M.Millions of people (many of whom haven’t given much thought to computer science) are experimenting with generative AI models like the very good conversational ChatGPT and the creative image generator DALL-E. While these products don’t reflect as much of a technological breakthrough as AI has penetrated the public consciousness, the traction they’ve found has led to massive investment streams.
For those of us who have long been bullish about the potential of AI to transform society, especially in key areas like health and medicine, the last few months have felt like science fiction.
However, while these capabilities are fun to explore (GPT-4, for example, scored 20 points above passing on the US medical licensing exam), the results largely highlight their shortcomings. With the ability to read, retain, and regurgitate all that data on demand, today’s AI is good at everything, but good at nothing.
There is no doubt that AI is poised to irrevocably change the way we look at disease prevention and treatment. The doctor hands over the documents to his AI clerk. Primary care providers will rely on chatbots for triage. A nearly limitless library of predicted protein structures facilitates drug development. But to truly transform these fields, we need to invest in creating ecosystems of models that learn as the best doctors and drug developers do today (e.g., “expert” AI). I have.
Getting to the top of the field usually starts with years of intensive uploading of information. It is often done through formal schooling, followed by some apprenticeship period. A year dedicated to learning, primarily directly, from the most accomplished practitioners in the field. This is an almost irreplaceable process. For example, much of the information that residents glean from listening to and watching good surgeons is not detailed in any textbook.
Intuition, which is often gained through education and experience, to help determine the best answer in complex situations is especially difficult to acquire. This is true for both artificial intelligence and humans, but for AI the problem is exacerbated by how we learn today and how engineers are currently tackling opportunities and challenges. By examining thousands to millions of labeled data points (“correct” and “wrong” examples), current advanced neural network architectures are able to determine which choices are better than others. You can find out why. AI should be trained using models stacked on top of each other, rather than learning only from large amounts of data and expecting a single generative model to solve all problems. First biology, then chemistry, layering on top of those underlying data points. For example, it is specific to healthcare and pharmaceutical design.
Medical students aspire to become doctors, but their coursework begins with the basics of chemistry and biology rather than the minutiae of diagnosing disease. Without these foundational courses, their ability to one day provide quality health care would be severely limited. Similarly, scientists designing new treatments spend years studying chemistry and biology, followed by a PhD and then working under the guidance of a professional drug designer. This style of learning helps develop a sense of how to navigate decisions involving nuances that really matter, especially on a molecular scale. health effects vary dramatically.
Developing these stacked AI models (simplified maps of complex data that help AI models understand patterns and relationships) with a hierarchy of latent spaces reflects the understanding or predictive ability of each primitive. will be This may initially be comparable to human education and teaching paradigms, but over time it will likely specialize in developing new types of AI learning expertise. These stacked models may develop in a manner similar to the human brain cortex. However, whereas humans have visual and motor areas, AI has biological and drug discovery areas. In both cases, it is a neural architecture specialized for a particular task.
Ironically, creating specialized AI for a particular domain, such as healthcare, may be easier to create than something similar to HAL 9000 with typical human-level knowledge across a wide range of disciplines. there is. In fact, we need domain-specific AI rather than comprehensive AI that can do everything the average human can do. Rather than a single expert AI, many AI models with diverse coding, data, and testing approaches can provide a second opinion (or third opinion, or fourth opinion) if needed. I hope that specialized AI will be created.
In parallel, AI must be uncoupled from online moorings and thrown into the atomic world. We need to equip the most skilled human experts with wearables to capture the nuanced real-world interactions that AI can learn from, as budding academic and industry stars are doing. I have. The most complex and uncertain aspects of dealing with health and medicine do not exist entirely in the world of bits.
Exposing these specialist AIs to the perspectives of top practitioners in various fields is essential to avoid duplicating dangerous biases. But AI is not as black box as the common imagination suggests. The human decision-making that we rely on today is arguably more opaque, as we have said before. Fear of violating human bias cannot limit our willingness to explore how AI can help democratize the expertise of human experts.
Given the neural networks that underpin artificial intelligence, these specialized AIs, through meta-learning (or learning to learn), acquire knowledge even faster than we might expect, bringing us humans on board. There is a possibility. The nature of AI allows it to do things that humans cannot. Break them down bit by bit and study them bit by bit. Create a learning flywheel by building a system that plumbs the inner workings of a specialized AI. Ultimately, specialist AI may move beyond the role of domain expert and into the next generation of specialists (both human and AI).
Other must-read articles from TIME
