The AI laboratory DeepMind developed the Flamingo family of models, capable of handling a larger workload with less costly and labour-intensive training.
Introducing Flamingo 🦩: a generalist visual language model that can rapidly adapt its behaviour given just a handful of examples. Out of the box, it’s also capable of rich visual dialog.
Read more: https://t.co/xEzqTizoJQ 1/ pic.twitter.com/GjlnDzbyOQ
— DeepMind (@DeepMind) April 28, 2022
The model is designed to combine text and image inputs to produce a text-only answer.
Flamingo was trained on a dataset created for multimodal machine-learning research. The dataset consists of 185 million images and 182 GB of text sourced from the public internet.
One of Flamingo’s components is the pre-trained language model Chinchilla LM with 70 billion parameters. DeepMind merged the algorithm with elements of visual learning. The engineers also added ‘intermediate components of the new architecture’ that keep data isolated and frozen, giving it an 80‑billion-parameter Flamingo VLM.
“A single Flamingo model can achieve top results across a wide range of tasks, competing with approaches that require task-specific fine-tuning on larger amounts of data,” the developers said.
According to the organisation’s representatives, Flamingo outperforms previous multi-step training approaches. The model is also more effective than fine-tuned algorithms that rely on larger amounts of data.
In the long term, Flamingo could reduce energy consumption during AI training and lessen the need for high-performance hardware. However, the company did not disclose details of how they achieved such results.
The developers emphasised that Flamingo can be quickly adapted to resource-constrained settings and to low-resource tasks such as assessing AI bias.
Earlier in April, DeepMind unveiled the 70‑billion‑parameter language model Chinchilla.
In February, the British AI lab demonstrated the AlphaCode, which can write code on its own.
In December 2021, DeepMind developed the large language model Gopher, containing 280 billion parameters.
Subscribe to ForkLog’s news on Telegram: ForkLog AI — all the news from the world of AI!