RAG and Prompt Engineering

To understand how to optimize LLM with RAG, it’s important to know that Prompt Engineering is the simplest and least technical method to interact with a large language model (LLM). This approach involves formulating a series of instructions that the model must follow to generate an appropriate output in response to a user’s query. 

Unlike RAG (Retrieval-Augmented Generation), prompt engineering requires fewer data, as it only uses those employed during the initial training of the model, and is less costly, utilizing only existing tools and models.

However, prompt engineering has some limitations. It cannot generate output based on updated or changing information since the model relies solely on pre-trained data, which can become outdated. Moreover, the quality of the output heavily depends on the prompt formulation itself; a poorly constructed or ambiguous prompt can lead to incoherent or inaccurate responses.

Despite these limitations, prompt engineering is an intuitive and cost-effective solution, ideal for extracting information on general topics that do not require a high level of detail. It’s particularly useful when quick responses are needed without the necessity to integrate updated or specific data. In such cases, prompt engineering offers an effective way to leverage LLM capabilities without the costs and complexity associated with methods like RAG.

Differences between RAG and Fine-Tuning a Model

To understand how to optimize LLM, it is necessary to recognize that fine-tuning and RAG are two different approaches to improving LLM performance. RAG focuses on integrating external data to enrich the LLM‘s content. In practice, RAG does not modify the base model but allows it to access a set of specific, up-to-date data in real time.

This approach uses minimal computational resources compared to fine-tuning, as it only requires a document or collection of relevant documents. RAG is particularly effective at reducing the risk of “hallucinations,” i.e., incorrect but convincing responses, as it allows the model to refer to verifiable sources.

Fine-tuning, on the other hand, involves modifying the pre-trained model through further training with a smaller, more specific dataset. This process enables the model to learn more targeted patterns and knowledge, which do not require frequent updates. However, fine-tuning requires a significant amount of data and computational resources, as well as substantial time and financial investment. It is ideal for highly specialized applications where the data is stable over time and where in-depth analysis is required.

Key Differences

• Resources: RAG requires fewer computational resources and can be quickly implemented with the addition of specific documents, while fine-tuning requires a large dataset and significant processing resources.
• Data Updates: RAG allows continuous information updates, incorporating new and verifiable data in real time. Fine-tuning, on the other hand, does not allow for frequent updates once the training process is complete.
• Reduction of Hallucinations: RAG has been shown to reduce hallucinations better than fine-tuning since it bases responses on verifiable external data, while fine-tuning requires a complex and costly process to achieve the same result.
• Specialized Applications: Fine-tuning is preferable in contexts where there is a large amount of specific data and adequate resources, and when it is not expected that the data will frequently change. RAG is more suitable for dynamic scenarios where access to up-to-date information is crucial.

Often, combining RAG and fine-tuning can offer the best of both worlds. Fine-tuning can be used to create a solid and specialized foundation for the model, while RAG can be implemented to keep information updated and relevant, improving the accuracy of responses over time.

In summary, the choice between RAG and fine-tuning depends on the specific needs of the project, the available resources, and the nature of the data being worked with.

Differences between RAG and Pretraining

Pretraining and RAG (Retrieval-Augmented Generation) represent two distinct approaches in the training and use of large language models (LLMs). Although both aim to improve the model’s capabilities, they differ significantly in method, required resources, and specific applications.

In the process of optimizing LLM, pretraining is the initial phase of training an LLM, during which the model acquires a basic understanding of language. This process involves using a large generic dataset to create an internal neural network, similar to the human brain’s learning process through the formation of neural connections.

Pretraining requires a huge amount of computational resources, often thousands of GPUs, and is a long and costly process. It is suitable when there is a very large dataset that can significantly influence the model, and there is a desire to provide a basic understanding of a wide range of topics and concepts.

RAG, on the other hand, integrates external, specific information into a pretrained LLM. This method allows the model to access updated and relevant data in real-time, improving the accuracy of responses and reducing the risk of “hallucinations” (erroneous but convincing responses). RAG uses fewer resources compared to pretraining and can be quickly implemented by adding specific documents. It is ideal for scenarios where information needs to be frequently updated and where quick access to specific data sources is required.

Key Differences

• Resources and Time: Pretraining requires a huge amount of computational resources and time, while RAG is much cheaper and faster to implement.
• Purpose and Applications: Pretraining provides a general knowledge base, useful for models that need to understand and generate language on a wide range of topics. RAG, instead, is used to integrate specific and updated information, improving the accuracy of responses for specialized applications.
• Information Updates: RAG allows the model to stay updated with new information in real-time, while pretraining does not foresee frequent updates after the initial phase.

In the next article, we will examine how Revelis is implementing the Retrieval-Augmented Generation technique in its solutions, sharing case studies and results achieved.

Author: Francesco Scalzo

L'articolo How to optimize LLMs with RAG proviene da Revelis.

In this article, we want to explain what RAG is and how to use it to optimize LLMs.

Retrieval-Augmented Generation (RAG)

The RAG method (Retrieval-Augmented Generation) combines data within a large language model with selected external data sources, such as specific repositories, text collections, and existing documentation. These resources are segmented, indexed in a vector database, and used as references to provide more accurate answers.

RAG is extremely useful because it allows the language model to use specific real-time information from one or more reliable sources. This approach offers significant savings by ensuring an optimal experience without the need to retrain or fine-tune models.

Additionally, it uses fewer resources because, after sending the query to the language model, it only returns the most relevant information, avoiding long and unfocused documents.

How the RAG Pipeline Works

The pipeline of a RAG system is divided into several stages, each crucial to the effectiveness of the overall process:

1. Data Ingestion: Data is collected from various sources such as databases, documents, or live feeds. This data is pre-processed and transformed into formats compatible with the embedding system, which converts them into numerical vector representations.
2. Embedding Generation: The transformed data is converted into high-dimensional vectors, representing the text in a numerical format understandable by the system.
3. Storage in Vector Archives: The generated embeddings are stored in specialized vector databases, optimized for quickly handling and retrieving vector data.
4. Information Search and Retrieval: When a user sends a query, the RAG system uses the vector representations to conduct efficient searches. It identifies the most relevant information by comparing the query vector with those stored in the databases.
5. Response Generation: Using the user’s prompt and the retrieved data, the language model generates a complete response. This process combines the internal knowledge of the model with the retrieved external information, enhancing the accuracy and relevance of the responses.
6. External Data Update: To keep information up-to-date, the system periodically updates documents and their embedding representations. This can be done through real-time automatic processes or periodic batches.

RAG and Indexing

Before effectively using a RAG system, it’s essential to have a solid data indexing process. This involves organizing and storing information in a Vector DB. These databases store information as numerical vectors representing the semantic meaning of the data. Indexing allows for the quick retrieval of relevant information, essential for the subsequent retrieval phase.

Vector DB: The Key to RAG

A fundamental component of the RAG pipeline is the Vector DB, a specialized database designed to manage and search semantic information in vector form. These databases are essential for RAG operation for several reasons:

• Semantic Representation: Vector DBs store information as numerical vectors representing the semantic meaning of the data. This contrasts with traditional databases that store data in structured or unstructured formats without capturing the intrinsic meaning.
• Efficient Search: Thanks to vector representation, Vector DBs allow for similarity-based searches. When a query is transformed into a vector, the database can quickly compare this vector with stored vectors to find the closest matches. This makes searching extremely efficient and precise.
• Scalability: Vector DBs are designed to handle large volumes of data. They can scale to contain millions or billions of vectors, ensuring that searches remain fast even as data volumes grow.
• Flexibility: These databases can be used for a wide range of applications, from document search to product recommendation, thanks to their ability to understand and compare the meaning of data.

User Query

The RAG pipeline begins with user input. The user asks a question in natural language, which the system must interpret and transform into a representation that can be processed by artificial intelligence models. This transformation occurs through vectorization, a process in which the question is converted into a numerical vector that captures the semantic meaning of the query.

For example, if a user asks, “What are the benefits of RAG?” the system translates this question into a vector representing the keywords and context of the query. This vector is then used to search for matches in the knowledge database.

Retrieval

Once the user’s query has been vectorized, the next step is retrieval. In this phase, the system compares the query vector with the vectors in the Vector DB. Using similarity measures, the system identifies the data chunks (e.g., paragraphs or document sections) that are most relevant to the query.

Retrieval is a critical phase because it determines the quality of the information used to generate the response. If the retrieved chunks are not relevant, the final response may be inaccurate or unhelpful. Therefore, the effectiveness of retrieval depends on the search engine’s accuracy and the quality of the indexing.

Augmentation

In the augmentation phase, the data chunks extracted during retrieval are combined with the user’s original query to create a response template. This phase can be seen as data preparation for the generative model.

The response template includes the key facts extracted from the documents and organizes them in a format that facilitates the generation of the final answer. This augmentation process ensures that all relevant information is considered and that the response is contextually accurate and complete.

Generation

The final phase of the RAG pipeline is response generation. In this phase, the generative model (typically an LLM) takes the response template and processes it to produce a coherent and natural response to the user’s query.

The generative model uses the facts and contexts provided during the augmentation phase to construct a response that not only directly answers the user’s question but does so in a way that is easy to understand. This response is formulated using advanced natural language generation algorithms, allowing the system to produce texts that appear human-written.

Advantages of RAG

Integrating external and verifiable data through the RAG architecture offers multiple advantages to LLM, significantly improving accuracy, flexibility, scalability, personalization, and control, without needing to retrain the entire model. RAGs improve the accuracy of responses by reducing the risk of “hallucinations,” or convincing but incorrect answers, as they provide LLMs with citable and verifiable sources by the user.

This dynamic integration allows models to stay up-to-date without needing retraining, making them ideal for large-scale applications and constantly evolving contexts.

In terms of convenience, RAGs allow for the introduction of new data into an existing LLM and the elimination or updating of information sources simply by uploading new documents or files. This significantly reduces the costs and time needed to retrain or fine-tune LLMs while improving efficiency, as queries provided to LLMs are reduced to the most relevant information.

RAG offer developers greater control, allowing them to customize information sources and tailor responses to users’ specific needs. This enhancement of the user experience increases trust in the provided responses. Additionally, the RAG architecture facilitates feedback, problem resolution, and application correction, ensuring a continuous flow of updated and domain-specific information.

From a data sovereignty and privacy perspective, RAGs mitigate risks associated with using confidential information for fine-tuning LLM, keeping sensitive data on-premises. They can also limit access to sensitive information based on different authorization levels, ensuring that only authorized users can access certain data. This approach further strengthens the security and privacy of the information used.

In the next article, we will compare RAG – Retrieval-Augmented Generation with other training and data processing methods, analyzing the advantages and limitations of each approach to understand when and why to choose RAG.

Author: Francesco Scalzo

L'articolo RAG: What It Is and How to Optimize LLMs proviene da Revelis.

In this series of articles, we will delve deeply into this methodology, starting with the fundamentals of LLMs and Prompt Engineering, and then comparing RAG with other training and data processing methods. We will conclude with an overview of the practical application of RAG in Revelis solutions.

LLM: what are Large Language Models

Large Language Models (LLM) are advanced artificial intelligence models trained on vast datasets of textual data, capable of understanding and generating human language. These models use deep learning techniques to predict the next word in a text sequence, enabling them to perform tasks such as automatic translation, sentence completion, and creative text generation.

LLMs are generally non-deterministic. They use sampling techniques like top-k sampling and nucleus sampling to generate responses, introducing variability and creativity into the output. However, they can be configured to behave deterministically by setting a “seed” for the random number generator, ensuring that the same input produces the same output every time.

Key Technical Features of LLMs:

• Size and Complexity: LLMs are characterized by a very high number of parameters. For example, OpenAI’s GPT-3 has 175 billion parameters. Parameters are the weights within the neural network that are optimized during the training process.
• Transformer Architecture: LLMs use the Transformer architecture, introduced by Vaswani et al. in 2017. This architecture relies on attention mechanisms that allow the model to weigh the importance of different words in an input sequence, improving effectiveness in context processing.
• Training on Extensive Data: LLMs are trained on vast corpora of textual data from diverse sources such as books, scientific articles, websites, and digital content. This allows the models to learn a wide range of linguistic and factual knowledge.
• Tokenization and Embedding: Before training, text is divided into tokens, which can be words, word fragments, or characters. Each token is then converted into a numerical vector (embedding) that represents its meaning in a multidimensional vector space.
• Generalization Capability: LLMs can generalize the information learned during training to respond to new queries. This is possible due to the model’s ability to capture complex patterns and relationships within the data.

Source image: AI4Business 

Large Language Models (LLMs) utilize machine learning and natural language processing (NLP) techniques to understand and generate human-like language. Although they are extremely useful for communication and data processing, they have some limitations. They are trained on generic data that does not include specific information related to the desired subject, such as a particular set of internal company data. Moreover, their knowledge is limited over time: the information used for training is not updated, so the material can become obsolete and no longer relevant. Finally, they tend to provide answers based on what they think the user wants to receive, which can lead to incorrect or outdated information, a phenomenon known as “hallucination.”

Differences Between Open Source and Closed LLMs

LLMs can be divided into two main categories: open source and closed.

Open source language models, such as GPT-J and BERT, are publicly accessible with open-source code. This approach offers numerous advantages.

Firstly, transparency: anyone can examine how the model is built and how it works. This not only fosters trust in the technology but also allows the developer community to contribute to the continuous improvement of the model. A crucial aspect of open source LLMs is the possibility of customization. Companies and developers can adapt the model to their specific needs, creating specialized versions for certain tasks. This flexibility is particularly useful in industries that require tailored solutions. Finally, from an economic standpoint, open source models can be used for free, significantly reducing implementation costs. This makes the technology accessible even to small businesses and startups that may not have large budgets.

On the other hand, closed language models, such as OpenAI’s GPT-4, offer distinct advantages, especially in terms of performance and optimization. These models are often developed with advanced resources and high-level infrastructures, ensuring superior performance compared to open source models. The companies that develop them can invest in dedicated hardware and optimization techniques that enhance the model’s efficiency and speed. 

In terms of security, closed models offer greater control over distribution and use. This can be crucial for companies that must comply with strict data privacy and security regulations. By using a closed model, companies can be assured that sensitive data is handled with the highest level of protection. However, access to closed models is generally limited and occurs via paid APIs. This can increase costs for the end user and limit the possibility of customization, as users cannot modify the underlying model. Despite this, for many companies, the benefits in terms of performance and support outweigh these disadvantages.

Pre-Training

Pre-training is a fundamental initial phase for large language models (LLMs). This process uses a vast corpus of textual data to provide the model with a general understanding of language. Through pre-training, the model learns the syntactic, semantic, and contextual structures of the language, developing a solid foundation to tackle various linguistic tasks.

Pre-Training Process

Data Collection:

• Diverse Corpus: The model is trained on a vast corpus of data from books, articles, web pages, and other text sources. This corpus must be large and varied enough to cover a wide range of linguistic styles and topics.

Tokenization:

• Text Segmentation: The text is divided into smaller units called tokens. Tokenization can be based on words, sub-words, or characters. For example, “tokenization” can be segmented into “token”, “ization”.

Model Training:

• Neural Network Model: A neural network, often based on the Transformer architecture, is used to train the model. During pre-training, the model learns to predict the next token in a text sequence (masked or autoregressive) using supervised learning techniques.
• Masking: In models like BERT, a percentage of the tokens is masked, and the model learns to predict the masked tokens based on the surrounding context.

Optimization:

• Optimization Algorithms: Algorithms like Adam or LAMB are used to minimize the loss function, adjusting the neural network’s weights to improve prediction accuracy.
• Learning Rate Scheduling: The learning rate is varied during training to enhance the model’s convergence.

Advantages of Pre-Training:

• General Language Understanding: Provides the model with a solid foundation to understand and generate text coherently and contextually.
• Reusability: Pre-trained models can be easily reused and adapted for specific tasks through fine-tuning, significantly reducing the time and resources needed compared to training from scratch.
• Efficiency: Allows for performant models with a generalist capability that can be subsequently refined for particular tasks, optimizing the use of computational resources.

Example of Pre-Training: 

Imagine wanting to create a virtual assistant capable of answering general questions. During pre-training, the model is exposed to millions of articles and books on various topics. Through this process, the model learns linguistic structures and contextual relationships, becoming capable of understanding and generating text fluently. Once pre-training is complete, the model can be fine-tuned with a specific dataset, such as a company’s FAQs, to specialize in answering customer questions.

Fine-Tuning

Fine-tuning is the process of additional training for a pre-trained large language model (LLM) to adapt it to a specific task or domain. This process allows the model’s capabilities to be refined using a targeted dataset, improving its accuracy and relevance for the desired activities.

How Fine-Tuning Works:

• Dataset Selection: Begin by selecting a dataset pertinent to the specific task. This dataset should represent the type of data the model will process after fine-tuning.
• Model Adaptation: The pre-trained model is further trained on this specific dataset. During this process, the model learns the peculiarities of the new domain, adjusting its parameters to optimize performance on the target task.
• Iteration and Evaluation: Fine-tuning is an iterative process. After each training cycle, the model is evaluated to check for improvements. The results are used to further refine the model through additional training iterations.

Advantages of Fine-Tuning:

• Improved Performance: Fine-tuning allows the model to become highly specialized for a specific task, significantly improving its performance compared to a generic model.
• Data Efficiency: Using a targeted dataset, significant improvements can be achieved with a relatively small amount of data compared to training from scratch.
• Adaptability: Fine-tuning enables a general model to be adapted to specific contexts, greatly expanding its utility in practical applications. For example, a general language model can be fine-tuned to analyze medical texts, legal documents, or customer interactions in a business context.

Example of Fine-Tuning Application:

Imagine wanting to use an LLM to provide customer support in a software company. The pre-trained model can understand and generate general text, but it may not be sufficiently competent in answering specific questions about the company’s products. Through fine-tuning, the model is trained on a collection of documents related to the company’s products, user guides, and frequently asked questions (FAQs). After fine-tuning, the model will be able to provide more precise and relevant answers to customer questions, improving the efficiency and quality of customer service.

Fine-tuning is a powerful technique for specializing large language models, making them more effective for specific tasks and particular domains. Through an iterative training process on targeted datasets, it is possible to significantly improve the model’s performance, ensuring more accurate and relevant responses to user needs.

Source image: AI4Business 

Prompt Engineering

Source image: cobusgreyling.com

Prompt Engineering is the practice of creating effective text inputs (prompts) to guide the responses of LLMs (Large Language Models). It is both an art and a science that requires a deep understanding of how language models work and the dynamics of text generation. The success of many applications based on LLMs depends on the quality of the prompts used, so there are fundamental principles to follow:

• Clarity and Specificity: A good prompt must be clear and specific. The more detailed the prompt, the more accurate and relevant the LLM’s response will be. For example, instead of asking, “Tell me a story,” a more effective prompt would be, “Tell me a story about a knight who saves a princess from a dragon.”
• Adequate Context: Providing sufficient context helps the model better understand the task. Including background information or specifying the desired style and tone can improve the quality of the response.
• Prompt Examples: Including examples within the prompt can help the LLM better understand what is being sought. For instance, if a certain type of output is desired, examples of sentences or paragraphs that reflect the desired format can be included.

Creating effective prompts requires a clear understanding of the goal. Before writing a prompt, one must ask what they want to achieve with the model: are they looking for informative answers, creative stories, or critical analysis? Once the goal is clear, the prompt structure should guide the model toward the desired response. Using complete sentences, avoiding ambiguity, and including relevant details can make a significant difference in the quality of responses.

There are various prompt engineering techniques:

• Zero-Shot Prompting: This approach involves asking the LLM to complete a task without providing specific examples. For example, “Translate this text into French.”
• One-Shot Prompting: In this case, a single example of task completion is provided. For example, “Here is a translation example: ‘Hello’ -> ‘Bonjour.’ Translate this text into French.”
• Few-Shot Prompting: Several examples of task completion are provided. For example, “Here are some translation examples: ‘Hello’ -> ‘Bonjour,’ ‘Goodbye’ -> ‘Au revoir.’ Translate this text into French.”

In the next article, we will delve into the technique of Retrieval-Augmented Generation (RAG), exploring how to effectively integrate information retrieval capabilities with LLMs to improve the performance and quality of generated responses.

Author: Francesco Scalzo

L'articolo Large Language Models (LLM) proviene da Revelis.

Generative artificial intelligence (Generative AI) represents a branch of artificial intelligence that focuses on creating new content, such as text, images, music, and videos. Unlike traditional artificial intelligence systems that rely on predefined models, Generative AI learns from large amounts of data to create new and original content.

We have already discussed the multiple potentials that generative AI offers to businesses and how they can leverage them to gain competitive advantages in their business. In this article, however, we want to explain the general functioning of Generative AI and in which other fields it can find application.

Imagine being able to write a complete article in seconds, or to create a realistic image of a non-existent animal. Generative AI makes all this possible and has the potential to revolutionize various sectors, from medicine to software design.

Generative AI: what is it in terms of the learning process

Learning in generative artificial intelligence (Generative AI) mainly occurs through two phases:

1. Training:

Pre-processing phase: Raw data is cleaned, transformed, and normalized to be used by the machine learning algorithm.

Model selection: The architecture and parameters of the machine learning model are selected based on the type of data and the desired objective.

Training phase: The model is “trained” on a large dataset (the training set). The algorithm learns to recognize relationships and patterns among the data and to use them to generate new content.

Validation phase: The model is tested on a validation dataset to evaluate its performance and its ability to generalize to new data.

2. Generation:

Inference phase: The model is used to generate new content. The algorithm uses the knowledge learned during the training phase to create new data similar to those in the training set.

Post-processing phase: The generated content can be further processed and refined to improve its quality and coherence.

Machine Learning Algorithms

Generative AI uses various types of machine learning algorithms, including:

• Supervised learning: The algorithm is trained on a dataset with examples of desired input and output;
• Unsupervised learning: The algorithm learns from a dataset without examples of desired output;
• Reinforcement learning: The algorithm learns to perform actions in a virtual environment and receives rewards or punishments based on the results obtained.

Deep Neural Networks

Generative AI primarily relies on deep neural networks, computational models designed to mimic the human brain, which need to be trained using large datasets to learn (even complex) patterns and provide accurate predictions.

Neural networks consist of several layers of artificial neurons capable of processing information and learning complex relationships among data by evaluating the surrounding environment. Just as the human brain is tasked with understanding the environment and its changes by providing responses suitable to contingent needs, artificial neural networks use datasets to derive identifying traits.

For those who want to delve deeper, I recommend this interesting article “What is Generative AI?” available on the NVidia website.

Generative AI: What it is in terms of types

Generative AI finds space in various fields, allowing the creation or modification of content of various kinds: from text generation to image and video generation to support for programming.

Generative Text AI

Generative text AI focuses on creating new texts, such as articles, poems, emails, scripts, and can be used in various applications including:

Content creation:

• Automatic writing of articles and blog posts: Text generators like Chat-GPT, Gemini, or Jasper can create complete articles and blog posts in seconds, providing only some basic information;
• Product description generation: Generative AI can be used to create accurate and informative product descriptions for e-commerce websites;
• Creating content for social media: Generative AI can create engaging and captivating social media posts.

Automatic translation:

• Generative AI can translate texts from one language to another accurately and fluently, overcoming the limitations of traditional automatic translation systems. Tools like DeepL and Google Translate use Generative AI to provide high-quality translations.

Text summarization:

• Generative AI can summarize long texts accurately and concisely, preserving the most important information.

Chatbots and virtual assistants:

• Generative AI allows the creation of chatbots and virtual assistants capable of interacting with users in a natural and realistic way. The technology can respond to user questions, provide information and support, and complete simple tasks.

Generative Image AI: Painting with Technology

Generative image AI focuses on creating new and original images. Thanks to systems like Midjourney, Leonardo AI, DALL-E, or Adobe Firefly, it is possible to create new images and modify existing ones.

Image Creation

• Generation of realistic images of people, animals, objects, and landscapes that do not actually exist;
• Creation of characters, settings, and objects for use in video games;
• Creation of more engaging and informative images for use on websites and e-commerce.

Image Editing

• Removal of unwanted objects from a photo;
• Improvement of the quality of blurry or damaged images;
• Colorization of black and white images.

Generative Video AI: Painting in Motion with Technology

Generative video AI focuses on creating new and original videos. Unlike traditional video editing systems that rely on manual editing, systems like Synthesia, HeyGen, NVIDIA Canvas, or the new Sora learn from large amounts of data to create realistic and fantastic videos. As with image generation, with generative video AI it is possible to create and modify videos.

Video Creation

• Generation of realistic videos of people, animals, objects, and landscapes that do not actually exist.
• Development of video games and therefore for the creation of characters, settings, and objects for video games;
• Creation of videos for e-commerce and therefore of products for e-commerce websites, making it all more engaging and informative;
• Generation of Avatars for the presentation of specific products or interaction with users in real-time;
• Generative AI can be used to create realistic and fantastic special effects for movies and TV shows.

Source of video: cdn.openai.com

Video Editing

• Removal of unwanted objects from a video;
• Improvement of the quality of blurry or damaged videos;
• Colorization of black and white videos.

Generative Artificial Intelligence in support of programming

Generative AI is opening new frontiers in the world of programming, offering tools and technologies capable of:

Increasing productivity:

• Automatic code generation: Tools like Github Copilot and Cursor can automatically generate code based on text descriptions or existing code snippets;
• Automatic code completion: AI can suggest automatic completions for variables, functions, methods, and other code elements;
• Automatic error correction: AI can identify and correct syntax and logic errors in real-time.

Improving code quality:

• Code reformulation to make it more readable, efficient, and secure;
• Automatic refactoring: AI can help reorganize code to make it more modular and maintainable.

Promoting creativity:

• Exploration of new solutions and approaches to programming problems;
• Development of new algorithms and programming techniques;
• Creation of innovative software: AI can help create innovative and never-before-seen software.

AI and Ethics: An In-depth Analysis

Artificial intelligence (AI) has brought enormous advances in various sectors, from medicine to finance. However, the rapid development of this technology has also raised several ethical issues that deserve in-depth analysis.

Some of the main ethical concerns related to AI include:

• Bias and discrimination:

AI systems can reflect and amplify biases present in the data they have been trained on. This can lead to discrimination against certain categories of people, such as women, people of color, or people with disabilities.

Example: An AI system used for hiring personnel might discriminate against women based on their gender or their name.

• Transparency and interpretability:

Decisions made by AI systems can be opaque and difficult to interpret. This can make it difficult to understand how an AI system made a particular decision and can hinder accountability in case of errors or harm.

Example: An AI system used for medical diagnosis might provide an incorrect diagnosis without providing sufficient explanations as to why that diagnosis was reached.

• Privacy and security:

AI systems can collect and analyze large amounts of personal data, raising concerns about the privacy and security of such data.

Example: An AI system used for facial recognition might collect biometric data without the consent of the individuals.

• Impact on jobs:

The automation of repetitive tasks by AI systems could lead to job losses in some sectors.

Example: The use of automated robots in factories could lead to the dismissal of manual workers.

Addressing these ethical issues is crucial for the responsible and sustainable development of AI. It is important to establish clear ethical principles to guide the development and use of AI. These principles should include non-discrimination, transparency, accountability, privacy, and security.

The responsible development of AI requires a commitment from all stakeholders, including researchers, developers, companies, governments, and citizens. Only by working together can we ensure that AI is used for good rather than for harm.

A recent news (March 13, 2024) published by Wired is the approval by the European Commission of the final text of the AI Act (Artificial Intelligence Act), which guarantees safety and respect for fundamental rights and promotes innovation.

The main objectives of the AI Act are:

• Classification of artificial intelligence applications based on risk;
• Definition of categories of high-risk artificial intelligence systems;
• Introduction of bans on some artificial intelligence applications;
• Protecting citizens from risks arising from artificial intelligence;
• Stimulating innovation in this sector;
• Promoting a global regulatory framework for artificial intelligence that ensures the safety and transparency of artificial intelligence systems.

Thanks to the implementation of this regulation, artificial intelligence systems will be subject to checks by the competent national authorities, while the European Commission will have the power to supervise and coordinate the implementation of the AI Act.

Finally, expert committees will be established to provide technical advice to national authorities and the European Commission.

More information and the approved text can be found here.

Conclusions

Generative Artificial Intelligence (Generative AI) represents a promising branch of artificial intelligence with disruptive potential in various sectors, with the potential to transform our way of living and working by improving our lives in many ways.

However, it is important to be aware of the challenges and work together to develop and use this technology responsibly and sustainably, maximizing benefits and minimizing risks.

Author: Francesco Scalzo

L'articolo Generative AI, what it is and what are the main tools proviene da Revelis.

The role of Revelis

As stated in the past weeks, Revelis has formed an internal task-force to contribute to the development of Artificial Intelligence solutions against the COVID-19 outbreak. This is a joint work with researchers from the Italian Research Council ICAR-CNR(CS).

In this blog post, we want to show you a model to forecast new daily confirmed cases and ICU workload in Italy.

Our work takes inspiration from the paper by Arianna Agosto and Paolo Giudici, where a model for financial contagion is adapted to predict the daily new cases caused by COVID-19.

It is possible to summarize our contribution in two main points:

• providing a novel mathematical formulation.
• taking into account containment measures.

The Model

The model is a variant of the well-known Poisson Auto Regressive Model; we assume a Poisson distribution for the target variable and its intensity is estimated through a log-linear model containing:

• target variable values in the previous k days,
• containment measures in the previous k days,
• an estimate of the Poisson parameter form at the time instant t-k..

The model has been trained on official data from Italian Civil Protection from 24th of February until 8th of April. Prior to modelling, we have transformed the target variable using moving average of 3 days.

Outcomes

As time goes by, performances are subjected to deterioration and, for this reason, the model has been updated many times. The following models have been trained before:

• a first model using data until the 9th of April
• a second model using data until the 16th of April

For readiness purposes, only the last model trained using data until the 25th of April is shown. After breaking down the moving average values into point estimates for each day, the results are listed in the following table:

On a medium-term perspective, we predict that after the second week of May ICU workload will reach a value around 1k. Unfortunately, as you can observe in figure that follows, confidence intervals blows up at that time and therefore it’s risky for policy-makers to rely on these predictions.

Conclusions

Long-term predictions are not enough reliable because of the confidence interval width, this might have been expected because Poisson distribution has a variance equal to its expected value and, as times goes, uncertainty sums up. However, it is reasonable to use this model to perform a short-term forecasting and to get a rough idea of when containment measures will reach their maximum efficacy. It is crucial to say that this work has not been peer-reviewed and it has limits given by the official data reliability.

L'articolo Artificial Intelligence: applications against COVID-19 proviene da Revelis.

The whole world is paying a high price both in terms of human lifes and social distancing needed to slow the epidemic peak. In this scenario public institutions, companies and independent researchers are involved in the fight against COVID-19. There are many research lines, so that the White House launched an open call through a Kaggle Challenge trying to get some meaningful results.

Revelis couldn’t stand aside and that’s why an internal task-force was created for this goal. Computer Scientists, Data Scientists and Researchers are working hard to achieve tangible outcomes using Big Data Analysis techniques. This activity involves also some researchers from ICAR-CNR.

One of the initial tasks of this task force was to analyze the state-of-the-art, open-data and everything that could be used for Artificial Intelligence solutions against the virus.

We decided to share this list of useful resources in order to allow everyone to benefit for their own project against COVID-19. This list is not exhausting and that’s why we kindly invite you to contact us in order to further expand it, writing an e-mail to info@revelis.eu

Resources are classified into five different categories (“Dataset”, “Research Papers”, “Dashboard”, “Tutorial” and “Others”), to allow an easier navigation.

Dataset

A good Artificial Intelligence Solution needs a dataset that is at the same time big enough and of a good quality. We have collected different kind of data going from time series to images and text.

Worldwide epidemic time series

At the Humanitarian Data Exchange site you can find open data about Confirmed, Deaths and Recovered cases all over the World, starting from 01/22/2020. Data are daily updated and contain informations about each Nation and, eventually, for Province.

In these days many scientists, like Prof. Ioannidis from Stanford University, are arguing that we need better information to guide decisions. Because of that we are adding other data sources to the list to provide a broad spectrum of choices. You can obtain worldwide time series also from Johns Hopkins, from the European Centre for Diseases Prevention and Control where you can also find a short tutorial to upload data into an R environment and the R Package nCov2019 that allows you to have historical data and other functionalities.

Italy epidemic time series

Official data about COVID-19 in Italy are made available the Italian Civil Protection. Beyond confirmed, deaths and recovered, you can retrieve info about the number of hospitalized individuals, the number of people in ICU and others.

Radiological Images

COVID-19 diagnosis using Deep Learning is an hot topic. The following is a list of websites where it is possible to find this kind of data:

• X-Rays images of ~30k unique patients having pulmonary diseases
• X-Rays images of pneumonia affected pediatric patients
• X-Rays images of pneumonia affected patients
• X-Rays images of COVID and normal patients
• X-Rays and CT scan images
• X-Rays and CT scan images of patients affected by different kind of pneumonias (including COVID)

Textual data

The White House released a dataset consisting of 29,000 research papers related to COVID-19. The goal is to find meaningful answers to various questions such as risk factors, virus genetics and many others.

Demographic data

Geo-spatial data are sometimes used in the context of epidemiological studies. Italian demographic data are available with a high geographical density.

Social Networks Data

Online Social Networks Sites are a data source with daily tons of news shared by common people, public institutions and newspapers. The crowtangle website allows to monitor the Facebook Pages activity related to COVID-19 in Italy.

Flight connection data

The virus transmission analysis requires to identify the connections between different areas in the whole world.  Some researchers are working on worldwide flights to better understand how the virus shifted from China to the World. All the worldwide connections between airports with an estimate of the average trips between them are available online.

Previous epidemics data

There is a research line that tries to understand COVID-19 trying to compare it with previous epidemics. Here you can find links to data related to:

• Zika outbreak
• H1N1 outbreak
• Ebola outbreak
• SARS outbreak

Data from Korea

Korean Ministry of Health and Welfare built a web-platform to make available official data coming from South Korea. You won’t be allowed to download data locally, but you will have at disposal a cloud computing platform to submit your applications.

Ultrasound data

Italian researchers from University of Turin are currently developing an AI system for the automatic diagnosis of COVID-19 pneumonia from ultrasound images. The database is available upon request.

Data from UK government

From the official website of the UK government, it is possible to retrieve information about the number of hospitalized individuals, the daily new cases, the cumulative confirmed cases and deaths. Data are recorded at a province level.

Data from CDC (US)

CDC has launched a public surveillance on US Covid activity through a weekly report of hospitalized individuals, emergency department visits and other info. It is also possible to obtain daily data on deaths and cases from this page.

Mobility Data from Apple

Apple is publishing daily reports on the request for directions in Apple Maps for the whole world. This could be a good proxy for actual mobility. It is possible to observe data through dashboards or download the raw version at this link.

Data from Clinical Studies

The Operation Research and Analytics Lab from MIT has started a project related to Covid-19 analytics. Among their contributions, we need to remark the collection of datasets from various clinical studies.

Research Papers

Many research papers have been written in the last weeks, confirming that researchers are making their best effort to provide a substantial progress in the fight against COVID. These works can be an important guide and that’s why we want to include some papers that we consider as relevant.

Deep Learning applied to CT scan images

Chinese researchers have used about 5,000 CT scan chest images to detect infectedCOVID-19 patients with a high accuracy. The paper “Artificial Intelligence Distinguishes COVID-19 from Community Acquired Pneumonia on Chest CT” describes the use of Convolutional Neural Networks to extract meaningful features from images to classify a pneumonia either from COVID or not.

Poisson Autoregressive forecasting model

Trend prediction is a relevant information for decision-makers. An approach to Covid-19 trend analysis is described by Arianna Agosto and Paolo Giudici.

Non-Pharmaceutical Interventions (NPI) Epidemic model

During Covid-19 epidemic emergency many governments decided to take drastic actions such as school closures and social distancing. Ferguson et. Al. made a study of the efficacy of this kind of measures.

Case-Fatality-Rate unreported cases estimation

An open discussion is the one about data reliability. It is not always possible to screen all the population and, in this paper, a methodology to estimate the unreported cases is set out.

Deep Learning for Scientific Discovery

Researchers are trying to figure out solutions to a variety of problems related to COVID-19. Deep Learning algorithms are able to catch non-linear relationships between data and that’s why it is used in many different application. This paper is a Deep Learning survey that contains: a high-level explanation of Deep Learning theoretical concepts, a diverse set of Deep Learning techniques for different data modalities, a set of methods to improve performances when you have less data (semi-supervised methods), model interpretability techniques and tutorials on how to implement a complete Deep Learning pipeline.

Overview on recent AI solutions to Covid19

The epidemic outbreak have stimulated a lot of people to work towards providing new insights about the virus using Artificial Intelligence. This paper highlights many studies regarding molecular, medical and epidemiological applications giving also hints for future research on the subject.

Monitoring COVID contagion growth

The EU Think-Tank CEPS is producing reports, which are regularly updated, about the COVID-19 contagion. Two of their contributors are Arianna Agosto and Paolo Giudici, authors of the model we used as base to develop our own. 

Infections and NPI assessment in 11 European countries

In this study, researchers from Imperial College provides estimates about the true number of infectious, the reproduction number and the impact of Non-Pharmaceutical-Interventions. 

Dashboard

Dashboards are visual tools useful to represent complex information in a concise and dynamic way, with the aim to understand the epidemic outbreak in a short time.

Italian Civil Protection Dashboard

COVID-19 diffusion data are available by a dashboard from Italian Civil Protection.

Real-time epidemic simulation

Online you can find an epidemic simulation dashboard, where is possible to evaluate the virus diffusion by setting different parameters.

Epidemic Modelling for US

Researchers from Northeastern University, University of Florida and other institutions used a spatio-temporal epidemic model to estimate the number of deaths and infections in US. It is possible to observe their prediction at this link, in a single-page there are dashboards with both predictions and a “what-if” scenario analysis that simulates the contagion growth without containment measures.

Community mobility report from Google

Google decided to open-source their estimation of mobility in the World. You can find data about mobility in parks, transit stations and so on. The spatial-granularity is national and it becomes regional for wide areas.

Epidemic Projections by MIT 

MIT researchers employed a novel SEIR model, explained in the details at this link, to predict several quantities such as Deaths, Hospitalized Individuals and Confirmed Cases. Projections are available at a State level.

Tutorials

Tutorials describes are practical applications of Artificial Intelligence tools.

Deep Learning for X-Ray analysis tutorial

You can learn how to diagnose the COVID-19 by using Convolutional Neural Networks with Python and TensorFlow.

Epidemic models with R tutorial

This RMarkdown is useful to learn epidemic models by using R.

SEIR Modelling using R

In this tutorial, it is possible to find how to exploit a SEIR model taking into consideration the spatial dimension using an origin-destination matrix. In this case, the analysis is focused on the Greater Tokyo Area.

Resources

In the following you can find resources that could not be classified in previous categories.

R Package for time series

The tscount package is useful for time-series analysis with counters.

BERT-related resources

BERT is a deep neural network used in various Natural Language Processing tasks to achieve state-of-the-art results. In this Github repository, you can find various resources such as pre-trained architectures and research papers about different NLP solutions using BERT.

Research papers with code

It happens quite a few times that you discover an interesting paper but you don’t have time to implement from scratch the software. At this website, you can find a collection of papers with the associated code.

Data aggregators

Many organizations are sharing resources on-line. You can find other AU resources at the following links:

• Open-Source-COVID-19
• Dimensions.ai dataset
• Towards Data Science
• https://github.com/soroushchehresa/awesome-coronavirus

Last update: 2021/05/18

L'articolo Artificial Intelligence against COVID-19 proviene da Revelis.