© IE Online Media Services Pvt Ltd
Tags:
Journalism of Courage
Mark Zuckerberg’s Meta, on Friday, said that it was releasing a series of new AI models from its research division – Fundamental AI Research (FAIR). These models include a ‘Self-Taught Evaluator’ that could likely offer the possibility of less human involvement in the entire AI development process, and another model that freely mixes text and speech.
The latest announcements come after Meta’s paper in August that detailed how these models would rely on the ‘chain of thought’ mechanism, something which has been used by OpenAI for its recent o1 models that think before they respond. It needs to be noted that Google and Anthropic, too, have published research on the concept of Reinforcement Learning from AI Feedback. However, these are not yet out for public use.
Meta’s group of AI researchers under FAIR said that the new releases support the company’s goal of achieving advanced machine intelligence while also supporting open science and reproducibility. The newly released models include updated Segment Anything Model 2 for images and videos, Meta Spirit LM, Layer Skip, SALSA, Meta Lingua, OMat24, MEXMA, and Self Taught Evaluator.
Meta has termed this new model capable of validating other AI models’ works as “strong generative reward model with synthetic data”.The company claims that this a new method for generating preference data to train reward models without relying on human annotations. “This approach generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces for evaluation and final judgments, with an iterative self-improvement scheme,” the company said in its official blog post.
Essentially, the Self Taught Evaluator is a new method that generates its own data to train reward models with the need for humans to label it. Meta says that the model generates different outputs from AI models and then uses another AI to assess and improve those outcomes. This is an iterative process. According to Meta, the model is powerful and performs better than models that rely on human-labled data such as GPT-4 and others.
The Spirit LM is an open source language model for seamless speech and text integration. Large Language Models are usually used to create systems that convert speech to text and vice versa. However, this could also lead to natural expressiveness being lost from the original speech. Meta has developed Spirit LM, its first open-source model that can work with both text and speech in a more natural way.
“Many existing AI voice experiences today use ASR to techniques to process speech before synthesizing with an LLM to generate text — but these approaches compromise the expressive aspects of speech. Using phonetic, pitch and tone tokens, Spirit LM models can overcome these limitations for both inputs and outputs to generate more natural sounding speech while also learning new tasks across ASR, TTS and speech classification,” Meta said in a tweet.
The Meta LM is trained on both speech and text data, making it possible to switch between the two effortlessly. Meta has created two versions of the model – Spirit LM Base that focuses on speech sounds, and Spirit LM that captures the tone and emotion in a speech such as anger, excitement to make it sound more realistic. Meta claims that this model can create more natural-sounding speech. It also learns tasks like speech recognition, converting text to speech, or classifying different types of speech.
