0 CommentsAt the Zuckerman Institute on Thursday, researchers and Columbia professors Dr. Rudy Behnia and Dr. Shuran Song spoke about their respective fields of animal vision and computer perception, and how we can apply those ideas to neuroscience as a whole.
The brain is an enigma. An immensely complex organ, it is capable of taking in vast amounts of information and processing it into the form of our thoughts and senses. Can we really hope to understand such a mystery? Some of Columbia’s foremost professors and researchers believe that this is a real possibility.
On Thursday at Columbia’s Zuckerman Institute, researcher Dr. Rudy Behnia spoke about her work involving the study of how human and animal brains interpret and create color vision. Dr. Shuran Song lectured on how neuroscience could be applied to improve robot and AI learning. The event, part of the Stavros Niarchos Foundation Brain Science Lecture series, consisted of two 15-minute talks and a short Q&A and discussion segment and was moderated by postdoc Dr. Vassiki Chauhan.
Dr. Behnia first spoke on her research concerning how the brain processes dynamic information from the surrounding environment, using fruit flies as a model. Humans are distinctly visual beings: we have evolved to be so in order to best ensure our survival and serve our purposes. But we are far from the only species that effectively use vision.
In the animal kingdom, there is a huge diversity in how the vision of various creatures has evolved in response to their environments. For example, prey animals often have binocular vision: they can see almost 360 degrees around themselves. On the other hand, predators usually have monocular vision—although their eyes can’t see as wide, they have much better depth perception and clarity. Some species have evolved to possess incredibly unique forms of vision, such as monarch butterflies, who can see polarized light to help them navigate better, and bees, who can intake ultraviolet light to accurately identify flowers and locate nectar.
Dr. Behnia’s lab focuses specifically on color vision. The different ways that species see color can provide a pathway to a thorough examination of the brain. This is an idea that can be seen through the example of the color purple. Purple is a color that almost everyone can identify, but it does not technically exist in nature. There is no such thing as a purple photon—a particle of light—instead, an object can be labeled as “purple” if it reflects both red and blue photons. So how is the brain able to come up with the idea of the color purple?
It turns out that our eyes separate colors and photons into different channels, in the process losing the identity of the colors due to lack of information. The brain then creates the idea of the colors by analyzing the relative activation of the channels using specific neurons. Dr. Behnia claims that looking further into these neurons can help us better understand color perception and the brain as a whole.
Her current project involves using fruit flies as a model. Although this may seem incomprehensible at first, it turns out that the flies are actually excellent subjects. They are cheap, grow quickly, have incredibly compact brains, and actually possess more photoreceptors than humans. Due to these factors, scientists at the lab have been able to make a neural map of the fruit fly brains, tracking how they intake visual information and which neurons are involved in the process. They have recently been able to use microscopy to further analyze how the brains are responding.
Currently, they have found that fruit fly brains possess very similar circuits to humans and that they also use relative activation to create the idea of color. The lab has even narrowed it down to three specific sets of neurons that are responsible for processing mixes of light. However, although it is very possible that humans have a similar mechanism of neurons, Dr. Behnia advocated for more research into the subject by looking at a greater diversity of species, something she hopes will let us finally unravel this mystery.
Dr. Shuran Song then stepped up to present the second lecture of the event. Her topic of discussion was how her lab is building and improving computer vision programs to let machines better see and interpret the world. She also discussed the intersections between her field of computer science and the overarching ideas of brain science and neuroscience. Living beings interpret the world around them through the general process of “see, understand, act.” Computers and robots use a similar process in their algorithms, but it can be better described as “intake from a sensor, construct a representation of the environment, perform action.”
The transition from getting information from a sensor to actually understanding the surrounding environment is what is known as perception, or “computer vision.” Currently, most robots are not very good at going through this step. They are what is known as “passive observers,” or as Dr. Song described them, “couch potato robots.” Because they do not actively interact with their environment, their understanding of the surroundings is incredibly limited. However, this is beginning to change. Machine learning and the rise of big data have accelerated the development of this field of research and Dr. Song’s lab is at the vanguard of those attempting to address this issue.
The goal essentially lies in the attempt to endow robots and computers with the ability to conduct actions in order to sense and plan for the future, abilities that most animals already possess. This is what is known as “interactive perception,” and it consists of three major parts. The first is obtaining additional considerations. Some objects, until they are interacted with, cannot be fully understood. For example, a deformable object cannot be entirely evaluated unless a robot is able to interact with it.
In a similar vein, the second part is discovering physical properties. By learning to actively observe the environment, robots can obtain information such as mass, friction, and more. The Song Lab has already put significant work into development concerning this part. They have created the DensePhysNet program, which lets robots interact with objects and attempt to predict their properties, learning to use certain sequences of actions to achieve better representations of those objects. Among the robots involved is TossingBot, which, exactly as the name suggests, throws objects into containers. This surprisingly complex action lets the robot deduce properties such as mass distribution and aerodynamics.
The last component of interactive perception is self-supervised learning. Often, the hardest part of training a robot consists of getting data for them to learn from. By learning to explore, robots will hopefully be able to acquire their data, training themselves in the process. In short, through her research, Dr. Song hopes to convert robots and their vision from being passive observers to being active explorers.
The talks ended with a short discussion from the moderator on the connections between these two very distinct, but also very related fields, and what we can learn from both of them. After taking a few questions from the virtual and physical audiences, the event concluded with a summary and closing remarks from the three panelists.
Loading …