-
Training Large-Vocabulary Neural Language Model by Private Federated Learning for Resource-Constrained Devices
Federated Learning (FL) trains models using distributed data. Differential Privacy (DP) provides privacy guarantees. The goal is to train a large neural network language model (NNLM) on compute-constrained devices while preserving privacy using FL and DP. However, DP-noise increases as model size grows, hindering convergence. Partial Embedding Updates (PEU) is proposed to decrease noise by…
-
Leveraging Large Language Models for Exploiting ASR Uncertainty
Large language models (LLMs) excel at text-based natural language processing tasks through creative prompt engineering and in-context learning. However, their performance on spoken language understanding (SLU) tasks relies heavily on speech-to-text conversion by an off-the-shelf automation speech recognition (ASR) system, constraining their accuracy in this setup.
-
Advancing Speech Accessibility with Personal Voice
Introduced in May 2023 and available on iOS 17 in September 2023, Personal Voice is a voice replicator tool designed for individuals at risk of losing their ability to speak, such as those with ALS. It creates a synthesized voice for use in FaceTime, phone calls, assistive communication apps, and in-person conversations, supporting speaking ability.
-
Imprisoned ex-PM Imran Khan appears via AI-generated rally
Former Prime Minister of Pakistan, Imran Khan, utilized AI to deliver a four-minute speech at a virtual rally while in prison. The AI-generated voice closely resembled his own, delivering a message of resilience and defiance against political constraints faced by his party. The rally gained over five million views despite reported internet outages. AI’s political…
-
The Ultimate Guide to Training BERT from Scratch: Final Act
This blog post serves as the conclusion to a series on training BERT from scratch. It discusses the significance of BERT in Natural Language Processing, reviews the previous parts of the series, and outlines the process of building and training a BERT model. The post emphasizes understanding the model’s inner workings and shares insights on…
-
2023 in Review: Recapping the Post-ChatGPT Era and What to Expect for 2024
The year 2023 saw significant developments in the Generative AI landscape, marked by the release of multiple LLMs and the emergence of LLMOps. While there were challenges in production, it was a year of experimentation and getting to know Generative AI. Looking ahead to 2024, the focus will likely be on successfully deploying Generative AI…
-
Simulating Exoplanet Discoveries with Python
The text is a comprehensive explanation of computer simulations and their applications in understanding and predicting astronomical events. It covers various scenarios of transit phenomena, including exoplanet transits, asteroid belts’ influence, and hypothetical scenarios like simulating an exoplanet with an exomoon and detecting alien megastructures. It also highlights the advantages of simulations in scientific research.…
-
Importance of Smoothness Induced by Optimizers in FL4ASR: Towards Understanding Federated Learning for End-to-End ASR
The paper explores training End-to-End Automatic Speech Recognition (ASR) models using Federated Learning (FL) and its impact on minimizing the performance gap with centralized models. It examines adaptive optimizers, loss characteristics, model initialization, and carrying over modeling setup from centralized training to FL.
-
Bootstrap Your Own Variance
The paper “Bootstrap Your Own Variance: Understanding Model Uncertainty with SSL and Bayesian Methods” was accepted at the Self-Supervised Learning workshop at NeurIPS 2023. It proposes BYOV, combining BYOL SSL algorithm with BBB Bayesian method to estimate model posteriors, showing that BYOV’s predictive standard deviation aligns well with a Gaussian distribution.
-
DataComp: In Search of the Next Generation of Multimodal Datasets
Multimodal datasets play a crucial role in recent AI advancements like Stable Diffusion and GPT-4. However, their design is not as researched as model architectures or training algorithms. To tackle this, DataComp introduces a testbed for dataset experiments using 12.8 billion image-text pairs from Common Crawl, allowing participants to create and evaluate new datasets.