Moez Draief

This paper proposes a #distributionally #robust #optimization (#DRO) model to minimize, under uncertain latency probability distributions, the expected worst-case overall task routing delay from source to target user equipment through satellite constellation. The model addresses undetermined #uploading and #downloading latency between automobiles, satellites, and user equipment by employing the #Wasserstein #ambiguity #set, allowing for unpredictable vehicle mobility and intermittent connections. By reformulating the problem into a tractable form, the authors determine the optimal routing path for #task #uploading, #satellite #constellation, and #task #downloading. Ultimately, the performance of the proposed DRO model demonstrates the model’s ability to address the challenges of integrated satellite-terrestrial network routing.----Kai-Chu Tsai, Lei Fan, Ricardo Lent, Li-Chun Wang, @Zhu Han More details can be found at this link: https://lnkd.in/gn2tEe9H

4

1 Comment

For all large language models (LLM) can do, it's important to recognise what they can't. This research paper addresses a critical question: How reliable is LLM's knowledge? The authors, including the Head of Aveni Labs, Dr. Alexandra Birch, and Dr. Barry Haddow, our Head of Natural Language Processing, introduce a novel metric called MOdel kNowledge relIabiliTy scORe (MONITOR) to assess LLM's factual reliability. MONITOR works by asking models the same question in different ways and seeing how likely the model is to give the same answer each time. This innovative approach helps identify situations where an LLM might be making things up rather than providing facts. Read more: https://hubs.la/Q02w7fJk0 #LLM #largelanguagemodels #naturalLanguageProcessing #machinelearning #artificalintelligence

5

1 Comment

In this paper, the aythors propose a series of #Resource #Allocation (#RA) strategic algorithms harnessing the Transfer Learning, #Growth-#Share (#GS) matrix, #Game #Theory (#GT), and service priorities to tailor the aforementioned trade-off. This endeavour renders the network more intelligent, self-sufficient, and resilient. Furthermore, they have seamlessly integrated Device-to-Device communication scenarios into their proposed algorithms, enhancing #Spectrum #Efficiency (#SE) and network capacity. The proposed integration aims to strengthen overall system performance and accommodate the evolving demands of future wireless networks. Their primary contribution lies in the development of the #GS-#GT-#based #Optimal #PathFinder (#GS-#GTOPF) algorithm to identify optimal paths based on SE using Deep Neural Networks. Thereafter, they formulate an enhanced version of it by integrating #service #priorities (#GS-#GTOPF-#SP). This refinement has been further advanced by reducing the #Computational #Time (#CT), resulting in #GS-#GTOPF-#SP-#rCT. Further improvement is achieved by introducing the angle criterion (#GS-#GTOPF-#SP-#rCT- #θ). ---- Vivek Pathak, Chethan R, Rahul Jashvantbhai Pandya, Sridhar Iyer, Vimal Bhatia More details can be found at this link: https://lnkd.in/etxP5fsd

9

This is exactly right. LLMs will achieve fluency and not intelligence. Fluency won’t be enough for AGI, and this is also a call for us in the wireless community to heed the advice of not just rolling with “LLM for this” and “LLM for that”. The irony is that in wireless, we often (rightly so) complain of the lack of data, and now we want to rely on tools like LLMs that require gazillion data points, that likely wont work (except within limited contexts and use cases). Our AGI-native networks paper, that I shared recently: https://lnkd.in/e9rtKEab hits exactly on these points, and provides an initial path towards common sense and some form of actual cognition in AI. It also showcases how that will be synergistic with wireless networks (6G and potentially beyond). Next-generation AI must meet next-generation networks to achieve a meaningful leap in both that goes beyond fluency. We should all realize this fact sooner rather than later. #6G #AGI Omar Hashash

124

3 Comments

This week at #ICLR2024, #Microsoft #Research present #MathVista, an open-source #benchmark for measuring foundation models’ #mathematical #reasoning capabilities in #multimodal scenarios. Learn about the work now from coauthor Michel Galley in the “Abstracts” podcast. https://msft.it/6042YpuIk #AI #GenerativeAI #LLM #Mathematics

9

Quantum Error Correction was a key topic of discussion in last week's Economist Impact "Commercialising Quantum" event, including in this panel moderated by Christopher Bishop and featuring Ilyas Khan from Quantinuum, Maud Vinet from Quobly, Kenny Lee from Google, and our Chief Commercial Officer Yuval Boger. What did we learn? - Algorithms that leverage logical qubits and quantum error correction are closer than ever, bringing us closer and closer to customer value. - Who will win the logical qubit race? While we feel very good about QuEra's position in this "race", the real winners will be the customers. - Different modalities will implement error correction differently. For instance, modalities such as neutral atoms that allow all-to-all connectivity can implement codes that cannot be readily executed on modalities with static qubits. - Better physical qubits lead to better logical qubits (as well as to more efficient encoding for a target logical error rate), and thus significant efforts will continue to improve single- and two-qubit physical gate fidelity.

31

1 Comment

There is a kind of wack-a-mole thing going on with false claims of achievement of AI general/ human-level intelligence (AGI) I recommend that you follow Prof. J. Mark Bishop if you are not already doing so.

1

Sam Altman takes nuclear energy company Oklo public to help further his AI ambitions https://buff.ly/4bztCN4 - 💼 Sam Altman is the CEO of OpenAI and chairman of nuclear power company Oklo - 📈 Oklo went public through AltC Acquisition Corp., trading on NYSE as "OKLO" - 💰 Oklo received over $306 million in gross proceeds from the transaction - 💡 Oklo's business model is based on commercializing nuclear fission for energy needs - 🔬 Oklo is working on building its first small-scale reactor in Idaho - 🌎 Altman sees nuclear energy as crucial for meeting AI energy demands - 🔥 Altman has invested in other nuclear energy and AI-related ventures - 🤔 Altman believes nuclear energy is the best way to meet AI energy needs without fossil fuels - 🚀 Oklo CEO targeting 2027 for first plant to come online - 📉 SPAC market has seen setbacks but AI-related companies are gaining interest on Wall Street

Scaling and more data alone won't achieve human-level intelligence in LLMs. Yann LeCun sees JEPA as a key step. Our new article covers: - How JEPA works - How it differs from transformers - Models based on JEPA: I-JEPA, MC-JEPA, V-JEPA https://lnkd.in/ehqXW6T7

1

In this paper, the authors propose a three-stage joint channel decomposition and prediction framework to acquire #channel #state #information (#CSI). The proposed framework exploits the two-timescale property that the base station (BS)-#reconfigurable #intelligent #surface (#RIS) channel is quasi-static and the RIS-user equipment (UE) channel is fast time-varying. Specifically, in the first stage, they use the full-duplex technique to estimate the channel between a BS’s specific antenna and the RIS, addressing the critical scaling ambiguity problem in the channel decomposition. They then design a novel deep neural network, namely, the #sparse-#connected #long #short-#term #memory (#SCLSTM), and propose a SCLSTM-based algorithm in the second and third stages, respectively. ---- Wangyang Xu, Jiancheng An, Yongjun Xu, Chongwen Huang, Lu Gan, Chau Yuen More details can be found at this link: https://lnkd.in/eb8u6ZXG

1

Exciting news in the AI world! Meta AI has just launched Llama 3, offering 8B and 70B models available for use today. These new models are a significant step forward, as Meta AI is currently training its 400B+ model, positioning it as a strong competitor to OpenAI's GPT-4 upon release.   Highlights of Llama 3 include:   ·     8B and 70B models are ready for use as of today. ·     Impressive 8k context length, enabling more complex processing tasks. ·     The models were trained with 15 trillion tokens on a custom-built 24k GPU cluster, showcasing Meta AI's commitment to large-scale AI development. ·     Llama 3 demonstrates remarkable performance on various benchmarks, with the Llama3-8B model even outperforming the Llama2-70B model in certain cases. https://lnkd.in/gKHMBbsk

12

In the rapidly evolving landscape of Large Language Models (LLMs), traditional benchmarks like the MMLU (Massive Multitask Language Understanding) have been the go-to for evaluating AI capabilities. However, a growing body of research suggests that relying solely on these scores to select LLMs for specific applications could be misleading and costly. 🔍 𝐖𝐡𝐲 𝐁𝐞𝐧𝐜𝐡𝐦𝐚𝐫𝐤𝐬 𝐌𝐢𝐠𝐡𝐭 𝐌𝐢𝐬𝐥𝐞𝐚𝐝: A recent study, "When Benchmarks are Targets: Revealing the Sensitivity of Large Language Model Leaderboards," underscores a critical vulnerability in the standard benchmarking approach. This research demonstrates how minor methodological changes can significantly shift LLM rankings by up to eight positions on leaderboards. Such volatility raises questions about the stability and reliability of benchmark-based assessments. 👩‍💻 𝐓𝐡𝐞 𝐋𝐢𝐦𝐢𝐭𝐚𝐭𝐢𝐨𝐧𝐬 𝐨𝐟 𝐌𝐂𝐐𝐬: Leaderboards often rely on multiple-choice questions (MCQs) to gauge model performance. While MCQs offer a straightforward and quantifiable way to test certain model capabilities, they fall short in measuring comprehensive performance. The research highlights that models might perform well on MCQs by recognizing patterns rather than genuinely understanding content, which doesn’t necessarily translate to effectiveness in practical applications. 🔄 𝐓𝐡𝐞 𝐬𝐭𝐮𝐝𝐲 𝐜𝐚𝐭𝐞𝐠𝐨𝐫𝐢𝐳𝐞𝐬 𝐚𝐝𝐣𝐮𝐬𝐭𝐦𝐞𝐧𝐭𝐬 𝐢𝐧𝐭𝐨 𝐬𝐞𝐧𝐬𝐢𝐭𝐢𝐯𝐞 𝐚𝐧𝐝 𝐮𝐧𝐬𝐞𝐧𝐬𝐢𝐭𝐢𝐯𝐞 𝐩𝐞𝐫𝐭𝐮𝐫𝐛𝐚𝐭𝐢𝐨𝐧𝐬 𝐭𝐨 𝐞𝐱𝐩𝐥𝐨𝐫𝐞 𝐭𝐡𝐞𝐢𝐫 𝐢𝐦𝐩𝐚𝐜𝐭 𝐨𝐧 𝐦𝐨𝐝𝐞𝐥 𝐫𝐚𝐧𝐤𝐢𝐧𝐠𝐬: - Sensitive Perturbations: Modifications like shuffling choice order, changing symbols, and alternative scoring drastically alter model standings, pointing to a fragility in how performance is currently measured. - Unsensitive Perturbations: Adjustments that tweak the relevance of in-context examples show minor performance variations and help mitigate bias, suggesting a need for more nuanced testing methodologies. 𝐅𝐨𝐫 𝐨𝐫𝐠𝐚𝐧𝐢𝐳𝐚𝐭𝐢𝐨𝐧𝐬 𝐥𝐨𝐨𝐤𝐢𝐧𝐠 𝐭𝐨 𝐢𝐧𝐭𝐞𝐠𝐫𝐚𝐭𝐞 𝐋𝐋𝐌𝐬 𝐢𝐧𝐭𝐨 𝐭𝐡𝐞𝐢𝐫 𝐨𝐩𝐞𝐫𝐚𝐭𝐢𝐨𝐧𝐬, 𝐭𝐡𝐢𝐬 𝐫𝐞𝐬𝐞𝐚𝐫𝐜𝐡 𝐬𝐞𝐫𝐯𝐞𝐬 𝐚𝐬 𝐚 𝐜𝐚𝐮𝐭𝐢𝐨𝐧𝐚𝐫𝐲 𝐭𝐚𝐥𝐞: - Custom Evaluation: Assess LLM performance using your specific data and scenarios rather than relying exclusively on generalized benchmarks. This approach ensures the model aligns with your unique requirements. - Holistic Assessment: Consider a broader range of evaluations that go beyond MCQs to include tests that measure real-world task performance and contextual understanding. - Continuous Benchmarking: As LLM technology evolves, so should the methods we use to evaluate it. Businesses must stay updated on the latest research and adapt their evaluation strategies accordingly. Do you think traditional benchmarks suffice, or is there a need for innovation in evaluation methods? #ai #genai #innovation #technology #llms

21

2 Comments

🧠 The #transformer model was indeed one of the biggest breakthroughs of our century which led to a boom in generative AI's capabilities. Understand how transformers work: https://lnkd.in/emAc_VmR #llms #genai #generativeai

139

In computational linguistics, researchers are focused on enhancing language models' ability to extract precise information from extensive textual data. Recent studies introduce "retrieval heads," specialized attention mechanisms that selectively focus on critical segments of large texts, improving targeted data retrieval. These heads, integrated into transformer-based models like LLaMA and Yi, significantly enhance accuracy and efficiency, particularly in handling long-context scenarios. The methodology involves thorough experiments across various models, including the Needle-in-a-Haystack test, which embeds specific information to measure precision. Results show models with active retrieval heads outperform those without, with accuracy dropping notably when these heads are masked. The empirical evidence underscores the effectiveness of retrieval heads in improving information retrieval precision and reliability. This research deepens our understanding of attention mechanisms in large-scale text processing, paving the way for more efficient and accurate language models. https://lnkd.in/dNNPvMXw https://lnkd.in/ddbHX6EA Innovation Hacks AI #aiinnovation #generativeai

1

📊 The Impacts of Data, Ordering, and Intrinsic Dimensionality on Recall in Hierarchical Navigable Small Worlds (HNSW) 📊 The behaviour of HNSW under real-world scenarios using vectors generated with deep learning models remains under-explored...until now! Our latest article explores just that. Key Insights: 🔎 Default Parameters: Many HNSW implementations use sub-optimal default parameters. Our research proposes higher defaults for improved performance, as adopted in Marqo V2. 🔎 Intrinsic Dimensionality: A crucial factor influencing recall, with higher intrinsic dimensionality generally resulting in poorer recall. 🔎 Local Intrinsic Dimensionality (LID): LID significantly affects recall, and our findings indicate that ordering data by LID can enhance recall performance. 🔎 Real-World Application: Through testing on e-commerce datasets, we found that category-based indexing order impacts recall, emphasizing the need for data randomization during indexing. 🔎 Recall Challenges: Under-configured HNSW parameters can lead to a drop in NDCG@10 by up to 18% on popular benchmarks. Additionally, the order of data insertion can impact recall by up to 17%. Recall in HNSW is not a trivial problem but we see there are clearly things we can do to improve our recall other than turning up parameters for the HNSW algorithm. In our article, we also explore some of our tips and best practices for any vector search system. Read the full article here: https://lnkd.in/eFqsZhsh

27

One of the more reasonable articles about the various types of ML/AI and where they might be useful in practical application. Finneran’s Second Law of Compute: The outcome benefit must outweigh the costs. Success = DV > DR Data Value greater than Data Risk. https://lnkd.in/g4Xe6BpX?

39

1 Comment

Unveiling Chain-of-Thought Reasoning: Exploring Iterative Algorithms in Language Models Quick read: https://lnkd.in/eEe4EdAN Paper: https://lnkd.in/eb_D4Dk8 AI at Meta

18

This paper on Qiskit just popped up on arXiv #quantumcomputing https://lnkd.in/eWtCAxkc

184

1 Comment