Cross-modal Knowledge Transfer of Large Models Proven (Gemini Notes 1/8)

by Zhi-Fei Li, Gao Jia, Wei Li, from "Brother Fei on AI"


In 1948, inspired by psychiatric patients, British doctor Ross Ashby invented a peculiar machine called the "Homeostat." He proclaimed that this device, costing about 50 pounds, was "the closest thing to an artificial brain ever designed by mankind." The Homeostat utilized four bomb control switch gear devices from the British Royal Air Force, used during World War II, as its base. Above these were four cubic aluminum boxes, with the only visible moving parts being four small magnetic needles on top of the boxes, swaying like compass needles in a small trough of water.

When the machine was activated, the needles moved in response to the electric current from the aluminum boxes. The four magnetic needles were always in a sensitive and fragile state of balance. The sole purpose of the Homeostat was to keep the needles centered, maintaining a "comfortable" state for the machine.

Ashby experimented with various methods to make the machine "uncomfortable," such as reversing the polarity of the electrical connections or the direction of the needles. However, the machine always found ways to adapt to the new state and re-center the needles. Ashby described the machine as "actively" resisting any disturbances to its balance through synaptic action, performing "coordinated activities" to regain equilibrium.

Ashby believed that one day, such a "primitive device" could evolve into an artificial brain more powerful than any human, capable of solving the world's most complex and challenging problems.

Despite Ashby's lack of knowledge about today's AGI evolution and the laughable idea of using four small magnetic needles as sensors for intelligence, his Homeostat fundamentally challenged everyone's understanding of "intelligence" - isn't intelligence the ability to absorb information from the environment in various modalities, and to modify behavior and responses based on feedback?

From the peculiar "Homeostat" to today, 75 years later, Google's Gemini, which claims to have surpassed human multi-modal task processing abilities, accelerates towards the evolution of billions of years of carbon-based intelligence through the injection of multi-modal native big data.

The acceleration speed of machine intelligence evolution today far exceeds our imagination. A year ago, OpenAI overturned Google's long-established AI position with its 'brute force aesthetic,' having constructed the Babel Tower of human languages. A year later, Google countered with Gemini, via a 'fight fire with fire' approach to building the first unified cross-modal model, setting another milestone in AGI evolution.

Despite initial skepticism over exaggerated video demos upon Gemini's release, it's undeniable that the dawn of a unified multi-modal approach is shining. What capabilities does Gemini confirm? How will Google's wheels of fate turn? Is time a friend to OpenAI or Google? What does multi-modality mean for Agents and embodied intelligence? Are the foundations for the emergence of AGI with consciousness already in place? How should we view the implications of Gemini for the AI future?


Cross-modal Knowledge Transfer of Large Models Proven Again

For humans, the ability to transfer knowledge across various domains and through different timespaces is more important than merely learning skills. If machines can master cross-modal knowledge transfer, they edge closer to "intelligence generality."
In July this year, Google introduced RT-2, a robotic system based on large models, sparking hope for general-purpose robots.  The system's robotic arm, leveraging the "common sense" of language models, demonstrated the ability to "pick up an extinct animal from a table," moving from common sense reasoning to robotic execution, showcasing cross-modal knowledge transfer. 
In December, the introduction of Gemini by this tech giant reaffirmed the cross-modal knowledge transfer capability of large models: the "common sense" of language models could be transferred to the training of other non-linguistic modalities added later. Language models are known to form the foundation of cognitive intelligence, and the most basic form of cognitive intelligence is "common sense."  Without common sense empowerment, the practical application of large multi-modal models would be challenging.  Gemini smoothly transfers this "common sense" to downstream multi-modal tasks.  Like RT-2, it achieves cross-modal integration through the transfer of text-derived knowledge — Gemini can connect ontology concepts to the understanding of auditory and visual objects, and eventually link them with action, forming an intelligent system ready for real world application. 
From the perspective of model training, compared to language models trained with massive internet data, downstream models (like robotic models) can be trained with very limited data through knowledge transfer.  This transfer-based training manages to address the long-standing issue of data scarcity in downstream applications.  For instance, to achieve the effects shown in the video (which raised doubts about Gemini's video comprehension or picture comprehension but did not affect the discussion on cross-modal knowledge transfer here), Gemini first needs some ontological knowledge — it understands the concept of a duck, knows the usual color of ducks, and what blue is. When it sees a "blue duck," it reacts similarly to humans, expressing the "common sense" that "blue ducks are uncommon." 
Gemini, through auditory and visual perception, identifies that the material of the blue duck is rubber and knows that rubber's density is less than water's. Based on this common sense and reasoning, when it hears a squeaking sound, it can predict that "the blue duck can float on water." 
From RT-2 to Gemini, we've moved to the "fusion" of multi-modal perceptual intelligence and cognitive intelligence. We've transitioned from isolated "five senses" modules of eyes, ears, mouth, nose, and body to a unified digital "human". 
Doesn't this imply that on the path to simulating human intelligence, the unified model is the right approach? 




(Gemini Notes Series to be continued)


Original from:

关于 Google Gemini 的八点启示

by Zhi-Fei Li, Gao Jia, Wei Li, from "Brother Fei on AI"



立委博士,问问副总裁,聚焦大模型及其应用。Netbase前首席科学家10年,期间指挥研发了18种语言的理解和应用系统,鲁棒、线速,scale up to 社会媒体大数据,语义落地到舆情挖掘产品,成为美国NLP工业落地的领跑者。Cymfony前研发副总八年,曾荣获第一届问答系统第一名(TREC-8 QA Track),并赢得17个小企业创新研究的信息抽取项目(PI for 17 SBIRs)。


您的电子邮箱地址不会被公开。 必填项已用 * 标注