Advancing AI: Google Brain and University of Alberta Unleash Power of Memory Augmented Large Language Models

Memory Augmented Large Language Models (MALLMs) are revolutionizing the way technology interacts with natural language, altering our conception of machine learning, and pushing the boundaries of computational capability. This is not only apparent in the advent of sophisticated transformer-based models like GPT-2 and GPT-3 but also with interactive language platforms such as ChatGPT. These platforms…

Written by

Casey Jones

Published on

July 5, 2023
BlogIndustry News & Trends

Memory Augmented Large Language Models (MALLMs) are revolutionizing the way technology interacts with natural language, altering our conception of machine learning, and pushing the boundaries of computational capability. This is not only apparent in the advent of sophisticated transformer-based models like GPT-2 and GPT-3 but also with interactive language platforms such as ChatGPT. These platforms have sparked widespread popularity by introducing in-context learning and chain-of-thought prompting, innovative strategies to complement Artificial Intelligence.

However, the golden age of Large Language Models (LLMs) is not devoid of limitations. Presently, transformer-based models are restricted to conditioning just a fixed length of input strings, commonly reducing their computational scope. Besides, contemporary models are beset with various unanswered questions, especially concerning the viability of employing an external feedback loop to enhance computational scope.

Recognizing the need to navigate these complexities, advancements in collaborative research led by Google Brain and the University of Alberta have shed light on potential solutions. One avenue of interest involves introducing an external read-write memory to Large Language Models (LLMs), creating a mechanism that could emulate any algorithm on any input.

Outlined extensively in their research paper, “Memory Augmented Large Language Models are Computationally Universal,” the Google Brain and University of Alberta researchers detail the significant progress made using the Flan-U-PaLM 540B model. The innovative model, coupled with external associative memory, imitates the working of a simple stored instruction computer, placing it at the forefront of computational linguistics.

Connecting the LLM and associative memory is conducted through a loop engaging regular expression matches. Essentially, this loop plays a similar role to a dictionary, increasing the accuracy in the computational function of the MALLMs.

As part of their methodology, the researchers developed a unique prompt-program, designed to simulate a Turing machine’s execution. Accuracy for each model is then appraised through analyzing a finite number of prompt-result patterns.

The study’s considerable strength lies in its methodology’s simplicity: no additional training or alteration to the pre-existing model is necessary. Their innovative approach places an emphasis on crafting a stored instruction computer that can be programmed with certain prompts.

The implications of this study are immense and the breakthrough is forecasted to usher in a new age of language models. By contrast to previous frameworks, this modern approach opens up a wealth of future opportunities, enriching the prospects of tackling complex computational tasks with state-of-the-art language models. In essence, this fusion of Google Brain and University of Alberta is a testament to the limitless potential of Memory Augmented Large Language Models and the transformative power they wield in the realm of machine learning. The journey of artificial intelligence is thus set to become even more fascinating, guided by machinery that can learn, evolve, and flaunt a language finesse that was once deemed far beyond its reach.