Lingaro Data Science and AI Conferences

This lecture demonstrates how Dagster can be utilized to quickly produce machine learning pipelines and how it can be integrated with current tools like Databricks. It also shows how reducing time to production can reduce costs, and that modeling both tasks and data assets can reduce technical debt, especially when it comes to AI products.

Timely and accurate delivery of goods is a crucial competitive advantage. However, generating accurate demand forecasts and efficiently deploying solutions across markets and customers are difficult for multinational CPG companies. This presentation focuses on Lingaro’s approach to these challenges.  

In our lecture, we present a platform approach that offers a common data model, high code reusability, and increased efficiency and scalability to address these challenges.  

Large language models (LLMs) hold immense potential, but their outputs can be unreliable. This presentation focuses on the challenges of output quality and safety how to improve these through the use of guardrails.  

Here's an overview of the presentation: 

• Overview and definition of the problems concerning LLMs 

• Safety issues in generated content 

• Introduction to and definition of guardrails 

• Guardrail usage 

• Conclusions and real-world deployment recommendations 

Artificial intelligence (AI) has the potential to broaden its own capabilities by integrating different modalities, creating more engaging and interactive experiences. Multimodal generative AI (GenAI) models are capable of perceiving, understanding, and generating content using multiple modalities such as text, speech, image, video, and sound.  

Our experts showcase how their implementation can unlock new business value, enabling innovative applications and improving existing ones. They also discuss the challenges and opportunities that these models present to businesses. The experts also share Lingaro’s unique experience and the use cases that can be applied in real business situations such as content creation and knowledge management. 

This presentation explains multimodal generative AI (GenAI), which can perceive, understand, and generate content in various formats such as text, image, video, and sound. Our experts demonstrate its impact and new possibilities in creating innovative solutions. We also discuss its implementation challenges, use cases, and opportunities. 
 
Here's an overview of the presentation: 

• Introduction to and definition of multimodal GenAI  
• Overview of available solutions and GenAI techniques 
• Use cases based on projects delivered by Lingaro 
• Trends, latest developments, and future trajectory of multimodal GenAI 

This presentation provides a comprehensive review of critical considerations for utilizing foundational models within commercial use cases, with a focus on computer vision and natural language processing (NLP). It also outlines a systematic framework comprising essential steps for verification.  

The presentation illustrates the process through examples of evaluation protocols, offering practical insights into assessing model performance and applicability in real-world scenarios. The analysis will concern mainly generative models, particularly text-to-image synthesis and large language models (LLMs). Through this detailed exploration, participants will gain a deeper understanding of the strategic and technical prerequisites for utilizing foundational models to drive innovation and efficiency in commercial applications. 

In recent years, large language models (LLMs) have demonstrated their usefulness and impressive capabilities in language processing. However, we still do not fully understand how their inner workings allow them to achieve such great performance.  

This talk discusses the importance and motivation of using interpretability methods and gives an overview of selected methods. By exploring their effectiveness in demystifying black box models, the audience will learn about tools that can potentially provide valuable insights into the decision-making processes of LLMs. The presentation will mainly focus on methods that are applicable to models that are first pretrained on a large corpus of unlabeled text data and then fine-tuned on a set of labeled data from a specific downstream domain, such the BERT or RoBERTa model. 

Where is the exciting frontier of computer vision heading? What is the impact of infusing it with cloud-based deep learning models and large language models (LLMs)?

Discover how these models are shaping the future of AI, escalating scalability, slashing infrastructure costs, and hitting the bull's eye with precision.

Ever wondered about the dynamism behind electricity demand forecasting? How different forecasting models are when executed, like naive seasonal, regularized linear regression, gradient boosting, ARIMA (autoregressive integrated moving average) and Facebook Prophet? Dive into this session as we put these to the test, specifically for the Spanish grid.

In this presentation, we’ll explore:

• The vast potential of multimodal generative AI models capable of handling multiple modalities, such as text, speech, image, and video.
• Challenges and opportunities for enterprises, covering state-of-the-art methods and frameworks like AutoGen, Composable Diffusion (CoDi), and NExT-GPT.
• Valuable insights into real business use cases, including content creation and knowledge management.

With AI regulations in the forge across the globe, the race is on. Recent developments in generative AI have heightened concerns about the socio-technical implications of AI systems — a crucial consideration for applied AI.

In this presentation, we’ll discuss:

• The business perspective on legal and ethical AI.
• The balance of costs and benefits in responsible AI development and explainable AI.
• Microsoft's Responsible AI vs. impending European Union regulations.

This holistic review coupled with case studies provides insights into the workings and applications of responsible AI in large organizations.

This presentation dives into the latest advancements in large language models (LLMs) and generative AI, and their impact on commercial practices. Continuing the exploration of generative AI models in business applications that began at DSS ML ’23, this presentation provides an update on recent changes in the field.

Key focus areas include:

• Successful implementation strategies for enterprises.
• Effective business models utilizing LLMs and generative AI.
• Updated industry hype curve and industries likely to be profoundly affected by generative AI and LLMs.

In this engaging talk, we will:

• Discuss challenges faced in establishing AI solution delivery frameworks.
• Explore the complexity of machine learning (ML) software, which goes beyond data management to incorporate business logic.
• The importance of interdisciplinary approaches and dynamic data governance for data science practices.
• Modifying tasks of the CRISP-DM to bridge gaps in ML operationalization using our proposed AI Delivery Framework.
• Showcase the AI Delivery Framework as a tool for AI project management and technical leadership.

As natural language processing (NLP) encounters ever-increasing complexity, applying computer vision-inspired (CV) strategies has also become more recognized. How can the attention mechanisms of BERT be harnessed to expand the horizons of data augmentation in NLP?

Key insights from this presentation include:

• Understanding data augmentation for NLP through mixed training samples.
• Application of CV techniques to NLP, insights into BERT’s role, novel applications in data augmentation, and visual examples of augmented outcomes.
• Empirical evaluations of the proposed techniques' effectiveness.

How can embracing structured coding standards foster an environment of innovation without chaos, and set the stage for data science initiatives to really excel?

In this presentation, Konrad will share:

• Structured coding standards and best practices in data science that reinforce innovation and efficiency.
• Overlooked aspects of the software creation process and practical solutions that address issues such as code inconsistency and poor readability.
• How strategies like linting, rigorous testing, and continuous integration foster team cohesion.
• Success stories highlighting the transformational impact of standardized coding practices on team productivity.

This presentation focuses on managing machine learning models in production and assessing new models after training. Key concepts include data drift, concept drift, and lack of generalization, which are demonstrated through a practical project use case.

This use case showcases the integration of Databricks with MLflow, introducing MLflow’s features such as model registry, models versioning, experiments, and deployments.

Lingaro's NLP and generative AI experts, delve into the practical application of image generative models in business.

In this engaging talk, they’ll unpack:

• Two business applications of generative models: concept generation for ads and PowerPoint slide generation based on varied format inputs.
• Explanation on what made these innovative projects successful.

Lingaro expert delivers a detailed technical presentation on modern data augmentation techniques, shedding light on:

• Notation, methods covered, and the universal mixing equation.
• Canonical method: Mixup.
• Canonical method: CutMix.
• Empirical evaluation and method similarities. Application to other tasks and modalities.

Take a closer look at a real-life case study on developing an efficient framework to put many-model ML solutions into action, utilizing the best practices in MLOps.

Join our expert to get valuable insights on how to leverage MLOps and effectively overcome these hurdles. Experience the journey of designing and building a pioneering framework that fuses MLOps best practices with Microsoft's Azure Machine Learning.

This session will help you understand how a versatile, user-friendly, and production-ready base for many-models machine learning can simplify the complex web of machine learning production.

Is AI really maturing at 57? How can this question be addressed from both a global standpoint and an organizational perspective?

This thought-provoking discussion exposes the gap between AI's scientific achievements and practical application.

This session presents an overview of:

• Automatic image generation for a defined group of clients.
  The Fluence project, automatic image generation, innovative framework, and generation examples.

• Fine-tuning of generative model, in embedding space, and with unfrozen weights.

Lingaro expert delivers a detailed technical presentation on modern data augmentation techniques, shedding light on:

• Notation, methods covered, and the universal mixing equation.
• Canonical method: Mixup.
• Canonical method: CutMix.
• Empirical evaluation and method similarities.
• Application to other tasks and modalities.

This session provides an overview of:

• Practical examples of the voice assistance.
• Experiment on open-source library Merlin.
• Experiment on FastSpeech 2 implementation.
• Experiment on Tacotron 2 implementation.
• MUSHRA (Multiple Stimuli with Hidden Reference and Anchor) and comparing the results of the experiments.