ShengYun (Anthony) Peng

Check out this fascinating research by Peichen Pan, Chang-Yu Hsieh, Tingjun Hou, et al. from Zhejiang University! 'A flexible data-free framework for structure-based de novo drug design with reinforcement learning' This work introduces a unqiue structure-based de novo drug design method that utilises Monte Carlo tree search alongside a carefully curated fragment library. This approach effectively overcomes the limitations of traditional atom-by-atom molecular generation methods. Read here for free via Diamond💎 Open Access: https://lnkd.in/eWtRSN9M

1

A Distributional Perspective of Reinforcement Learning 🌟 TonKien Check this out: https://lnkd.in/gqmiccTC In this seminar we discuss Distributional Reinforcement Learning (RL), a cutting edge approach to reinforcement learning that models not just expected returns, but the entire distribution of possible returns. In this seminar we will introduce the foundational principles, key methodologies and the most recent advances in distributional RL. About us: 📚 We share knowledge about AI through detailed seminars, covering various domains such as Reinforcement Learning, LLM, MLLM, Vector Search, Computer Vision, Web Crawling, and more. 🌍Our goal is to foster an educational community eager to learn and apply AI technologies across different sectors. 🌟 We definately can make mistake as it's human-being so we are really appreciated reading your feedbacks, feel free to express your opinion!!

1

Can we leverage the strengths of both Transformer and Mamba to excel in tasks requiring both long-term context and efficient sequential processing? The Bamba-9B model does exactly that! Bamba-9B is an inference-efficient hybrid Mamba2 model. It not only achieves competitive performance against state-of-the-art transformer models such as LLama3.1-8B but also delivers 2.5x throughput improvements and 2x lower latency! Bamba is the result of a fantastic collaboration between IBM, UIUC, CMU, and Princeton. And the best part? It's publicly available for everyone to explore and experiment with! - Tech blog: https://lnkd.in/gSgNZnhf - GitHub: https://lnkd.in/gf6M_M8E Raghu Ganti, Mudhakar Srivatsa, Davis Wertheimer

24

Interested in solving combinatorial #scheduling problems with differentiable models and gradient descent on GPUs, TPUs, or other accelerators? Coming at ICML 2024 #ICML, Mingju Liu (Mingju's website: https://lnkd.in/gbyWAwT9 ) present our paper, "Differentiable Combinatorial Scheduling at Scale," where we introduce a fully differentiable modeling approach for combinatorial scheduling. This new formulation allows for a fully differentiable encoding of SDC-scheduling #sdc #scheduling, which can be extended to broader combinatorial problems under linear programming (LP) models. Great thanks to the scheduling/HLS expert Prof. Zhiru Zhang from Cornell for many insightful discussions and advice, and congrats to our leading student authors Mingju Liu Yingjie Li. If you are interested in solving discrete combinatorial problems with differentiable models using ML infrastructures, please check out our paper and GitHub repository, and feel free to reach out! Paper: https://lnkd.in/gFCyGXZX Code: https://lnkd.in/gUik3NuA #ICML2024 #Scheduling #Opensource

125

2 Comments

Predictions based on Machine Learning and Statistics often rely on maximal likelihood, which can trap LLMs in biases inherited from the majority opinions of training data. This problem has been substantially addressed through conditional statistics, allowing for mitigation of biases and preventing the cancellation of minority perspectives. https://lnkd.in/gMXKqwyB

7

Great learning materials on Equivariant Neural Networks put together by Wenhan Gao, along with toy code to play with. If you are interested in Group CNNs, Steerable CNNs, Spherical Tensors, Cartesian Tensors, Irreps, Tensor Product, Spherical Harmonics, Group Averaging, Frame Averaging, etc, can check this for an easy start: https://lnkd.in/enfwyenj Other topics that Wenhan is working on are here: https://lnkd.in/ebmjMwm4 #Equivariance, #Symmetries, #AI4Science, #SciML, #GeometricDeepLearning

34

𝐌𝐮𝐬𝐭-𝐫𝐞𝐚𝐝, 𝐟𝐫𝐞𝐞 𝐛𝐨𝐨𝐤 𝐭𝐡𝐚𝐭 𝐩𝐫𝐨𝐯𝐢𝐝𝐞𝐬 𝐚 𝐩𝐚𝐭𝐡𝐰𝐚𝐲 𝐟𝐫𝐨𝐦 𝐆𝐀𝐈 𝐭𝐨 𝐀𝐆𝐈! ​ Yann LeCun has expressed skepticism about the potential of LLMs to achieve Artificial General Intelligence (AGI), citing limitations in memory, planning, and grounding in real-world understanding. However, we argue that 𝗟𝗟𝗠 𝗖𝗼𝗹𝗹𝗮𝗯𝗼𝗿𝗮𝘁𝗶𝘃𝗲 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲 (𝗟𝗖𝗜)—particularly with multiple multimodal LLMs working together—offers a promising architecture to address these challenges and progress toward AGI. ​ Two key architectural innovations underlie LCI: ​ 𝐒𝐞𝐩𝐚𝐫𝐚𝐭𝐢𝐨𝐧 𝐨𝐟 𝐔𝐧𝐜𝐨𝐧𝐬𝐜𝐢𝐨𝐮𝐬 𝐏𝐫𝐨𝐜𝐞𝐬𝐬𝐢𝐧𝐠 𝐚𝐧𝐝 𝐂𝐨𝐧𝐬𝐜𝐢𝐨𝐮𝐬 𝐑𝐞𝐚𝐬𝐨𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐏𝐥𝐚𝐧𝐧𝐢𝐧𝐠: Inspired by the human mind’s dual-layer structure, LCI divides its architecture into foundational and adaptive layers. The foundation layer, akin to unconscious processing, leverages extensive datasets to build robust pattern recognition, encoding essential responses much like human instincts. The adaptive layer, mirroring conscious thought, enables rapid adaptation and contextual reasoning. This dual-layer approach allows LLMs to perform complex tasks with foundational knowledge while adjusting to new information on the fly, much as a child learns new concepts within a developed cognitive framework. ​ 𝐃𝐢𝐬𝐭𝐢𝐧𝐜𝐭 𝐑𝐨𝐥𝐞𝐬 𝐟𝐨𝐫 𝐄𝐚𝐜𝐡 𝐋𝐋𝐌-𝐀𝐠𝐞𝐧𝐭: Each LLM-agent is assigned a specific role, such as executive (knowledge processing), legislative (behavior guardrails), and judicial (adapting to cultural norms). ​ Additionally, our validation of the LCI framework has been thoroughly tested through real-world deployment in applications across various domains, including healthcare, sales planning, investment, and debiasing news. These empirical deployments highlight the practical viability and adaptability of the LCI framework, showcasing its ability to enhance reasoning and decision-making through collaborative intelligence. ​ The eleven aphorisms presented in this free book illustrate the philosophical foundation of the LCI framework, backed by empirical studies. Through LCI, we envision a clear pathway to AGI, rooted in collaborative intelligence and flexible, context-aware adaptation. https://lnkd.in/gn74DkNj (free) https://lnkd.in/gszSvfqP ($6.98 printing fee)

9

✨ Jiayuan Hao is a second-year PhD student in biostatistics program. Hao’s area of interest is environmental statistics and environmental health. “I would like to be a researcher in biostatistics developing advanced statistical tools for complicated environmental health studies after graduation,” shares Hao. 🖥️ This award enabled Hao to train in Deep Learning models more efficiently. Similarly, accessing APIs from OpenAI broadened Hao's perspective on the application of AI techniques in health-related research. 💡 “Last year, I stepped out of my comfort zone and began genomics research. Although, I initially found it incredibly challenging, my persistence in seeking out resources and mentors helped me overcome the steep learning curve. It shattered my self-imposed limitations and instilled in me a newfound confidence. The key takeaway was a reaffirmation of my belief: with dedication and perseverance, I can excel in any field of my choosing.” 💭 Hao encourages other students to consider applying for the Student Success Award. "It represents an excellent chance to reflect on the challenges they have overcome in the past year and to set new objectives for the future. This process is not just about recognition but also about personal growth and envisioning a path forward."

10

We are happy to announce the addition of knowledge distillation to torchtune, a PyTorch library for easily authoring, fine-tuning and experimenting with LLMs. Knowledge distillation is a technique for imparting knowledge from a larger teacher model to a smaller student model. This was a key step in Llama 3.2 pretraining. Check out how we can leverage knowledge distillation in post-training to distill Llama 3.1 8B to Llama 3.2 1B using torchtune: https://hubs.la/Q02YyqPP0

816

1 Comment

Decoding Arithmetic Reasoning in LLMs: The Role of Heuristic Circuits over Generalized Algorithms A key question about LLMs is whether they solve reasoning tasks by learning transferable algorithms or simply memorizing training data. This distinction matters: while memorization might handle familiar tasks, true algorithmic understanding allows for broader generalization. Arithmetic reasoning tasks could reveal if LLMs apply learned algorithms, like vertical addition in human learning, or if they rely on memorized patterns from training data. Read the full article: https://lnkd.in/eXvGdZVH Paper: https://lnkd.in/emyu3B7U

3

[Paper Update] I am pleased to announce that our paper, "LoAS: Fully Temporal-Parallel Dataflow for Dual-Sparse Spiking Neural Networks," has been accepted at #MICRO2024! In this work, we delve into accelerating sparse matrix-sparse matrix (spMspM) operations for dual-sparse SNNs, where both the spike and weight matrices are sparse. Deploying dual-sparse SNNs on spMspM accelerators using existing dataflow and architecture designs is suboptimal due to repetitive access to expensive memory and sparsity-handling units across timesteps. To address this, we spatially unroll and parallelize all timesteps, creating a fully temporal parallel (FTP) dataflow. To enhance the efficiency of the FTP dataflow, we developed an FTP-friendly spike compression mechanism. Neurons that fire no output spikes across timesteps, termed silent neurons, are not processed. Additionally, we proposed an innovative SNN-tailored inner join circuit that is more power and area-efficient without sacrificing throughput. These innovations are integrated into LoAS, a low-latency inference accelerator for dual-sparse SNNs. 📰 Preprint (early version): https://lnkd.in/e3ivB54C 💻 Sample source codes: https://lnkd.in/e_UtXZAY Authors: Ruokai Yin Youngeun Kim Collaborator: Di Wu #SpikingNeuralNetworks #SparseMatrices #TemporalParallelDataflow #Neuromorphic #MICRO2024

155

3 Comments

Exploring In-Context Reinforcement Learning in LLMs with Sparse Autoencoders Read the full article: https://lnkd.in/ePyEX-YC Paper: https://lnkd.in/eeE_ErWt

2

Our paper "ChatDev" has been successfully published in the ACL2024 Main Paper List, one of the top conferences in NLP! 🎉 This work, in which I contributed as the third author during my research internship at Tsinghua University, is among the first to apply a large language model multi-agent framework to software development. We've received significant support from the community, and it’s an honor to be recognized at such a prestigious platform. ------ Check out the paper here: https://lnkd.in/gewnGbUV See the ACL Main Paper List here: https://lnkd.in/gBquKsEN Additionally, we have made numerous updates and expansions on our GitHub repository. I encourage everyone to take a look, contribute, and engage in discussions: GitHub - https://lnkd.in/gRkAG3su ------ Thank you all for your support and attention! Let’s keep pushing the boundaries of AI together. P.S. Feel free to share this post to spread the news! #AI #ACL #NLP #LLM #multiagent #Artificial_Intelligence

42

3 Comments

Simulink with super-powers...

5

Dylan Patel - Inference Math, Simulation, and AI Megaclusters https://lnkd.in/gA3WVQh7 Summary generated by Gemini 2.0 Flash, edited by me: AI Mega Clusters: - Microsoft's Arizona data center, a 48 MW building with ~32,000 GPUs - Next-gen data centers: today some companies already have 300 MW datacenters, Microsoft plans for a 1.5 GW site with a million GPUs - Challenges in scaling due to limitations in power delivery and the need for large-scale power conversion. (context: avg US household uses 1 KW, small town 1MW, mid-sized city 1GW) Multi-Data Center Training: - Shifting from single to multi-data center training to scale model size, as most single sites are limited by their power grid. - Transitioning from synchronous to asynchronous gradient descent methods to overcome the latency and reliability challenges of multi-data center training. - Challenge: preventing one slow or failed GPU from impacting the efficiency of the entire system Data Center Examples and Approaches: - Google's data centers use liquid-cooled TPUs, and they are now connecting multiple data centers together to scale training runs. - Elon's approach: converting a washing machine factory into a data center using mobile generators and mobile chillers to quickly scale out compute. The first to deploy water-cooled Nvidia GPUs in large scale. Using Ethernet instead of Infiniband. - Amazon is building multiple data centers at the same time and wiring them together with fiber. - Meta's data center evolution from one focused to serving recommended content across social media platforms, to a new more dense design with water-chillers to support AI workloads. Inference Math and Tokenomics: - Inference providers are token factories, and token generation costs are significantly impacted by various optimizations (quantization, batching, parallelism, etc.). - Cost of serving token decreases as you can batch more users together; tradeoff between increasing batch size, slower generation, less expensive to host vs decreasing batch size, faster generation, more expensive to host - KV cache size grows with sequence length impacting the number of users that can be served and the time it takes to serve tokens. - OpenAI's O1 model increases sequence length and thus is more difficult to batch. - The cost of a server is mainly driven by the chip's price, while hosting cost is secondary but still very important. Future and Challenges: - Emphasis on the need for new architectures beyond transformers, but also the importance of systems level optimizations for efficiency. - The growing demand and expenditure for data centers raise the question of long-term sustainability and potential infrastructure problems. - There is a push for transforming the current architecture of inference (SSMs?), but no current model is more efficient than Transformers.

6

1 Comment

checkout this insightfull paper which is explain about efficient classical computation of single qubit marginal measurement probabilities to simulate certain classes of quantum algorithm

4