In a bid to optimize content generation for uses such as natural language processing (NLP), researchers from Meta, Facebook’s parent company, have put forward a novell approach. The MegaByte approach, unveiled in a recently published paper, is poised to improve the creation of lengthier content and potentially make transformer-based architectures obsolete.
The MegaByte architecture distinguishes itself by employing a multi-scale decoder capable of modeling sequences exceeding one million bytes with end-to-end differentiability. Unlike models like OpenAI’s ChatGPT that excel with short outputs but falter with longer or intricate sequences, MegaByte aspires to deliver superior generation performance while simultaneously curbing running costs.
Transformers on the Chopping Block?
The research team at Meta has expressed discontent with transformer-based architectures, a mainstay in the field of NLP since its inception by Google researchers in 2017. While such systems, including renowned models like ChatGPT, GPT-4, and BERT, have found broad acceptance for NLP tasks, they are deemed resource-intensive when working on complex inputs such as books or podcasts.
MegaByte’s innovative approach involves dividing inputs and outputs into “patches” rather than individual tokens. Each patch is accorded its localized response, which the model consolidates with other patches to create the final output. This ‘patches’ technique bypasses the hurdles of self-attention scaling and reduces computation time as calculations are performed parallelly, not sequentially.
Endorsed by Industry Leaders
Meta’s newly proposed architecture garnered commendation from Andrej Karpathy, Tesla’s AI director, who labelled it as “promising“. According to Karpathy, the prospect of bypassing tokenization in large language models should be a common aspiration, as byte-level sequences created naively are overly lengthy.
Despite the initial excitement, MegaByte is still in its nascent stages. The experiments conducted using it, as detailed in Meta’s paper, are significantly smaller in scale than those of current state-of-the-art language models. The team behind MegaByte anticipates future research to explore scaling the architecture to larger models and datasets.