Prof Bill Buchanan OBE FRSE’s Post

Quote article "new proofs hows that for the quintessential path-finding problem, one algorithm is close to ideal: Assuming worst-case traffic patterns, it’s the best approach on every possible street grid. What’s more, the algorithm is nearly 70 years old and a staple of the undergraduate computer science curriculum. The new work will be presented with a best-paper award at the 2024 Symposium on Foundations of Computer Science next week." #AI #tigergraph #neo4j #machinelearning

In CS224n this quarter, my team and I experimented with a new approach to retrieval: GRAFT (Graph Retrieval Augmented Fine-Tuning). We sought to address the limitations of naive Retrieval Augmented Generation (RAG) when answering multi-hop global questions that require pulling information from multiple chunks of text. Traditional RAG can struggle with identifying relevant and interdependent information spread across various sources. For instance, consider a question like, "What are the main themes explored in Taylor Swift's album 'The Tortured Poets Department,' and how do they reflect her personal experiences?" This type of question requires integrating details about themes, personal experiences, and specific tracks across different parts of the document. Naive RAG often retrieves and generates responses that miss the nuanced connections between these chunks. Our solution, GRAFT, leverages the structure of a knowledge graph to capture both local and global relationships within the data. By constructing a graph of entities and their relationships, we enable more accurate and contextually relevant retrievals. We're excited to continue exploring this approach and are wrapping up experiments! More updates to come :)

For those of us in the wayfinding systems business, have a look at this article.

New to this edition are articles emphasizing the importance of diversity in research, not only cultural diversity among study participants but also in methodology (including quantitative, qualitative, and mixed methods); the role of big data; using technology such as artificial intelligence and apps; and strategies to improve transparency and accessibility, including open science practices, replication, and preregistration. Learn more and save 15% + free domestic shipping using code APABOOKS15: https://bit.ly/3MeiDO7

Highly recommended. Focused on algorithmic computation, with secondary exploration of second, third, fourth, and fifth level decision-making. https://a.co/9Kl5E5X

🌲Advanced RAG Series: Retrieval This is the fourth post in a series focused on retrieval by @divyanshu_van It dives into different retrieval algorithms that can be part of your RAG system Check it out: https://lnkd.in/dsp5JSUP

Turftopic (0.9.0) now comes with an implementation of Dynamic Semantic Signal Separation, meaning that you can now study the fluctuation of underlying semantic axes over time in a corpus. I have found it immensely useful for my work, and it has so far managed to uncover some patterns in corpora, I haven't seen other models do. What's also really nice about S³ in a dynamic context, is that, since it can describe both ends of an axis, you can see what the "opposite" of a given topic is in discourse over time. You can check this plot out interactively in our documentation (https://lnkd.in/d2UvrzFT)

Working on a project, I was wondering : what open licences are used for AI systems? Looking to the 604,464 models on Hugging Face https://lnkd.in/ekPtvGwt, as expected (it does not mean I agree with) most licences are software licences (209,000). For example, French Mistral 7b and Lighton's Alfred 40b are licensed under Apache 2.0, the most chosen from far. There are around 17,000 Creative Commons (a range of licences for open data with gradual exceptions), with twenty-two ODbLs (an open database licence popularized through OpenStreetMap) and eleven Etalab - Service du Premier ministre licences, including eight from IGN (Institut national de l'information géographique et forestière) and one from Direction interministérielle du numérique (DINUM). To be continued...

I just deployed a new DeepLLM app that reviews or summarizes technical papers, try it out at: https://lnkd.in/gUgk2gbh or fetch it from https://lnkd.in/gSvxXRbB . Here is a screenshot of its output on a nice paper by Blum & Blum about an analog of consciousness for computational processes.

I am pleased to share this recently published work, on the study of explainable intelligent systems that we carried out with José N. Paredes Maria Vanina Martinez Gerardo Simari In particular, we present the implementation of a recently proposed application framework (and make available its source code) for developing interpretable and explicable intelligent systems, and explore user-centered mechanisms for building explanations based both on the kinds of explanations required (such as counterfactual, contextual, etc.) and the inputs used for building them. In order to validate our approach, we develop two use cases, one as a running example for detecting hate speech in social platforms and the other as an extension that also contemplates cyberbullying scenarios. We share the free access link: https://lnkd.in/dDPkmmr9