Jay Shah

Exploring Parameter-Efficient Fine-Tuning Strategies for Large Language Models Quick read: https://lnkd.in/gWKCv38m Researchers from Northeastern University, the University of California, Arizona State University, and New York University present this survey thoroughly examining diverse PEFT algorithms and evaluating their performance and computational requirements. It also provides an overview of applications developed using various PEFT methods and discusses common strategies employed to reduce computational expenses associated with PEFT. Beyond algorithmic considerations, the survey delves into real-world system designs to explore the implementation costs of different PEFT algorithms. As an invaluable resource, this survey equips researchers with insights into PEFT algorithms and their system implementations, offering detailed analyses of recent progressions and practical uses. The researchers categorized PEFT algorithms into additive, selective, reparameterized, and hybrid fine-tuning based on their operations. Major additive fine-tuning algorithms include adapters, soft prompts, and others, which differ in the additional tunable modules or parameters they utilize. Selective fine-tuning, in contrast, involves selecting a small subset of parameters from the backbone model, making only these parameters tunable while leaving the majority untouched during downstream task fine-tuning. Selective fine-tuning is categorized based on the grouping of chosen parameters: Unstructural Masking and Structural Masking. Reparametrization involves transforming model parameters between two equivalent forms, introducing additional low-rank trainable parameters during training, which are then integrated with the original model for inference. This approach encompasses two main strategies: Low-rank Decomposition and LoRA Derivatives. Hybrid fine-tuning explores different PEFT methods’ design spaces and combines their advantages. Paper: https://lnkd.in/gh8qfi4k

26

Researchers at CMU Introduce TriForce: A Hierarchical Speculative Decoding AI System that is Scalable to Long Sequence Generation Quick read: https://lnkd.in/gr_he4kw Researchers from Carnegie Mellon University and Meta AI (FAIR) Introduce TriForce, a hierarchical speculative decoding system designed for scalable long sequence generation. TriForce utilizes the original model weights and dynamic sparse KV cache via retrieval as a draft model, serving as an intermediate layer in the hierarchy. Maintaining the full cache allows for superior KV cache selection using retrieval-based drafting, characterized as lossless compared to eviction-based methods like StreamingLLM and H2O. The hierarchical system addresses dual memory bottlenecks, pairing a lightweight model with a StreamingLLM cache for initial speculations to reduce drafting latency and accelerate end-to-end inference. TriForce introduces a hierarchical speculative decoding system with retrieval-based KV cache selection. The hierarchical system addresses dual bottlenecks, enhancing speed-up. Retrieval-based drafting segments the KV cache, highlighting relevant information. Lightweight models with StreamingLLM cache accelerate initial speculations, reducing drafting latency. TriForce utilizes model weights and KV cache to enhance LLM inference speed for long sequences. The implementation utilizes Transformers, FlashAttention, and PyTorch CUDA graphs, maintaining full layer sparsity while minimizing kernel launching overhead. Paper: https://lnkd.in/ga45hspw Machine Learning Department at CMU

26

Effect of Correlated Errors on Quantum Memory Smita Bagewadi, Avhishek Chatterjee https://lnkd.in/es4_QN4c Recent improvements in LDPC code based fault-tolerance for memory against i.i.d. errors naturally lead to the question of fault-tolerance against errors with long-range correlations. We introduce a classical correlation model based on hidden random fields for modeling such errors. We show that this proposed model can capture certain correlation patterns not captured by the joint (system and bath) Hamiltonian model with pairwise terms. Towards that, we derive a converse result for retention time in the presence of an error distribution, which is from the proposed class and exhibits quadratically small correlations. On the other hand, we show that for a broad subclass of error distributions within the proposed model, Tanner codes can ensure exponential retention time when the error rate is sufficiently low. The proposed model is analytically tractable due to the existence of a rich probability literature and thus, can offer insights complementary to the joint Hamiltonian model with pairwise terms.

3

We are happy to announce that our paper "SAUC: Sparsity-Aware Uncertainty Calibration for Spatiotemporal Prediction with Graph Neural Networks" has been accepted by ACM SIGSPATIAL 2024! 📜 Paper link: https://lnkd.in/efgH22K5 📑ArXiv: https://lnkd.in/evvjYd2A This paper addresses the challenge of uncertainty quantification in spatiotemporal deep learning, which is often overlooked in favor of deterministic predictions. Sparse, highly-granular spatiotemporal datasets pose additional difficulties. To tackle this, the authors propose the Sparsity-aware Uncertainty Calibration (SAUC) framework, which adjusts uncertainty for both zero and non-zero values. They first transform deterministic spatiotemporal Graph Neural Networks (ST-GNNs) into probabilistic models and then calibrate them using quantile methods. Experiments show that SAUC improves variance fitting and reduces calibration errors by 20% in sparse datasets like traffic accident and crime predictions, demonstrating its theoretical and practical value. We extend our heartfelt thanks to our esteemed collaborators: Dingyi Zhuang, Bu Yuheng, Guang Wang, Jinhua Zhao! #UncertaintyQuantification #SpatiotemporalData #GraphNeuralNetworks #DeepLearning #ACMSIGSPATIAL #AIResearch #DataScience #SparseData

14

#IJCAItutorial T25: Advances in Fairness-aware Reinforcement Learning: Theory and Application #IJCAI2024 🗣️Pratik Gajane, Mykola Pechenizkiy, Yingqian Zhang ➡️ https://fair-rl.github.io/ Abstract: The tutorial offers a comprehensive overview of fairness in reinforcement learning (RL), spanning various applications and theoretical frameworks. We will start by delving into the crucial issue of fairness in algorithmic decision-making using the paradigm of reinforcement learning across various high-impact real-world domains. We will provide a comparative analysis of the fairness conceptualizations in the (relatively mature) field of fair-ML with the emerging field of fair-RL. Moving forward, we will introduce fundamental concepts of reinforcement learning, including problem settings and performance measures for both utility and fairness, while also reviewing mathematical results and empirical works on fairness-aware (deep) RL solutions. Furthermore, we will explore the considered fairness notions, discuss integrating fairness into reinforcement learning as a multi-objective optimization problem, and examine the trade-off between fairness and utility across various applications. The tutorial will conclude with a discussion of future challenges in fair reinforcement learning and their prospective solutions, exploring interdisciplinary insights from successful fair machine learning research and applying them to the realm of fair reinforcement learning. #FairRL #RL #AI #ResponsibleTech

10

🎙 Introducing the paper presented by Yuxiang Wang from Wuhan University at #ICDE2024: Generative and Contrastive Paradigms Are Complementary for Graph Self-Supervised Learning. In this paper, the team from Wuhan University, Centre for Perceptual and Interactive Intelligence (CPII) Limited, OceanBase and Peking University proposed a graph contrastive masked autoencoder (#GCMAE) framework which combines masked autoencoder (#MAE) and Contrastive Learning (#CL) to enhance graph self-supervised learning (#GSSL). The results show that #GCMAE consistently provides good accuracy, and the maximum accuracy improvement is up to 3.2% compared with the best-performing baseline. 🏄‍♂️ Learn more: https://lnkd.in/gXgy2az6

1

This week in #TheBatch: 🎨 All about Claude’s Artifacts feature 🌱 AI growth hinders carbon emissions goals 🧠 GaLore, a memory-efficient method for fine-tuning and pretraining Plus: Andrew Ng shares his concerns about the proposed California regulation SB 1047. Read The Batch now: https://hubs.la/Q02G5YpM0

49

✨ Quantum computing holds immense potential for solving complex problems, but current hardware limitations pose a hurdle. Here's how we're tackling them ➡️ Congzheng Liu, Ph.D.'s post describes our work on innovative methods like #MatrixSparsity enhancement and #EuclideanScaling to significantly reduce problem size, making quantum optimisation more achievable with today's technology. This is a leap towards unlocking the true power of quantum computing. Learn more in our blog: https://lnkd.in/dggjtGfi #QuantumAdvantage #KnapsackProblem #QuantumComputing

14

Robust determination of cubic elastic constants via nanoindentation and Bayesian inference https://lnkd.in/gEcRgQJq

5

0 or 1 Prediction aka Binary Classification What loss function would you use? Binary Cross-Entropy or Cross-Entropy (Multi-class) ? At beginning building a Binary Classification task at any Vision Task, i would go with the Binary Cross-Entropy. Why? simpy its a binary task :') Then, realize that when the model predict a wrong prediction. It would be really hard to inspect what CAM result from each class has, as Binary is only have 1 branch where 0 and 1 are fused together. However when use Cross-Entropy loss, you could easily target your interested class to inspect what its actually the network highlight to reason a wrong prediction. As within this couple weeks, i am focusing to write my dissertation and re-read some papers, the Paper from AWS AI by Xu et al. with the title "On Improving Temporal Consistency for Online Face Liveness Detection System" reminded me my previous statement and describe it more details why you should use Cross-Entropy loss. Here are 3 reasons why you should consider your classification supervision using multi-class (CE) instead of Binary CE: 1. The discriminative features to distinguish different types of presentation attacks can be learned 2. The embedding space of liveness class can be squuezed into a more compact space than using a binary classfication, which helps decrease the false positive rate 3. Fine-grained analysis can be conducted when the model makes mistakes. #BinaryClassification #CrossEntropyLoss #MachineLearning #DataScience

14

🔍 We encourage you to explore one of the April TCAS-I Paper Highlight videos "Training Neural Networks with In-memory-computing Hardware and Multi-level Radix-4 Inputs". 📝 Authors: Christopher Grimm; Jinseok Lee; Naveen Verma Training Deep Neural Networks (DNNs) takes an immense number of operations and therefore a lot of computational power. While numerous accelerators exist to reduce the power consumption for running pre-trained DNNs (inference), training proves to be a more challenging task. Two reasons for this are that training requires higher precision computations and computations over much more data than inferencing. In-memory computing (IMC) is an approach that can overcome the challenges of managing large amounts of data, but to do so it leverages analog computations for energy and area efficiency, which limits its achievable precision. In this paper, we explore and test novel training methods capable of obtaining high performance on DNN accelerators with significant levels of analog noise. We demonstrate the possibility of using IMC for training DNNs within 2% of performances obtained by GPUs at 3 orders of magnitude greater energy efficiency. Watch the video to learn more 👇 🔗 Read the paper on IEEE Xplore here: https://loom.ly/9NeufvY #IEEECASS #TCASI #AprilHighlight

4

Physics-Based Deep Learning: Insights into Physics-Informed Neural Networks (PINNs) https://lnkd.in/gzWs-iFX #deeplearning #ai

17

🚀 Control Systems in AI 🚀 Control systems often face challenges with high-dimensional, noisy data. The ACDF algorithm, improves the learning of action-dependent latent states, and ensures more accurate representation and control in complex environments. Paper 🔗 - https://lnkd.in/gi82aUqn 💡 Keen on AI breakthroughs? DM us to explore opportunities!

9

Advancing Ethical AI: Preference Matching Reinforcement Learning from Human Feedback RLHF for Aligning LLMs with Human Preferences Quick read: https://lnkd.in/gr-238BN Paper: https://lnkd.in/gJddxRws

15

🚀 Concurrent Learning: Revolutionizing High-Throughput Crystal Structure Prediction 🔍 A recent collaboration between Jilin University and the Institute of Applied Physics and Computational Mathematics (Beijing) has led to the development of a concurrent learning scheme for crystal structure prediction (CSP). By integrating the CALYPSO method with DP-GEN, this approach significantly reduces the computational cost of CSP while maintaining high accuracy. This breakthrough research, now published in Physical Review B, highlights the potential of combining machine learning potentials (MLPs) with CSP algorithms for large, complex systems. From predicting stable Mg-Al binary compounds to discovering new hydrogen-rich structures in the Li-La-H ternary system, this method demonstrates remarkable efficiency and reliability in high-throughput CSP tasks. 🔗 https://lnkd.in/gaC_RZ9X #DeePMD #DPGEN #CrystalStructurePrediction #MachineLearning #CSP #PhysicalReviewB #MaterialsScience #DeepModeling

2