Syntiant Corp.’s Post

Learn how to compress models with Hugging Face Transformers library and Quanto library with this interesting short course from deeplearning.AI! Check it out here: https://lnkd.in/daGR6re6 #LLM #huggingface #QLora

I’m thrilled to share the latest issue of The AI Blueprint, packed with cutting-edge AI developments and practical applications! 🌐 Erasing Clouds from Satellite Imagery Using GANs Discover how Generative Adversarial Networks (GANs) are revolutionizing remote sensing by removing clouds from satellite images, ensuring clearer and more accurate data for various applications. 🔍 An LLM Journey: From POC to Production Dive into the challenges and solutions involved in taking large language models from proof of concept to full-scale production. Essential insights for anyone working with LLMs! 🍎 Apple Intelligence Debuts at WWDC Apple's latest AI-driven features for iPhone, iPad, and Mac promise to transform our interaction with technology. Learn about the exciting new capabilities introduced at WWDC. 📊 7 Ways to Monitor Large Language Model Behavior Explore effective strategies for monitoring LLM behavior to ensure reliability and accuracy in your AI applications. 🛠 Effective Strategies for Managing ML Initiatives Optimize your machine learning projects with practical strategies to streamline processes and improve outcomes. Stay ahead in the AI game by exploring these fascinating topics and more in this week’s issue! 📈 Read the full newsletter here: https://lnkd.in/d-yVsjj3 Stay innovative, Avi

#Apple's getting into #AI! 🍏 They're adding #generativeai features to their software platforms, coming later this year. No flashy details yet, but it seems like a move to keep up with Google and Samsung.🤖 ▶Read more: https://lnkd.in/eU3Q3yc3 Follow our page Start With WCPGW #AppleAI #ChatGPT #PrivacyControl #TechNews #chip #chipsecurity #semiconductorsecurity #chip #SemiconductorTechnology #AIethics #RiskManagement #AI #AIRisks #AIethics #HardwareSecurity #Semiconductor #Chips #WCPGW #PrivacyTalk #DataAwareness

1/ US-startup Useful Sensors has developed Moonshine, an open-source speech recognition model optimized for real-time applications on resource-constrained hardware, achieving up to five times faster performance than OpenAI's Whisper. 2/ Moonshine scales processing time proportionally to audio input length, eliminates overhead from zero-padding shorter data, and maintains Whisper-like accuracy at a reduced computational cost, despite having a smaller model size. 3/ Moonshine models show strong performance in benchmarks, slightly outperforming Whisper in word error rate, but with room for improvement on very short audio segments. It's available as open source.

Heading out to Greece for INTERSPEECH 2024. Below are the talks and posters from NVIDIA AI NeMo team. We have cool works about fast RNN-T inference, Canary-1B, ASR, AST, TTS, speech codecs, enhancement, and SpeechLLM. 1️⃣Monday 🎙️12:30 Speed of Light Exact Greedy Decoding for RNN-T Speech Recognition Models on GPU 📜15:00 Fast Context-Biasing for CTC and Transducer ASR models with CTC-based Word Spotter 2️⃣Tuesday 🎙️10:40 Schrödinger Bridge for Generative Speech Enhancement 🎙️17:40 Codec-ASR: Training Performant Automatic Speech Recognition Systems with Discrete Speech Representations 3️⃣Wednesday 📜13:30 Improving Robustness of LLM-based Speech Synthesis by Learning Monotonic Alignment 📜16:00 Less is More: Accurate Speech Recognition & Translation without Web-Scale Data 📜16:00 Instruction Data Generation and Unsupervised Adaptation for Speech Language Models 📜16:00 DeSTA: Enhancing Speech Language Models through Descriptive Speech-Text Alignment Looking forward to meeting everybody F2F!

Large language models are getting bigger and bigger, demanding more energy and computational power. For LLMs that are cheap, fast, and environmentally friendly, they’ll need to shrink, ideally small enough to run directly on devices like cellphones. Researchers have long compressed networks by reducing the precision of parameters—a process called quantization—so that instead of taking up 16 bits each, they might take up 8 or 4. Now researchers are pushing the envelope to a single bit. #llm #quantization

🤖 AI Insights: Updates from TechRepublic's AI Feed #TechRepublic #AI 🎩

The previous post was about hallucinations in speech to text model. Here is NVIDIA blogpost about hallucination detection in text to speech model. Do you know any systematic study of hallucinations in text to speech models?

🍎 From everything that Apple announced during the WWDC, I want to single out one feature that I find most interesting. It's the overall visual interface of Apple Intelligence. I felt this was missing from most implementation of AI in consumer technology so far. 🌟 The playground, the scratch math, the genomji, all of these are accessible through distinct use custom interactions and visual cues. To be honest, writing the exact prompt in a text box always felt a bit restrictive. This visual interaction will surely help the adoption of AI for everyday uses. 🥇 Kudos to Apple design for keeping up the practice of refining the design of existing technologies. 🏆 Another +1 to Apple for integrating the AI with the existing assistant, Siri, and not introducing a new assistant alongside the existing one.

🚨 Our paper titled "Exploring Federated Self-Supervised Learning for General Purpose Audio Understanding" has been accepted at the ICASSP-2024 workshop on Self-supervision in Audio, Speech, and Beyond (SASB-2024). In this study, we conducted an initial investigation into various self-supervised learning methods (predictive and feature-matching) within the context of federated cross-device settings. We explored different aggregation methods and examined their impact on downstream audio tasks. Notably, we discovered intriguing patterns, including both strong and weak correlations between federated self-supervised pretext tasks and the centralized downstream audio tasks. Work done with Kin Wai Lau, yuyang xie, Ruby Ma, and Jiajun Shen at TCL AI Lab, TCL Preprint: https://lnkd.in/gx8eK2NN #ICASSP2024 #SASB2024 #federatedlearning #Flowerdev #Audiounderstanding #selfsupervision #SSL