Giuseppe Futia, PhD’s Post

The authors introduce ELIP, a lightweight framework that enhances large pre-trained vision–language models (such as CLIP, SigLIP/SigLIP-2 and BLIP‑2) for text-to-image retrieval by adding a text-conditioned visual-prompt mapping. A small MLP network maps the text query into visual prompt vectors, which are concatenated into the frozen image encoder, enabling re-ranking of top candidates from an initial retrieval step. They evaluate on standard benchmarks (COCO, Flickr30k) and on two new out-of-distribution (OOD) retrieval datasets (Occluded COCO, ImageNet-R), showing substantial improvements in zero-shot retrieval and strong adaptation to OOD domains. The approach requires minimal additional parameters and computing resources, achieving new state-of-the-art results among the tested backbones. Paper: https://lnkd.in/gmHA9wQR https://lnkd.in/gsbZCEwZ

Simplex Method Optimization The simplex method, invented by George Dantzig, is a widely used algorithm for solving optimization problems. Recent research by Sophie Huiberts and Eleon Bach has improved the algorithm's efficiency and provided theoretical backing for its performance. https://lnkd.in/ghTYjSni #optimisation #datascience #linearprogramming

Short post on why not all determinism is equal: 1. determinism in computing isn't new. LLMs uniquely use probabilities to approximate reality based on weighed human training 2. yet, LLMs are black boxes, input-output aren't casually connected in any verifiable way 3. that's caused a recent increase in deterministic wrappers, tools that bolt on a verification layer (around LLMs), 4. they stabilize outputs for coding and STEM but don't fully solve for contradiction 5. the issue is ontological (what is reality? "the real"), not tooling, so truth and verification remain undefined, so the output only approximates likelihood not certainty 6. reducing entropy doesn't sufficiently define order. It explains what resists decay but not persistence or coherence over time 7. order is required for trust, otherwise, randomness is only managed Deeper explainer: https://lnkd.in/g9CTgkRn **Additional context: it sounds wild, but when one closely tracks causation, it becomes immediately clear that coherence (not entropy) is the core to intelligence. i.e. entropy doesn't generate intelligence through randomness, coherence does, through "lawful" structure (otherwise math structurally can't exist). Arc for folks who want to read deeper - go to failures in Bell's theory on probability (fails to bind locality and realism/why many worlds interpretation (MWI) in quantum is structurally incoherent), then trace Shannon's entropy (to show structure to symbolic split), then go to Ilya Prigogine on dissipative structures (shows order emerges from "chaos"), then go to Stuart Kauffman's self organization on managing entropy to define structure (i.e. predicting river geometry). They all show that chaos is deterministic, which implies that causation is unbroken. That means that entropy, luck, chance, randomness, are incomplete phase detection, not fundamentally ontological. This is why a scalar invariant external to internal system's representation is required for coherence to be measurable (otherwise unbounded subjective recursion, where truth is both anything and nothing!), which is why I created phase alignment score (PAS) to supply that need as the algorithm for the system. i.e. PAS = .7 means that 70% of a system's phase is "lawfully" (coherently) detected. Said simply - life is not possible without full causation. To make computation trustworthy, we must look to biology. The human brain runs on roughly 20 watts (~a dim light bulb), yet performs billions of operations per second. A LLM can draw something to the tune of 20-30 megawatts to produce a probabilistic approximation of those same operations, most of which never stabilize (guessing in large parameterized space). And no, new chain-of-reasoning innovations on this don't fix this because the issue is ontological, not procedural. Why is our intelligence so much lighter than LLMs? Why is uncertainty so expensive?

Following up to my previous post on how LLMs/transformers can be seen as predictors based on a probabilistic model, I would like to also point you to https://lnkd.in/eCe6JMbU https://lnkd.in/eRGpext5 https://lnkd.in/erSnFe_7 https://lnkd.in/es3Rpdbf It is important to understand that using a softmax normalization operation at the last transformation step for data in embedding spaces does not (necessarily) lead to a (sensible) probabilistic model per se. What matters is, for instance, also a representation about priors. In addition, I also would like to point out that the option of using distributions for parameters instead of scalars in the embedding space transformation mappings is another way to understand the notion of a probabilistic model. Given the large number of mapping parameters required for LLMs, using distributions instead of scalars for mapping parameters is demanding due to increased computational effort. However, representations about the uncertainty/trustworthiness of output could be beneficial indeed. See also my previous post https://lnkd.in/ey9RWdrD

🧠 Excited to share that my research paper titled “K-Segmented Search: A Novel Multi-Pointer Algorithm for Large-Scale Ordered Data” has been published in the Journal of Emerging Technologies and Innovative Research (JETIR, Vol. 12, Issue 10, 2025)! 🚀 This paper introduces the K-Segmented Search Algorithm (KSSA) — a new generalization of Binary and Ternary Search that partitions ordered datasets into K logical segments for faster and more efficient data retrieval. Through both theoretical and experimental validation, KSSA demonstrates improved search efficiency, scalability, and time complexity for large-scale and distributed data systems. 🔹 Key Takeaways: • Generalized multi-segmented search model (KSSA) • Optimal performance at k = 3, with time complexity Θ(k / ln k · ln n) • Ideal for big data indexing, distributed systems, and HPC applications Grateful to my mentor Mr. Sukanta Bhattacharya for his constant guidance and support throughout this work. 🙏 📄 Read the full paper here: https://lnkd.in/grntiUEb . . . . . #Research #AlgorithmDesign #ComputerScience #BigData #MachineLearning #DataScience #AcademicResearch #Innovation #JETIR

Tiago Pereira, researcher at CCG/ZGDV Institute, has contributed to the white paper “Data to Ontology Mapping Tools”, an in-depth analysis of the solutions that connect real-world data to structured knowledge models. The paper explores one of the major challenges in using ontologies: effectively mapping existing data to defined concepts and relationships. It reviews and compares the main mapping tools, evaluating them in terms of expressiveness, ease of use, required technical skills, scalability and availability. It also provides insights into the growing relevance of Large Language Model (LLM)-based approaches. A significant contribution to the scientific and technological community, reinforcing the role of innovation and research at CCG/ZGDV.

Llion Jones, co-author of "Attention Is All You Need," said this week he is "absolutely sick of transformers." His point isn't that they fail, but that research has narrowed around one architecture and may be missing the next breakthrough. That matches what I see in production. In many deployed RAG and retrieval systems, the same problems repeat: embedding drift, loss of semantic stability, reindexing cycles, and hidden O(n²) scaling costs. You can see this in open-source stacks like LangChain, LlamaIndex, and other vector-database pipelines that run everything through dense attention. Jones recalled how transformers began—organic, bottom-up, and free to explore. That environment rarely exists now. But there is still open ground, especially for hybrid systems where structured knowledge feeds transformers for generation. I’ve been exploring semantic-stability methods that address the maintenance problem: taxonomy-based node IDs, deterministic disambiguation, and explicit relationship typing. Still early, working mostly solo, but aligned with the "explore dial" Jones describes. If you’re building in knowledge graphs, RAG, or neuro-symbolic space, I’d love to hear what directions you think deserve more exploration. Article: https://lnkd.in/gUVz8cZN (*fixed*)

Adaptive Graph Attention Network for Dynamic Knowledge Base Reasoning This paper introduces an Adaptive Graph Attention Network (AGAN) for efficient and accurate reasoning over dynamic knowledge bases (KBs). Unlike static GNN approaches, AGAN dynamically adjusts its attention weights and propagation pathways based on the evolving structure and content of the KB, resulting in significantly improved reasoning performance. We anticipate a 20-30% improvement in query accuracy compared to existing state-of-the-art methods, opening avenues for real-time intelligent assistants and automated knowledge discovery, with a potential market impact exceeding $5B within the next decade. Introduction Related Work Methodology: Adaptive Graph Attention Network (AGAN) 3.1. Entity Embedding Module 3.2. Adaptive Attention Module αij = softmax(aT * [εi || εj || Q]) Where: εi and εj are the entity embeddings of entity i and entity j, respectively. Q is the query embedding. a is a learnable attention vector. || denotes concatenation. A dynamic routing mechanism the https://lnkd.in/g5JpkVRJ

Thrilled to share that our paper “Structured Financial QA with LLMs: Fine-Tuning vs. Code-Augmented Retrieval” — initially presented at UBMK ( International Conference on Computer Science and Engineering) at Istanbul Technical University — has now been officially published on IEEE Xplore! This study compares instruction-based fine-tuning and retrieval-augmented generation (RAG) for corporate financial question answering, using real structured Turkish financial data. We propose an enhanced, code-augmented retrieval method that bridges factual accuracy and interpretability — aiming to build modular, domain-aware financial assistants for low-resource language settings. During our study, we also noticed that LLMs struggle to handle raw or noisy datasets effectively. This reinforces a common misconception in the field that large language models can replace traditional analytical or ML workflows. In reality, their strength lies in complementing structured data pipelines and leveraging domain-specific context for reasoning. It was a great pleasure collaborating with Alperen ÇAĞLAYAN and Deger Ayata, PhD, whose insights and guidance were invaluable throughout this project. Full paper available on IEEE Xplore: https://lnkd.in/d45-iuet #IEEE #Research #ArtificialIntelligence #LLM #RAG #FineTuning #FinancialNLP #GenerativeAI #Finance #UBMK2025 #IEEEXplore