Multi-modal Large Unified Models Finally Surpass Specific Single-modal Models
Humans perceive, cognize, and generate emotions and consciousness through the integration of multiple senses. Gemini is also practicing this approach, processing multiple modal inputs, integrating them in the brain, and then expressing through various modal outputs. This comprehensive "simulation" of human intelligence by such models is rapidly evolving.
Previously, multi-modal model training resembled a system composed of separate eyes, ears, arms, and brains, lacking strong coordination. However, the direction represented by Gemini feels significantly different: it's as if the large model has become a complete digital person, where hands, eyes, brain, and mouth work in harmonious silicon unity. Gemini is the first true end-to-end multi-modal system.
In the past, models optimized for a single modality usually outperformed those handling multiple modalities simultaneously. The common practice was single-modality model training. Even GPT-4 primarily "concatenates" different modalities into an overarching framework, rather than being a unified multi-modal model.
The exciting aspect of Gemini is that it was designed from the start as a native multi-modal architecture. The training process interweaves various modal data from the beginning. If previous large models were like attaching sensory organs or mechanical arms to a brain externally, Gemini is like growing its own eyes, ears, and arms internally, allowing for fluid and natural interaction.
Whether in terms of model architecture, training process, or final output, Gemini achieves a seamlessly integrated multi-modal experience.
For the first time, Gemini demonstrates that a unified model can handle all modalities, and perform even better than models focused on a single modality! For example, compared to the Whisper model, which is optimized for voice recognition, Gemini shows a significant improvement in accuracy.
This signifies the dawn of the era of unified multi-modal models.
In fact, Gemini is not the first model to demonstrate that different modalities can mutually enhance performance. This was also evident in PaLM-E, where "PaLM-E, trained across different domains including general vision-language tasks at internet scale, showed a marked improvement in performance compared to models performing single tasks in robotics."
Another example of modalities enhancing each other is the multilingual processing ability of large language models. If we consider different languages as distinct "modalities," the practice of large language models has proven that processing native data of all languages together (through tokenization and embedding) managed to lead to the successful construction of a human language tower of Babel.
The overwhelming amount of English data in the training of large language models also benefits the model's understanding and generation of languages with limited data, reaffirming the transfer of linguistic knowledge. It's akin to a person skilled in tennis also being able to improve their abilities in squash or golf through related skills.
Since the rise of large models in February this year, many have gradually embraced the belief that "unified multi-modal models will surpass single-modality models." However, this belief hadn't been confirmed on a large scale until Google's Gemini showcased the prospects of this belief, reshaping and solidifying it for many.
In the future, specialized models for tasks like voice recognition or machine translation may become less significant. Many generative tasks such as TTS and image generation are also likely to be unified under large models. Some may complain about the high cost and slow speed of large unified models, but these are purely technical challenges. In practice, we can distill unified models to specific modalities or scenarios.
We firmly believe that unified cross-modal large models will become the mainstream pathway to achieving AGI.
Furthermore, "modalities" are not just sound, images, videos, etc. Olfactory, gustatory, tactile, temperature, and humidity sensors are also different modalities for gathering environmental information, all of which can in time be encompassed by unified models.
Ultimately, various modalities are merely carriers of "information." They are a form of rendering, a presentation style, a means for an intelligent entity to interact with the physical world. In the eyes of a unified model, all modalities internally can be represented by unified multi-dimensional vectors, enabling cross-modal knowledge transfer and the intersection, alignment, fusion, and reasoning of information.
When the barriers between modalities are breached, revealing the core beneath various renderings, we see the origin of cognition — language.