features that distinguish spam emails from non-spam ones. Discriminative models are widely used in classification problems. Above representation shows model for DNN for "Cat" or "Dog".

3.B.????? Generative Model: These models learn the underlying probability distribution of the data, which allows them to generate new data instances that are similar to the training data. Generative models are widely used in tasks that involve generating new content, such as text, images, or music. Some examples of generative models include Naive Bayes classifiers, Hidden Markov Models, and Generative Adversarial Networks (GANs).

?

4.????? Generative AI: Subset of Deep Learning that uses techniques (such as deep learning algorithms or transformers) to generate new content.?Typically, the models leverage massive pre-trained knowledge to generate content. Large Language Models are a type of machine learning model for Natural Language Processing (NLP). These models are trained on vast amounts of text data and can generate human-like text (and many more things like images, code, etc.) by predicting the likelihood of a given word sequence. By learning the statistical patterns of language, these models can produce coherent and contextually relevant sentences. An important breakthrough in Large Language Models came with the advent of Transformer models. Introduced in a paper titled "Attention is All You Need" by Vaswani et al. in 2017, the Transformer model revolutionized NLP tasks with its novel attention mechanism, allowing it to focus on different parts of the input sequence when producing an output. The model's ability to handle long-term dependencies in text, along with its scalability, propelled the Transformer to become the foundation for many subsequent models.

?So, if ML, DL and Gen-AI are related, what differentiates these techniques?

• Data Required: First, the amount of data required. While machine learning can work with small amounts of data, Deep Learning needs large amounts of data for training. GenAI LLM requires even larger data for training. Popular way to estimate minimum data required to train a regression machine learning algorithm is called “10 times rule of machine learning” which states that you need at least ten times as many rows (data points) as there are features (columns) in your dataset. As compared to that, Deep Learning requires millions of data points to train Artificial Neural Networks. And the GenAI LLMs require billions of data points to train. As an example, GPT3 which was trained on 45TB of data, 10 Billion parameters and 175 Billion tokens, Llama is trained on 2 Trillion tokens, PaLm 2 is trained on 3.6 Tillion tokens and GTP4 is rumored to be trained on 100 Trillion. Please note that most of the tech companies who created these models do not officially publish these numbers and hence they are best guesses only – given to help you understand the requirement of data to train these models.

?

• Algorithms: Algorithms are pieces of code that help explore, analyze, and find meaning in complex data sets. The algorithms used in ML or Gen-AI are different. Each algorithm is a finite set of unambiguous step-by-step instructions that a machine can follow to achieve a certain goal. In a machine learning model, the goal is to establish or discover patterns that people can use to make predictions or categorize information.

Below is a list of popular, but nowhere comprehensive, list of algorithms used in various AI techniques: ?

Machine Learning:

• Linear regression, also called least squares regression, is the simplest supervised machine learning algorithm for predicting numeric values. E.g. Predicting the total number of sales based on the amount of advertising or the season of the year.
• Support vector machines (SVMs) are a kind of parametric classification model, a geometric way of separating and classifying two label classes. E.g. Predicting the total number of sales based on the amount of advertising or the season of the year.
• Decision trees (DTs) are a non-parametric supervised learning method used for both classification and regression. Decision trees are powerful tools that can support decision making in different areas such as business, finance, risk management, project management, root cause analysis tools and etc.
• Random forest model produces an ensemble of randomized decision trees and is used for both classification and regression. E.g. Predicting the creditworthiness of a loan applicant in banking, diagnosing patients in health care, identifying potential markets for stocks in stock market, Predicting the preference of customers in e-commerce, etc.
• K-means clustering problem attempts to divide n observations into k clusters using the Euclidean distance metric, with the objective of minimizing the variance (sum of squares) within each cluster. It is an unsupervised method of vector quantization, and is useful for feature learning, and for providing a starting point for other algorithms. E.g. Categorizing students into grades based on their scores, Creating sets of similar customers for targeted marketing, Finding optimal locations for pizza stores based on delivery areas.

?

Deep Learning:

• Convolutional neural networks (CNNs) are a type of deep neural network often used for machine vision. They have the desirable property of being position-independent. It uses algorithms like Gradient descent, Newton’s method, etc. It is typically used in OCR and image recognition, Detecting objects in self-driving cars, social media face recognition, Image analysis in medicine, etc.
• Recurrent neural networks (RNNs): While convolutional neural networks do a good job of analyzing images, they don’t really have a mechanism that accounts for time series and sequences, as they are strictly feed-forward networks. Recurrent neural networks (RNNs), another kind of deep neural network, explicitly include feedback loops, which effectively gives them some memory and dynamic temporal behavior and allows them to handle sequences, such as speech, Language Modelling, machine translation, etc.

?

• Generative AI: Gen-AI uses Large Language Models based on Transformers. These are neural networks that solely use attention mechanisms, dispensing with recurrence and convolutions entirely. Attention units (and transformers) are part of Google’s BERT (Bidirectional Encoder Representations from Transformers) algorithm and OpenAI’s GPT-2 algorithm (transformer model with unsupervised pre-training) for natural language processing. Transformers continue to be integral to the neural architecture of the latest large language models, such as ChatGPT/Bing Chat (based on GPT-3.5 or GPT-4) and Bard (based on LaMDA, which stands for Language Model for Dialogue Applications).

I plan to deep dive into these techniques in the subsequent articles. But before that, let’s take a quick look at some of the key concepts that we would be using to compare or differentiate these as well as to understand key terms used in the context of AI:

1. Neural Network: An artificial neural network is a system of hardware and/or software patterned after the operation of neurons in the human brain. Neural networks mimic how the human brain operates, enabling computer programs in AI, machine learning, and deep learning to identify patterns and address common issues. The structure of artificial neural networks is based on the concept of nodes, which contain input, hidden, and output layers. There are connections between each node or artificial neuron, and each one has a threshold and weight that go along with it. Changes in weight values allow artificial neural networks to learn.
2. ?GPT (Generative Pre-trained Transformers): Generative means that it can generate the next best word, pixel, audio note, code, etc. that makes sense in the context of what the AI is being asked to output. Pre-trained refers to it being trained on millions, billions, or trillions of pieces of information. The transformer is the foundational component in the LLM stack and is in use by all current LLM models. Transformers work so well in LLMs because they model the relationship between words and concepts in text. They can distinguish the context of how words are used and ensure that fidelity and nuance are maintained. For example, a word like “novel” can refer to a book or to something that is new or original and the transformer is able to distinguish the intended meaning. Transformers are also able to learn to represent words and phrases in a more conceptual way which gives it the ability to generate natural-sounding output.
3. ?Feature: In artificial intelligence, a feature is a measurable property of the object you’re trying to analyze. In datasets, features appear as columns. Each feature, or column, represents a measurable piece of data that can be used for analysis. Depending on what you’re trying to analyze, the features you include in your dataset can vary widely.?For example, if you have dataset about a City population, then things like age, gender, wealth, car or home ownership, number of kids, etc. could be feature or columns of the dataset.
4. ?Tokens: Large language models (sometimes referred to as GPT's) process text using tokens, which are common sequences of characters found in a set of text. The models learn to understand the statistical relationships between these tokens and excel at producing the next token in a sequence of tokens. A helpful rule of thumb is that one token generally corresponds to ~4 characters of text for common English text. This translates to roughly ? of a word (so 100 tokens ~= 75 words). E.g. “Who is Virat Kohli” translates to 18 characters and 6 tokens.
5. Parameter: a parameter is a value that controls the behavior of a machine learning model. Parameters play a crucial role in controlling the output of the model. For example, in a text generation model, parameters can control the style, tone, and content of the resulting text. By adjusting the parameters of the model, one can influence the output of the model and generate text that meets specific criteria. In general, the more parameters a model has, the more accurate and powerful it is. With more parameters, a model can learn a wider range of relationships and patterns in the data. This can lead to better performance and improved accuracy in predictions. However, having too many parameters can also be a drawback. With too many parameters, a model can become overfit, meaning that it is too closely tied to the training data and may not generalize well to new data. Overfitting can lead to poor performance on test data and can limit the model’s ability to make accurate predictions.
6. Prompt Engineering: In Generative AI, there are two ways you can improve the response outcomes – Prompt Engineering and Fine-tunning. Prompt engineering is all about effectively communicating our needs to AI applications, ensuring they can assist us in a precise and accurate manner. The prompt converts your natural language questions into computations that are understood by the LLM. Sometimes the above is sufficient to elicit the information you need in the format you need it in, but sometimes your tasks might be more specific, or maybe you've asked for something entirely new to the LLM. What do you do when you don't get consistently good results even with the above? This is where you can add even more context to the prompt by giving it examples of what you're looking for through techniques like Zero-shot, Single-shot, Few-shot, Chain-of-thought, etc. prompting methods.
7. Fine Tuning the Model: The second method to unlock the full potential of Generative AI is fine-tuning. It empowers individuals to take benefit of pre-trained AI models and tailor them to specific domains, making it an invaluable skill for those seeking to harness AI effectively and align it with their unique requirements. In essence, fine-tuning allows us to refine pre-trained models such as GPT-3.5 or GPT-4 for precise applications by subjecting them to a more targeted dataset that closely aligns with the specific task at hand (i.e., an application focused on giving legal opinions based on existing knowledge base). Fine-tuning involves furnishing the model with a dataset containing labeled examples pertinent to the specific task at hand.
8. Hallucination: Gen-AI hallucination is where a large language model (LLM) makes up false information or facts that aren’t based on real data or events. Hallucinations are completely fabricated outputs from large language models. Even though they represent completely made-up facts, the LLM output presents them with confidence and authority. As you go from lower end of models to more sophisticated models, hallucinations tend to reduce – e.g. according to the GPT-4 technical report, the new model is 19% to 29% less likely to hallucinate when compared to the GPT-3.5 model.
9. Jailbreak: Most of the LLMs are trained on vast datasets, can generate coherent and contextually appropriate responses and also have safeguards against inappropriate response. However, persistent expert users can identify vulnerabilities and exploit them to “jailbreak” AI chatbots. Jailbreak attacks are User Prompts designed to provoke the Generative AI model into exhibiting behaviors it was directed to avoid or to break the rules set in the System Message. These attacks can vary from intricate role-play to subtle subversion of the safety objective.
10. RAG (Retrieval Augmented Generation): RAG is an architecture that augments the capabilities of a Large Language Model (LLM) like ChatGPT by adding an information retrieval system that provides the data. Adding an information retrieval system gives you control over the data used by an LLM when it formulates a response. For an enterprise solution, RAG architecture means that you can constrain natural language processing to your enterprise content sourced from vectorized documents, images, audio, and video. ?For e.g., if an insurance company create a chat BoT using LLM, it would want to constraint the response to only its insurance policies and user data, rather than generic information. The company can achieve this through RAG.

11. Small Language Models (#SMLs): In the world of artificial intelligence (AI), bigger has often been seen as better. The advent of large language models (LLMs) like GPT-4 has sparked both awe and concern, as these giant AI models have demonstrated remarkable natural language understanding and generation capabilities. But in the shadow of these behemoths, a quiet revolution is taking place. Recent research suggests that smaller language models, once thought to be mere steppingstones to their larger counterparts, are starting to outperform—or at least match—the performance of LLMs in various applications. At Ignite 2023, Microsoft AI announced the newest iteration of the Phi Small Language Model (SLM) series termed Phi-2. This comes at a time when many industry members are voicing their opinions that smaller models are going to be more useful for enterprises in comparison to Large Language Models (LLMs). A large language model uses its huge complement of parameters to memorize information contained in its pretraining and fine-tuning datasets. It wouldn’t need to memorize so much — and thus wouldn’t need so many parameters — if it had access to documents on demand, which is possible and key requirement in enterprise environment, anyway.

The author is currently working with Microsoft. The views are personal. All brand, trademarks and IP rights are acknowledged.