How to kill any AI project – 8 easy and powerful ideas – Part III
Huh – you read part one (data dilemma, software engineering challenges, and privacy matters) and part two (change, complexity, and cost of AI) of our #AIAdventures series and are hungry for more? Well, we bet you rather fancy the solutions over the problems. So, before wrapping this series up, we have some more:
- Multimodality – We have more than just [insert any type of data here] and only when combining them, it is possible to draw insights.
- Explainable AI – Getting an answer is not enough, it has to be justified and traceable.
And there’s a rather new obstacle on the field that you wouldn’t want to miss: The 2021 Proposal for a Regulation of the European Parliament and of the Council laying down harmonised rules on Artificial Intelligence (Artificial Intelligence Act) and Amending certain Union legislative acts – or short: the AI Act. We’ll touch it briefly in the end but it is a long story for another time 😉.
So let’s dive in!
1. Diverse data types: beyond text to tables, structured data, figures, and videos
When it comes to unstructured data, the challenges posed by diverse file formats and information scattered across different systems are a longstanding issue for many organizations. Traditional solutions often fall short, focusing on specific data types and demanding heavy investments for setup, leaving users to start anew (i.e. start the training process anew) when adding new information. It’s even worse if you have to integrate data in various formats to create a unified source of knowledge (try adding a video to a PDF archive).
The heart of semantha lies in its ability to generate a universal representation of information based on its meaning – called the “fingerprint.” This fingerprint is not tied to any specific file format, allowing users to connect disparate data sources seamlessly. Crucially, semantha preserves the connection to the source files, facilitating easy traceability for users back to the original data. When it comes to document analysis, semantha excels in automatically slicing and dicing documents, recreating document structure, and linking artifacts such as tables and figures to their relevant sections. The resulting fingerprints (they come in various shapes and sizes and represent sentences, sections, entire files, etc.) remain agnostic to the source, providing a unified understanding of the information represented. And since the fingerprints have standardized format – no matter the data source – semantha can perform analyses cross-modal and search your multi-modal knowledge base.
semantha integrates into existing systems through its REST API and leaves the process leading to external systems. Whether it’s an established document management system or a custom-built middleware, you’d leverage external processes to feed content to semantha and trigger analyses. This approach also means that you do not open a new data silo but rather point to existing systems and storage.
Real-world applications showcase semantha’s versatility. In requirements analysis, it easily ingests the nuanced structure of requirements, detecting differences across versions, and ensuring critical information is identified, irrespective of its placement within a document.
In knowledge searches, including in videos, semantha quickly pinpoints information within multimedia content and documents alike, streamlining the process of finding specific details in a sea of documents.
What sets semantha apart is not just its current capabilities, but its future-proof design. The ingestion modules can be effortlessly extended for unforeseen formats, new AI techniques, or additional sources. There’s no need for extensive training or re-training, making it easy to adapt the system over time. semantha’s unique blend of versatility, adaptability, and ease of use positions it as a frontrunner in the realm of AI-driven unstructured data analysis.
2. We need explainable AI
In AI projects, project leads often find themselves bridging the gap between technological innovation and the need for decision makers to understand and trust the solutions proposed. They have to convince decision makers that not only can an AI system provide accurate results, but its decision-making process is transparent and comprehensible/auditable. This need for explainability is crucial, as complexity often breeds skepticism, eroding the trust necessary for green-lighting a project.
This is where semantha not only delivers results but also excels in explaining how it got to these results. One of its key features is the library. Answers point back to entries in the library that led to a specific answer. This transparency is heightened with the integration of a knowledge graph, effectively modeling the problem domain and enhancing both problem-solving capabilities and explainability.
The pitfalls of complex AI models, for example Large Language Models (LLMs), are evident in niche use cases where plausible-sounding answers may be entirely incorrect. semantha, on the other hand, takes a different approach. It will tell you when it doesn’t know the answer. With every answer rooted in the library and knowledge graph, the system ensures that the provided solutions are not mere “plausible sounding” but robustly grounded in available information.
Beyond these assurances, semantha offers a unique approach to maintenance and adaptability. Unlike traditional AI models that require extensive retraining when updating knowledge bases, semantha generates a fingerprint from input data and stores it with metadata. Fingerprinting eliminates the need for recurrent training: Whether adding, updating, or removing information from the library, semantha ensures real-time adaptability without compromising efficiency. The collaborative nature of semantha further addresses concerns related to organizational divisions. Project leads can set up and share libraries across different divisions, fostering knowledge-sharing and uniformity. At the same time, they can create isolated library copies to empower teams to evolve independently, tailoring solutions to specific needs without jeopardizing the integrity of the overarching AI infrastructure.
3. We have to wait for the AI Act to be final
Bad news: Laws are rarely final. And it won’t be the last step of regulation either. So let’s quickly see what awaits us:
A new EU law regarding the creation, deployment, and use of artificial intelligence is waiting in the wings. There are things which are forbidden, but – coming from a European background – we believe that you don’t look for AI support in these areas (that would be analyzing biometrics and emotions, automated decision making for the well-being of people etc.). As this post is in the making, there are still things to finalize and the EU is still preparing the application of the AI Act; we’re not yet sure how this will turn out. We’re pretty clear, but you never know… Ah, and decision makers hate uncertainty. So, what to do about that?
Project managers leading AI initiatives should take a different perspective: They can transform regulatory challenges into catalysts for success. Let’s draw parallels to privacy regulations like the GDPR: a proactive attitude can become the cornerstone for kicking off projects and driving them forward: As a project manager you can strategically prepare for the (necessary and justified) questions coming from the regulatory (non-functional) requirements. NFRs are not new, they are an inherent feature of any software project anyways. So get yourself a comprehensive understanding of the regulatory landscape, decipher the nuances of risk assessments (ask for help, there are experts on that!), and get up to speed to the “new” compliance criteria.
By anticipating questions related to the AI Act and similar regulations, you can be an architect of transparency and governance. NFRs can be reframed as the pillars upon which project success rests. Instead of viewing these requirements as impediments, you can approach them as valuable tools to improve robustness and reliability of AI systems. A focus on documentation, conformity checks, and risk mitigation strategies not only ensures compliance but also underpins your commitment to delivering high-quality AI solutions.
In summary, the (two-way!) influence between AI projects and regulatory frameworks need not be a show stopper. Rather, it’s a very navigable terrain if you take on proactive strategies. Yes, there’ll be documentation involved and scrutiny with suppliers becomes more important. But joining forces with industry leads and expert providers, you’re more than prepared.
Get started with semantha today
In our series we covered a lot – semantha is an enterprise-ready AI platform that addresses the challenges of unstructured data. It comes as a versatile cloud platform but can also be deployed on-premises and it’s easy to get it up and running. Today, we discussed how semantha can draw information and gain insights from various data sources and from various formats, be it text documents or videos. And semantha provides you with full traceability from your request to your company’s knowledge base ensuring audit-readiness.
In one of our next articles, we will discuss the implications of the EU’s AI Act on using semantha and shed some light on the resources that we provide to our customers to ensure safe and compliant adoption of AI in their processes. If you want to learn more and get to know details for your specific situation, don’t hesitate and schedule a call today. We look forward to outlining a customized approach for your use case.