A naukri.com initiative
ML News
Medium
2d
216
Image Credit: Medium
Mastering Prompt Design in Vertex AI: My Journey into the Art of Prompting
- Prompting is essential for designing intelligent conversations with large language models (LLMs) like PaLM and Gemini, impacting the quality of the output.
- The course on Prompt Design in Vertex AI covers foundations of prompt design, strategies like zero-shot and chain-of-thought, best practices, and ethics in generative AI, offering both conceptual and practical knowledge.
- Insights include the importance of prompt clarity, effectiveness of few-shot prompting, benefits of chain-of-thought prompting for reasoning, and the iterative nature of prompt design for optimal results.
- Real-world applications include designing chatbots, improving content generation workflows, enhancing AI system safety, and accelerating prototyping through code generation and summarization prompts, making the skills immediately practical for developers, testers, and product managers.
Read Full Article
13 Likes
Global Fintech Series
2d
232
Image Credit: Global Fintech Series
A Technical Look at Transaction Monitoring Systems in AML Compliance
- Transaction monitoring systems are crucial in AML compliance, helping detect suspicious financial activities like money laundering and fraud.
- These systems analyze customer transactions in real time or batch mode using predefined rules or machine learning algorithms.
- Key components include data integration, customer profiling, rules engine, machine learning, alert management, and reporting.
- Challenges include high false positive rates, data quality issues, evolving threats, regulatory complexity, and explainability of AI models.
- Innovations in transaction monitoring include hybrid models, graph analytics, real-time monitoring, adaptive learning systems, and privacy-preserving techniques.
- Graph analytics help uncover hidden money laundering networks, while real-time monitoring allows faster responses to suspicious activities.
- Adaptive learning systems continuously refine alert accuracy, and privacy-preserving techniques maintain data privacy in AML efforts.
- Modernizing AML infrastructure with machine learning and real-time processing can lead to smarter and more accurate transaction monitoring.
- Financial institutions investing in these technologies can strengthen their defenses and enhance trust with regulators and customers.
- Transaction monitoring systems play a critical role in AML compliance and are evolving to meet the growing sophistication of financial crime.
Read Full Article
14 Likes
Medium
2d
589
Image Credit: Medium
Best AI & ML Training In Bangalore For Working Professional?
- Artificial Intelligence (AI) and Machine Learning (ML) are integral parts of various applications like voice assistants, fraud detection, and self-driving cars.
- AI and ML involve collecting and analyzing large sets of data to make predictions based on patterns identified by machine learning models.
- Professionals from any background can enter the AI ML field without extensive coding knowledge, as beginner-friendly tools and platforms are provided for training.
- Eduleem School of Design & IT in Bangalore offers one of the best AI and ML training programs. It is designed to be practical, industry-focused, and suitable for working professionals aiming for a career change.
Read Full Article
18 Likes
Arxiv
2d
285
Image Credit: Arxiv
LAS: Loss-less ANN-SNN Conversion for Fully Spike-Driven Large Language Models
- Spiking Large Language Models (LLMs) offer an energy-efficient alternative to conventional LLMs through event-driven computation.
- Researchers have been developing ANN-to-SNN conversion methods to create spiking LLMs while maintaining energy efficiency, but face challenges with extreme activation outliers and incompatible operations.
- A new approach called Loss-less ANN-SNN Conversion for Fully Spike-Driven LLMs (LAS) is proposed to address these challenges by introducing novel neurons to handle activation outliers and nonlinear operations.
- Experimental results show that LAS achieves loss-less conversion and even improves accuracy on tasks like the WSC task. Source code for LAS is available at https://github.com/lc783/LAS.
Read Full Article
17 Likes
Arxiv
2d
186
Image Credit: Arxiv
Analog Foundation Models
- Analog in-memory computing (AIMC) is a promising compute paradigm that aims to enhance speed and power efficiency of neural network inference beyond traditional von Neumann-based architectures.
- Challenges like noisy computations and strict input/output quantization constraints hinder the performance of off-the-shelf Large Language Models (LLMs) when deployed on AIMC-based hardware.
- A new method has been introduced to adapt LLMs for execution on noisy, low-precision analog hardware, allowing advanced models to maintain performance comparable to 4-bit weight, 8-bit activation standards despite noise and quantization restrictions.
- The models developed through this approach can also be quantized for inference on low-precision digital hardware, displaying improved scaling behavior compared to models trained with 4-bit weight and 8-bit static input quantization.
Read Full Article
11 Likes
Arxiv
2d
289
Image Credit: Arxiv
Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing
- Graph unlearning is essential for protecting user privacy by removing user data influence from trained graph models.
- Recent developments in graph unlearning methods have focused on maintaining model performance while deleting user information, but changes in prediction distribution across sensitive groups have been observed.
- Study shows that graph unlearning introduces bias, and a fair graph unlearning method, FGU, is proposed to address this issue.
- FGU ensures privacy by training shard models on partitioned subgraphs and fairness by employing a bi-level debiasing process, achieving superior fairness without compromising privacy and accuracy.
Read Full Article
17 Likes
Arxiv
2d
281
Image Credit: Arxiv
Energy-Efficient Federated Learning for AIoT using Clustering Methods
- This study focuses on the energy implications of federated learning (FL) within Artificial Intelligence of Things (AIoT) scenarios.
- The research examines the energy consumed during the FL process, highlighting pre-processing, communication, and local learning as the main energy-intensive processes.
- The study proposes two clustering-informed methods for device/client selection in distributed AIoT settings, aiming to speed up model training convergence.
- Through extensive numerical experimentation, the clustering strategies show high convergence rates and low energy consumption compared to other recent approaches in the literature.
Read Full Article
16 Likes
Arxiv
2d
174
Image Credit: Arxiv
Training Deep Morphological Neural Networks as Universal Approximators
- Researchers investigate deep morphological neural networks (DMNNs) and emphasize the importance of activations between layers.
- They introduce new architectures for DMNNs with different parameter constraints, showcasing successful training and improved pruning capabilities compared to linear networks.
- This study is the first successful attempt to train DMNNs under specific constraints, although the networks' generalization capabilities are limited.
- Additionally, a hybrid network architecture combining linear and morphological layers is proposed, demonstrating faster convergence of gradient descent with large batches.
Read Full Article
10 Likes
Arxiv
2d
150
Image Credit: Arxiv
Out-of-distribution generalisation is hard: evidence from ARC-like tasks
- Out-of-distribution generalisation is crucial for human and animal intelligence.
- To achieve OOD through composition, an intelligent system must identify task-invariant input features and composition methods.
- Testing on an OOD setup is not sufficient; confirming that features are compositional is also essential.
- Exploration of tasks shows that some neural networks struggle with OOD, while novel architectures with appropriate biases can be successful.
Read Full Article
9 Likes
Arxiv
2d
313
Image Credit: Arxiv
Robust Federated Learning with Confidence-Weighted Filtering and GAN-Based Completion under Noisy and Incomplete Data
- Federated learning (FL) offers a solution for collaborative model training while preserving data privacy in decentralized client datasets.
- Challenges like noisy labels, missing classes, and imbalanced distributions affect the effectiveness of FL.
- A new methodology is proposed to address data quality issues in FL by enhancing data integrity through noise cleaning, synthetic data generation, and robust model training.
- Experimental evaluations on MNIST and Fashion-MNIST datasets show improved model performance, especially in noise and class imbalance scenarios, ensuring data privacy and practicality for edge devices.
Read Full Article
18 Likes
Arxiv
2d
195
Image Credit: Arxiv
A Generative Neural Annealer for Black-Box Combinatorial Optimization
- A generative, end-to-end solver for black-box combinatorial optimization has been proposed.
- Inspired by annealing-based algorithms, a neural network is trained to model the Boltzmann distribution of the black-box objective.
- The network's conditioning on temperature allows capturing a range of distributions, aiding in global optimization and improving sample efficiency.
- The approach shows competitive performance on challenging combinatorial tasks with limited or unlimited query budgets.
Read Full Article
11 Likes
Arxiv
2d
56
Image Credit: Arxiv
Community-based Multi-Agent Reinforcement Learning with Transfer and Active Exploration
- Proposed a new framework for multi-agent reinforcement learning (MARL) with agents cooperating in a time-evolving network with latent community structures and mixed memberships.
- Community-based framework allows agents to belong to multiple overlapping communities, maintaining shared policy and value functions aggregated based on personalized membership weights.
- Designed actor-critic algorithms that exploit this structure, enabling structured information sharing without needing access to other agents' policies.
- The framework supports transfer learning via membership estimation and active learning by prioritizing uncertain communities during exploration, with convergence guarantees established under linear function approximation.
Read Full Article
3 Likes
Arxiv
2d
321
Image Credit: Arxiv
Self-Consuming Generative Models with Adversarially Curated Data
- Recent advances in generative models have made it hard to differentiate between real and synthetic data.
- Self-consuming loops in training with synthetic data can lead to model collapse or instability.
- Data curation based on user preferences can drive models to optimize those preferences, leading to converging distributions.
- Study explores the impact of noisy and adversarially curated data on generative models, proposes attack algorithms for adversarial scenarios, and conducts experiments to demonstrate algorithm effectiveness.
Read Full Article
19 Likes
Arxiv
2d
329
Image Credit: Arxiv
Lossless Compression for LLM Tensor Incremental Snapshots
- During Large Language Models (LLMs) training, a significant amount of tensor data is checkpointed periodically for recovery purposes in case of failure.
- The paper focuses on optimizing the checkpointing process by analyzing checkpoint data and maximizing the use of lossless compression techniques to reduce the data volume.
- An effective compression solution named Language Model Compressor (LMC) has been developed, based on byte-grouping and Huffman encoding, offering better performance than existing alternatives like BZ2 with significantly reduced compression time.
- LMC's 16-core parallel implementation achieves high compression and decompression throughput, leading to reduced CPU resources and enabling higher-frequency checkpoints during model training.
Read Full Article
19 Likes
Arxiv
2d
206
Image Credit: Arxiv
Adversarial Attack on Large Language Models using Exponentiated Gradient Descent
- Large Language Models (LLMs) are widely used but still vulnerable to jailbreaking attacks despite being aligned using techniques like reinforcement learning from human feedback (RLHF).
- Existing adversarial attack methods on LLMs involve searching for discrete tokens or optimizing the continuous space represented by the model's vocabulary.
- A new intrinsic optimization technique, using exponentiated gradient descent with Bregman projection, has been developed to ensure optimized one-hot encoding stays within the probability simplex.
- This technique has proven to be more effective in jailbreaking LLMs compared to other state-of-the-art techniques, as demonstrated on five LLMs and four datasets.
Read Full Article
12 Likes
For uninterrupted reading, download the app