Amir Tasbihi

I and Amin Ahmad recently appeared on The Ravit Show with Ravit Jain! We discussed the crucial role of reliability, transparency, and accountability in creating trustworthy #AI solutions. Catch the episode to learn more about how we build AI that people can trust! https://gag.gl/CEE9QC

34

2 Comments

𝐄𝐠𝐨-𝐜𝐞𝐧𝐭𝐫𝐢𝐜 𝐯𝐢𝐝𝐞𝐨 𝐚𝐧𝐧𝐨𝐭𝐚𝐭𝐢𝐨𝐧 Knowing where your body from an ego-centric video is an important research problem that is still not yet solved ⇨ SAM2 ability to identify object boundary & propagate single click frame annotation is for such use case surprisingly good! ➤ Video object segmentation is getting its own little revolution (the age of rotoscoping does not seem so far away)... #projectaria #egocentricvision #sam2

42

1 Comment

🚀 Responsible Retrieval Augmented Generation (ResRAG) Last Part! 🌟 Dive into the world of ResRAG Operations: Stack and Pipelines! 📘 Key highlights of ResRAG Ops Stack and Components include: - Data Extraction - Ontology Design - Vector Database - Graph Database - LLM Orchestration - GraphRAG Orchestration Explore the ResRAG Ops Creation Pipeline: - Ontology Design - Data Collection - Entity Extraction - Relationship Extraction - Graph Construction - Graph Enrichment - Graph Security - Graph Vectorisation - Graph Storage Discover the ResRAG Ops Serving Pipeline: - User Query Understanding - Privacy Check - Domain Detection - Access Check - Graph Retrieval - Response Generation Next, I will go deeper into explaining the different aspect of ResRAG. #AI #MachineLearning #DataScience #ResponsibleAI #SemanticWeb #KnowledgeGraph #neo4j Neo4j PoolParty Semantic Suite WhyHow.AI Gerald Dipplinger, Andreas Hladky, Kristian Landegren

27

3 Comments

The 2024 Stanford AI Index Report was out a while ago.. a few fun or not-fun revelations/interpretations depending on where you sit : The training costs of state-of-the-art AI models have reached unprecedented levels. OpenAI's GPT-4 used an estimated $78 million worth of compute to train, while Google's Gemini Ultra cost $191 million.. a staggering increase in the computational resources and financial investment required to develop the latest AI models.. Only the beginning, huh.. what say!? Despite AI surpassing human performance on many benchmarks, it still lags behind on more complex tasks like competition-level mathematics, visual commonsense reasoning and planning.. suggests AI still has significant limitations compared to human intelligence.. Only a matter of time, huh.. what say!? The report reveals a significant lack of standardization in responsible AI reporting across leading developers like OpenAI, Google, and Anthropic.. cannot systematically compare the risks and limitations of top AI models, plus fear mongering by top tech companies to keep industry hostage thanks to fearful and old world risk averse leaders calling the shots is at an all time high.. Only job left for human beings then is to call the bluff if they can - call the bluff, huh.. what say!? Notably, the U.S. dominated the development of notable AI models in 2023, producing 61 compared to 21 from the EU and 15 from China.. only metadata here guys, money does what money can huh.. what say!? Industry is dominating frontier AI research, producing 51 notable machine learning models in 2023 compared to only 15 from academia. Industry-academia collaborations also reached a new high. Investment in generative AI technologies surged to $25.2 billion in 2023, despite an overall decline in AI private investment.. Only 25.2B eh, maybe they got the currency wrong on that one huh.. what say!? & closer to heart, AI is transforming FMCG operations and innovation. The report highlights how leading FMCG firms like Unilever, P&G, and Nestlé are leveraging AI to enhance demand forecasting, inventory management, supply chain optimization, and product development. AI is driving significant productivity gains and quality improvements for FMCG workers. Studies cited in the report show AI enables FMCG employees to complete tasks 26-73% faster and improve output quality by up to 40%... only 1 look at the AI Products’ marketplace & that adds up huh, what say Shreyasi Singh Mriganshi Sharma Komal Falodiya Sweta A. Rekha Lath Abhilasha Sinha Quratulain Naseer Darshika Tripathi Krishna Soni S P Ravishankar Vijay Reddy Gopi Krishna M Neelesh Singh Rajpurohit Munna Kumar

25

3 Comments

Wow! 🙏 to Fused app builder (https://lnkd.in/gYtEUKHM powered by streamlit) and some LLM coding support, here is my "Cube Factory" fused app for effortless GeoML visualization and data curation. Goal is to give time back to Data Scientists to do actual research rather than low level data engineering. (Funny to think < 3 weeks ago I had never touched a front end app before, nor even thought myself having the capability....) Combine resources like Sentinel 1 SAR imagery, 13 bands of Sentinel 2, Digital Elevation, Scene Classification and Cropland Data Layers sandwiched into datacubes for ML ingestion. Choose your survey period and region of interest graphically by directly drawing on the map. 🚀 🌎 #machinelearning #geospatial #datascience #earthobservation #transformers

11

1 Comment

🚀 Have you heard of #GSoC? Especially if you're a college student with a programming background, you should know about this incredible opportunity! 📢 Google Summer of Code (#GSoC) - a global program by Google that lets students to contribute to open-source projects from various organizations and gain invaluable experience, while earning a generous stipend. It's a matter of pride and a standout addition to your resume. 💯 🧑‍💻 Organizations and mentors participating in GSoC guide the students to complete projects, within a stipulated timeframe, ensuring quality and expertise. 📢 For #GSoC2024, the acceptance of such contributors was announced just a couple of days ago. The post below is about the same from the organization called DBpedia Association, which works in the areas of #SemanticWeb, #KnowledgeGraphs, and #NeuralExtractionsFromText. 🤝 As part of the DBpedia team, along with Tom Soru, Ziwei XU and I will be mentoring Abdulsobur Oyewale on the project titled "Towards A Neural Extraction Framework" (https://lnkd.in/d2fJeXY8). 💡 I'm sure it will be a fun project to be associated with and also a learning/brainstorming opportunity to upgrade myself with the latest trends in #InformationExtraction. #OpenSource #Coding #StudentOpportunity #TechMentorship #CareerGrowth #mvp #mvpbuzz GDG Pune GDG Cloud Pune GDSC GDSC WoW #opensource

10

1 Comment

OpenAI’s o1 model, codenamed “Strawberry,” marks a significant leap in AI reasoning capabilities. Here’s why it’s a game-changer: 1️⃣ Enhanced Problem-Solving Power: Trained with a new optimization algorithm, o1 can break down complex, multi-step tasks—like coding and math—achieving 83% accuracy on Olympiad-level math problems, surpassing GPT-4o’s 13%. 2️⃣ Higher Costs & Slower Speeds: o1’s reasoning abilities come at a premium—4x more expensive and slower than GPT-4o, making it best suited for intricate queries. 3️⃣ Human-Like Reasoning: With a “chain of thought” approach, o1 processes queries step-by-step, mimicking human-like thinking, though it sometimes overdoes it on simpler tasks. #OpenAI #GenerativeAI #AIReasoning

33

📢 Introducing Mixllama3-8x8b-Instruct-v0.1 - an experimental MoE (Mixture of Experts) Large Language Model (LLM) based on AI at Meta's LLaMA-3-8B. MixLLaMA3-8x8b is a combination of 8 fine-tuned LLaMA 8B models, each specializing in a specific set of tasks. By leveraging the strengths of each expert model, Mixllama3-8x8b aims to deliver enhanced performance and adaptability across a wide range of applications. Check out this model at Hugging Face: https://lnkd.in/eqeVeWMf #machinelearning #datascience #largelanguagemodels #AI #deeplearning #huggingface #MixtureofExperts #LLM #llama #generativeai

25

3 Comments

RIG: Enhancing AI Accuracy with Real-Time Data Retrieval Google's new DataGemma model leverages RIG (retrieval interleaved generation) to reduce hallucinations and improve the accuracy of AI-generated responses. RIG integrates large language models (LLMs) with Data Commons, an open-source database of public data from reputable sources like the UN, WHO, CDC, and Census Bureau. Key features of RIG: • Pauses response generation to search for current or specific data from reliable external sources when needed • Seamlessly incorporates newly acquired data into the response • Adapts in real-time while generating responses, refining output iteratively • Addresses all aspects of complex queries through multiple retrievals at various stages Compared to RAG (retrieval augmented generation), which performs retrieval once before generating the answer, RIG offers a more dynamic approach. However, RIG may be slower in performance due to multiple data retrieval points. Evaluation of the RIG approach shows significant improvements in factual accuracy, with the DataGemma-RIG-27B-IT model performing particularly well by using Data Commons to dynamically fact-check and validate statistical information during response generation. RIG has the potential to solve critical problems faced by AI language models, including hallucinations and response quality, making them more reliable and trustworthy for real-world applications. #AI #MachineLearning #DataScience #NLP #GoogleAI #DataCommons #Hallucinations #FactualAccuracy

3

🚀 Unlocking AI with Vector Databases & RAG 🌐 🔹 Vector Databases: A cutting-edge solution for efficient data retrieval, storing data as vectors for fast, accurate searches. 🔹 Embeddings & Dimensions: Transform data points into vectors with multiple dimensions, capturing rich semantic features for nuanced comparisons. 🔹 Similarity Metrics: Utilize cosine similarity (or alternatives like Euclidean distance and dot product) to measure how closely data points align. 🔹 Retrieval-Augmented Generation (RAG): Boost AI performance by retrieving relevant context before generating responses, leading to smarter, context-aware interactions. 🔹 LLMs (Large Language Models): Powerful AI models that understand and generate human-like text, driving innovation in natural language processing and beyond. 💡 As AI continues to evolve, vector databases and RAG are crucial for extracting valuable insights from massive datasets. The future of AI is here, and it’s exciting! Feel free to play - https://lnkd.in/gSiaUJQG #AI #MachineLearning #VectorDatabases #Embeddings #RAG #LLMs #NLP #DataScience #Innovation #Tech

26

4 Comments

Alright, I am back with my next AI Generated Video. This time, I used LTX Studio from Lightricks. The tool was very cool to use and allowed me to go truly multi-modal by incorporating video, voice-overs, and sound effects. It uses a combination of open-source, licensed services, and other proprietary technologies. Is does permit the use of the content for social media so I did dot my “i’s” and crossed my “t’s”. #responsibleAI This week’s inspiration comes from Booz Allen Hamilton’s open-source solution, aiSSEMBLE. The aiSSEMBLE solution represents a pioneering leap in the field of AI, leveraging a lean manufacturing-inspired approach to accelerate impactful delivery of AI at enterprise scale—reducing pilot purgatory, expanding the field of AI to clients via an open architecture, and ensuring repeatability for future AI solutions. #aiSSEMBLE We have the chance to standardize AI, where we can, through re-use. There are too many missions and applications that need to be solved; we need to be agents of change that embrace interoperability; we need to move faster and efficiently! Tell me about a time when you embraced re-use and how it accelerated your results! Bill Vass John Larson Justin Neroda Jessica Reinhardt Bryan Castle More details below: Press release: https://lnkd.in/gSuG6PVf aiSSEMBLE Homepage: https://lnkd.in/g_5sHg3H

30

If you want to save a ton computational resources and time for AI development, leverage hyper-Networks for efficient model customization. Hyper-networks offer a great solution to a common challenge: Adapting large, pre-trained models to new tasks without the need to retrain them entirely. Instead of modifying the original model's weights which can be both time-consuming and computationally expensive we introduce a smaller, trainable network alongside the frozen base model. This hyper-network, acts as a bridge between new inputs and the base model's existing capabilities. It processes additional information such as depth maps and integrates it with the base model at strategic points so we can guide the base model to produce outputs that consider these new inputs, all without altering its original parameters. The advantages of this approach are huge: • Efficiency: Training only the hyper-network reduces computational costs and accelerates development cycles. • Stability: Keeping the base model frozen preserves its learned knowledge and prevents potential degradation of performance on tasks it already handles well. • Flexibility: We can customize models for specific tasks or datasets without the overhead of full retraining, enabling rapid iteration and experimentation. Imagine enhancing a powerful text-to-image model to incorporate new information. Retraining the entire model would be impractical. By using a hyper-network, we can efficiently adapt the model to accept new inputs, expanding its capabilities while conserving resources. Understanding this has reinforced the importance of looking beyond traditional training methods. Interested in going deeper? You can read my full blog on this in the comment section down below.

29

3 Comments

🥁 Exciting News in the AI Sphere! 🥁 Llama3 is officially out! Today marks the release of both 8B and 70B models of Llama3, promising a substantial leap forward in AI capabilities. With an impressive 8k context length and trained on a massive 15 trillion tokens, Llama3 represents a significant technological advancement.

9

1 Comment

Good takeaways from our session with the one and only David Jayatillake about Semantic Layers. Thanks Data Action Mentor for having us!

8

1 Comment

Topic Modeling For Humans. Gensim is an open-source library for unsupervised topic modeling, document indexing, retrieval by similarity, and other natural language processing functionalities, using modern statistical machine learning. #nlp

2

Are multimodal systems the future of AI? We explore the recent paper by The New York University. Cambrian-1 MLLMs lead with advanced vision models and the Spatial Vision Aggregator (SVA). Read more: https://lnkd.in/ede2d-3h

1

I processed todays drone imagery in both WebODM and Luma. One for a photogrammetric model and one for Gaussian Splats (3DGS). @luma ai finished first and below is the 3DGS animation delivered by the Luma software. Again, if you are not familiar with 3DGS this is not a video I captured with my drone, it's an animation based on a 3D model created from 293 images I captured in three orbital projects that I flew at varying altitudes. I can turn around and create any animation / video I want out of this model using the LUMA web interface. It's amazing but differs from photogrammetry in that it's not based in real world coordinates and volume, distance, and area measurments cannot be taken from it. It provides some amazing flexibility for site inspection since you can pretty much look wherever you want and piece it together in a fly through. Amazing stuff. Stay tuned for the photogrammetry output. #photogrammetry #3dgs #luma #gaussiansplatting #3dmodelling #remotesensing #gis #geospatial Pix4D Luma AI Jonathan Stephens

187

9 Comments