Oluwatobi Olabiyi

Get real! Solving a complex-value linear system using real arithmetic. Sometimes, our favorite software is not complex-friendly. For instance, the IEEE-754 standard for floating-point arithmetic does not natively support complex numbers. Also, TensorFlow has limited capabilities with complex numbers in Graph XLA mode. In such cases, when solving a linear system, a natural option is to convert it to its real counter-part and then apply classic techniques operating in the real field. Surprisingly, the proper way to do this is not well known—that's why I decided to write the next blog post 😊 https://lnkd.in/eiW4ufk5 For more detailed info, you can check out the paper by Dai et al. https://lnkd.in/esHwDbbD

49

4 Comments

How will the integration of large language models transform the safety and reliability of autonomous robotic systems? 🧐 Researchers at Stanford University and NVIDIA have developed a pioneering two-stage framework that leverages large language models (LLMs) for detecting and mitigating robotic system failures in real time. Their work, which won the Outstanding Paper Award at the Robotics: Science and Systems conference, addresses the challenge of detecting out-of-distribution (OOD) inputs that can hinder the performance of autonomous vehicles. The framework overcomes computational limitations by using a two-stage reasoning pipeline, enabling low-latency reactivity in the first stage and leveraging the full generative capabilities of LLMs in the second stage. By rapidly detecting anomalies and slowing down robot actions, the system allows the LLM to reason about safety-preserving interventions, significantly enhancing the trustworthiness of autonomous robotic systems. In tests, the researchers found that their framework outperformed other methods and showed that smaller LLMs could perform as well as larger ones for anomaly detection. This suggests that the deployment of both fast and slow reasoning can improve the practicality of using LLMs for robotic planning tasks. How do you think the integration of large language models in robotic systems will impact the future of autonomous technology? Let's discuss! #RoboticSystems #AIIntegration #FutureTechnology

2

0059: Mapping Molecular Hamiltonians To Qubit Operators with Jordan-Wigner Transformation In quantum computing, the efficient mapping of molecular Hamiltonians to qubit operators is a critical step in solving electronic structure problems, especially in quantum chemistry. One of the most established methods for achieving this is the Jordan-Wigner Transformation (JWT). The JWT translates fermionic operators, which describe the behavior of electrons in molecules, into qubit operators that can be used in quantum algorithms. The central challenge in electronic structure calculations is the fermionic nature of electrons, which obey the Pauli exclusion principle and anti-commutation relations. Fermionic operators are used to create and annihilate electrons in specific quantum states, but these operators do not directly translate to qubit-based quantum computing. This is where JWT comes in: it provides a systematic way to represent these fermionic operators as qubit operators that obey the commutation relations of quantum gates. The Jordan-Wigner Transformation maps each fermionic operator to a string of Pauli matrices on qubits. Specifically, each fermionic creation or annihilation operator is mapped to a combination of Pauli-Z, Pauli-X, and Pauli-Y operators that span the qubits, ensuring that the anti-commutation properties of fermions are preserved in the qubit space. By mapping molecular Hamiltonians using the Jordan-Wigner Transformation, we can represent the electronic interactions and energy of a molecule in a form that quantum computers can process. Once this mapping is complete, variational algorithms like the Variational Quantum Eigensolver (VQE) or Quantum Phase Estimation (QPE) can be employed to estimate the ground-state energy of the molecule. These algorithms aim to find the eigenvalues of the molecular Hamiltonian, which directly correspond to the molecular energy levels. The benefit of using JWT is that it scales linearly with the number of qubits, making it an efficient method for near-term quantum devices. However, it has some limitations, such as the need for long-range qubit interactions, which may pose a challenge for hardware implementations. Other transformations, like the Bravyi-Kitaev Transformation, offer alternatives with different trade-offs in qubit connectivity and operator complexity. The Jordan-Wigner Transformation is foundational in quantum chemistry simulations on quantum computers. It bridges the gap between the fermionic nature of molecules and the qubit-based framework of quantum computing. As the field progresses, leveraging such transformations is essential for advancing quantum simulations of complex molecular systems, which could lead to breakthroughs in fields like drug discovery, material science, and energy research. Check comment for links to code repository and resources in Quantum Chemistry.

69

8 Comments

An interesting topic

17

Watch Gen AI in action generating the RTL code for SHA-256 HW implementation ----------------------------------------------- In my course, Generative AI for Algorithm-to-Silicon Mapping, we’re diving into intersting experiments with Gen AI. In this session, I generate RTL Verilog code for part of the architecture design of SHA-256 algorithm, showcasing how today’s Gen AI tools can produce high-quality RTL code along with test stimulus, almost autonomously. This is amazing to experiment that how Gen AI optimizes circuits within the complex space of digital design —watch the code generation in action with Gen AI tools. For years, digital system design and verification have steadily evolved, with the abstraction level pushed higher and higher. But now, generative AI, large language models (LLMs), and agentic approaches are setting entirely new standards—enabling complex digital designs to be realized and verified with minimal manual coding effort. By thoughtfully architecting designs with systematic planning and phased execution, we're able to streamline intricate digital projects like never before.

62

1 Comment

Our goal is to simplify AI and accelerate your journey from prototype to deployment. It was great working with Siddhika Nevrekar on this blog post, which highlights our vision and the efforts to democratize AI. Read more here: https://lnkd.in/gmBijf7B

12

Attractor networks are building blocks of many dynamic neural networks. The #architecture proposed in this paper is a step towards overcoming the von Neumann bottleneck by accelerating matrix-vector multiplication in #Resistive #memory #devices #nonvolatile #neuralnetworks #AI

3

Introducing PMNet🗼: A new scalable pathloss map prediction method for 6G networks! Accepted at IEEE TWC, PMNet leverages ML to generate city map-scale pathloss maps quickly and accurately. Paper 📄: arxiv.org/abs/2312.03950 Code 💻 : github.com/abman23/pmnet - Need accurate pathloss predictions, faster network planning, and adaptation to new scenarios? It predicts pathloss with an RMSE of 10⁻² in milliseconds - PMNet’s transfer learning allows it to adapt 5.6x faster with 4.5x less data, outperforming traditional methods like 3GPP in various environments. Big thanks to my co-author Prof. Andreas F. Molisch for making this journey possible!

55

1 Comment

One of this year’s People to Watch is Yann Lecun. He is best known for his work in deep learning and the invention of the #convolutional #network method. Yann is currently a part-time professor at #NYU and a VP and Chief AI Scientist for #FacebookAI Research. https://ow.ly/fKuv50RxWYq #artificialintelligence

1

Putting the NVIDIA GH200 Grace Hopper Superchip to good use: superior inference performance and economics for larger models https://lnkd.in/g7XC8t9Q

3

It is very important to reduce KV cache memory consumption during long context inference. We introduce ThinK, a method that exploits substantial redundancy across the channel dimension of the KV cache. Our findings reveal that even KV caches optimized by eviction methods, such as SnapKV and H2O, it can still benefit from pruning approximately 40 to 50% of their key cache channels. Our approach results in achieving over a 20% reduction in KV cache memory compared with SnapKV and H2O. Paper: https://lnkd.in/gxpeVdC9

154

From models to systems 90ies were an interesting time for computing - with the new iteration of processors appearing every year or two. 286, 386, all the way to Pentiums. However, at that point development of single processors slowed down (heat generation in CMOS), and computers started to transition to dual and quadruple cores. And then the time of the cloud arrived. I wonder whether the same transition is starting with the ML world. The models for individual tasks saturate at slightly above average human level if they are largely supervised learning. In other words, the model cannot do better then the data it has been trained on. Models based on algorithmic tasks - from chess and Go to math and protein folding - can do better (advantage of multiple random rollouts for A* and MCDTs). But here the range of functionalities is limited. So, are we entering the era of model integration into functional workflows? LLM frontend with BO backend and KG intermediate data layers, for example?

77

1 Comment

When did “compute” become a noun? I've been hearing about “compute and memory” or the “cost of compute”... and I thought the noun was “computing." Maybe it’s some recent jargon. Maybe it’s the tendency of some verbs to become nouns? For example “ask.” Or “interconnect,” which, in electronics, has been a noun for as long as I remember. Merriam-Webster has “compute” only as a verb: https://lnkd.in/gShtHigW and Cambridge Dictionary agrees: https://lnkd.in/gqk6b3JA   On the other hand, Oxford English Dictionary says it's a noun since the 1500s: https://lnkd.in/gSTN8ay6. And Collins Dictionary lists it as a noun replacing “computation”: https://lnkd.in/gKaNv645 Maybe we should ask a few counties in Pennsylvania to decide this for us?!

57

9 Comments

Best Paper Award at the Intelligent Systems (IntelliSys) Conference, Amsterdam, 5-6 September 2024. Syed Muhammad Salman Bukhari, Muhammad Hamza Zafar, Syed Kumayl Raza Moosavi, Noman Mujeeb Khan, and Filippo Sanfilippo. "FireNet: A Hybrid Deep Learning Approach for Enhanced Fire Detection in Remote Sensing Imagery." In Intelligent Systems Conference, pp. 1-15. Cham: Springer Nature Switzerland, 2024. We are especially proud of the work conducted by our talented team member Syed Muhammad Salman Bukhari! In this study, we address the critical task of early bushfire detection in remote areas through UAV imaging, a vital effort for protecting inhabitants, infrastructure, and ecosystems within the smart cities paradigm. We introduce FireNet, a deep learning framework that employs a hybrid Inception Residual Neural Network (Inception-ResNet) model, specifically designed for the rapid and precise classification of fire and non-fire regions in real-time image processing scenarios. FireNet synergises the strengths of both Inception and ResNet architectures, enhancing the model’s feature extraction capabilities while significantly reducing computational overhead. A distinctive aspect of FireNet is the adoption of the HardSwish activation function, demonstrating superior performance over the conventional Rectified Linear Unit (ReLU) in our fire detection cases. Through rigorous evaluation of a robust dataset, FireNet achieves an impressive accuracy of 96%, with a corresponding AUC of approximately 96%. These results not only affirm FireNet’s efficacy in accurately identifying fire and non-fire areas but also underscore its importance as a crucial tool for real-time fire monitoring and rescue operations, especially in the context of smart cities. With its outstanding efficiency and accuracy, FireNet represents a significant leap forward in the domain of fire detection technologies, highlighting the role of deep learning in enhancing urban resilience against fires. The Science and Information (SAI) Organization Artificial Intelligence, Biomechatronics, and Collaborative Robotics UiA - Mechatronics / Mekatronikk CIEM - Centre for Integrated Emergency Management UiA - Faculty of Engineering and Science University of Agder (UiA) hashtag #DeepLearning hashtag #FireDetection hashtag #UAV hashtag #Imaging hashtag #InceptionResNet hashtag #SmartCities hashtag #IntelliSys

12