What Led To The Evolution Of Modern Human Eyes?
Although scientists are unable to travel back in time to examine the environmental factors that influenced the development of the many vision systems seen in nature, they are able to investigate this evolution in artificial intelligence agents thanks to a novel computational framework created by MIT researchers.
Similar to a “scientific sandbox,” the framework they created allows researchers to replicate various evolutionary trees as embodied AI creatures evolve eyes and learn to see over many generations. By altering the world’s structure and the activities AI agents perform, such identifying items or locating food, the user accomplishes this.
This enables them to investigate the reasons behind the evolution of simple, light-sensitive patches as eyes in one species and complex, camera-like eyes in another.
Using this paradigm, the researchers’ tests demonstrate how tasks influenced the agents’ eye evolution. For example, they discovered that the evolution of complex eyes with numerous separate units, such as those of insects and crustaceans, was frequently prompted by navigation requirements.
Conversely, agents were more likely to develop camera-like eyes with irises and retinas if they concentrated on object discrimination.
With the help of this approach, researchers may be able to investigate “what-if” scenarios pertaining to vision systems that are challenging to investigate empirically. It might also serve as a roadmap for the development of innovative sensors and cameras for wearable technology, robots, and drones that strike a balance between functionality and practical limitations like manufacturing feasibility and energy efficiency.
Even while we will never be able to go back in time and fully understand how evolution occurred, we have constructed a setting in this study that allows us to virtually reproduce evolution and investigate the environment in a variety of ways. Kushagra Tiwary, a graduate student at the MIT Media Lab and co-lead author of a report on this topic, says, “This way of doing science opens up a lot of possibilities.”
Co-lead author and fellow graduate student Aaron Young, graduate student Tzofi Klinghoffer, former postdoc Akshat Dave, who is currently an assistant professor at Stony Brook University, Tomaso Poggio, the Eugene McDermott Professor in the Department of Brain and Cognitive Sciences, an investigator at the McGovern Institute, and co-director of the Center for Brains, Minds, and Machines; co-senior authors Brian Cheung, a postdoc in the Center for Brains, Minds, and Machines and an incoming assistant professor at MIT; and others from Rice University and Lund University. The study was published in Science Advances today.
Constructing A Sandbox For Science
The study started out as a discussion among the researchers about finding novel vision systems that might be used to many domains, such as robots. The researchers chose to employ AI to investigate the various evolutionary possibilities in order to test their “what-if” hypotheses.
“I was driven to pursue science as a child by what-if concerns. We have a special chance to develop these embodied creatures with AI, which enables us to pose questions that are typically unanswerable, according to Tiwary.
The researchers transformed every component of a camera, including sensors, lenses, apertures, and CPUs, into parameters that an embodied AI agent could learn in order to create an evolutionary sandbox.
These basic blocks served as the foundation for an algorithmic learning mechanism that an agent would employ as its eyes changed over time.
We were unable to model the entire universe atom by atom. Determining which ingredients we required and which we didn’t, as well as how to distribute resources among those many components, was difficult, according to Cheung.
According to their framework, this evolutionary algorithm may decide which elements to evolve depending on the agent’s task and the environment’s limitations.
Every setting has a specific goal, such as monitoring prey, identifying food, or navigating, that is intended to replicate actual visual challenges that animals must solve in order to survive. A single photoreceptor that scans the environment and a corresponding neural network model that interprets visual data are the agents’ initial configurations.
Each agent is then trained throughout the course of its lifetime via reinforcement learning, a trial-and-error method in which the agent is rewarded for achieving the task’s objective. Constraints such as a pixel count limit for an agent’s visual sensors are also incorporated into the environment.
In the same way that physical limitations in our environment, such as the laws of light, have influenced the design of our own eyes, Tiwary claims that these limitations also influence the design process.
Agents develop many aspects of reward-maximizing vision systems across many generations.
Their system simulates evolution computationally by using a genetic encoding process, in which the development of an agent is controlled by the mutation of individual genes.
Morphological genes, for example, influence the location of the eyes and how the agent perceives its surroundings; optical genes control the amount of photoreceptors and how the eye interacts with light; and neural genes regulate the agents’ ability to learn.
Testing Theories
Upon conducting trials inside this framework, the researchers discovered that tasks significantly impacted the agents’ evolving vision systems.
Agents working on navigation tasks, for example, developed eyes that maximized spatial awareness through low-resolution sensing, whereas agents working on object detection produced eyes that prioritized frontal acuity over peripheral vision.
When it comes to processing visual information, a larger brain isn’t always better, according to another investigation. Physical limitations, such as the quantity of photoreceptors in the eyes, limit the amount of visual data that can enter the system at once.
According to Cheung, “a larger brain eventually does not help the agents at all, and in nature that would be a waste of resources.”
The best vision systems for particular applications will be investigated by the researchers using this simulator in the future, which may aid in the creation of task-specific sensors and cameras. To let users explore more options and pose “what-if” questions, they also wish to incorporate LLMs into their framework.
“There is a genuine advantage to posing more creative queries. Cheung says, “I hope this encourages others to develop more expansive frameworks, where they are looking to answer questions with a much wider scope, rather than concentrating on specific questions that cover a particular area.”
The Center for Brains, Minds, and Machines and the Defense Advanced Research Projects Agency (DARPA) Mathematics for the Discovery of Algorithms and Architectures (DIAL) program provided some funding for this work.