Why does fruit fly courtship "position" change? New Princeton study: By looking at the neurons behind it, Nature sub-magazine.

The thirteen fish and goat are reported from the foe temple. The official account QbitAI has new ways, breakthroughs and discoveries on the road to crack the brain.in order to explore the mysteries of the brain, scientists have tried various methods: directly observing the living brain → controlling neurons by light pulses → constructing complex devices and virtual reality environment.in fact, as long as we can observe animal behavior accurately and efficiently, we can explore the neurons in the brain.as Bob Datta, a neurobiologist at Harvard Medical School, said: we don't understand the output of the brain. To understand these high-density neural codes, we need to have a deeper understanding of behavior.this study from Princeton University, author Mala Murthy and her team used deep neural network model to successfully predict when and how male flies would sing a "love song" to "heart flies".these methods correspond to different strategies of male Drosophila. Based on this information, the researchers identified the neurons that can control male Drosophila to make different decisions.author Mala Murthy said: This is an important breakthrough.we expect this model to be widely used to link neural activity with natural behavior.this study was also published in nature neuroscience.the neural network for observing behavior, predicting brain state and capturing behavior changes of Drosophila melanogaster is generalized linear model (GLM) + hidden Markov model (HMM).this is an unsupervised approach.during courtship, male flies sing to female flies in three different patterns: one is sinusoidal, the other is pulse.△ three forms of courtship sound in Drosophila melanogaster, the researchers collected 276 pairs of wild type data, including 2765 minutes of courtship interaction.and the data set was used to train polynomial GLM to predict the singing behavior of male flies in the whole courtship process.on this basis, the researchers created a glm-hmm model, which combined hidden states when predicting singing patterns according to feedback cues.(Note: feedback cues are defined by the movements and interactions between males and females) glm-hmm allows each state to have a polynomial GLM associated with it, which is used to describe the mapping relationship between feedback cues and the probability of specific actions.at the same time, each state has a separate polynomial GLM, which can produce another mapping relationship: feedback cue - the probability of transition from the current state to the next state. that is, this probability varies with the feedback received by the male fly, and the researchers can determine which feedback cues affect the final conversion probability at each time point. the main difference between glm-hmm model and other models for predicting behavior is that it allows each state to use different regression weights to predict behavior. in Drosophila courtship experiment, three states of courtship in Drosophila were summarized by glm-hmm model. the first state is called close. the average distance between the male and female flies is shorter, and the male flies are slowly moving towards the female flies. in this state, the singing of male flies is mainly sinusoidal. the second state is called chasing. the male flies approach the female flies more quickly. in this state, the singing of male flies mainly takes the form of pulse. the third state is called "whatever". the average distance between male and female flies is relatively long, and the male flies move slowly and move away from the female flies. in this state, the male flies do not make much noise. next, the researchers explored which neurons drive the transition between states by activating various neurons. three types of neurons have been focused on: P1A, pip10 and vpr6. the results showed that the possibility of Drosophila entering the "close" state was greatly increased when pip10 was activated, but the activation of P1A and vpr6 neurons had no significant effect. that is, the researchers identified the pair of neurons that manipulate the activity of the fly's brain behind courtship behavior. behavioral biology cross-border to crack the brain, scientists have never stopped studying animal behavior for decades. in 1973, nature published a study describing the process of combing hair in a mouse and estimating the probability of certain actions in different environments. but at the time, researchers needed to capture all the movements of mice, because they didn't know which movements were more important. then, some scientists started to do research in the opposite way. they put the animals in a controlled environment and made them make simple binary decisions, such as whether to turn left or right in the maze. but human interference also impaired the understanding of natural behavior. as a result, scientists have modernized the field by "thinking more quantitatively about behavior.". automation of data collection and data analysis has become a key factor in this transformation. in the 1980s, scientists began to improve computer vision algorithms to solve animal behavior problems. in the next few decades, people developed a system that could mark the position of animals in each frame of the video, distinguish a variety of organisms, and even start to recognize some parts and directions of the body. nevertheless, these projects did not achieve the efficiency that scientists needed. until the advent of machine learning. through deep learning, researchers have begun to track joints and major body parts of almost all animals by training neural networks. all you need is some labeled frames. but it's not easy to put these marks on animals. five years ago, in order to mark mice, researchers at the Norwegian University of science and technology shaved their hair, lightly smeared glass beads on their backs, and smeared luminous ink and polishing agent on the joints of animals. but sometimes these markers are not bright enough to be tracked and interfere with animal behavior. finally, they decided to reconstruct spinal motion with small pieces of reflective tape attached to three points on the animal's back, and a small helmet with four other pieces of tape to track head movements. however, many scientists do not want to make such a mark on animals. as a result, a variety of application systems were born. the first is deeplabcut developed by Harvard University, which was launched last year. they redesigned a neural network that can already classify a lot of objects. other application systems include leap, sleep and deepposekit. the birth of these attitude tracking applications greatly simplifies the work of data collection. Harvard behavioral biologist Benjamin de bivort said: now we can think about other issues. but scientific research is such that it can never stop at the status quo. despite these useful applications, they all rely on supervised learning, that is, trained to infer the position of body parts from manually marked data. scientists hope to achieve this process in an unsupervised way. unsupervised methods themselves have the potential to reveal hidden behavioral structures. an interesting example emerged in 2008. at that time, researchers identified four basic units that make up the worm's movement. These basic units can be stacked together to capture the entire movement of animals. in 2013, Datta took this approach to a whole new level with his Xbox Kinect technology. the animal's three-dimensional dynamic behavior seems to be naturally divided into chunks, lasting an average of 300 milliseconds. based on this discovery, he and his team built a deep neural network to identify these small pieces by breaking down the activities of animals, thus predicting future behavior. behavioral biology, in this way, has become a member of the family of brain cracking methods: in-depth understanding of animal behavior is conducive to uncovering complex and high-density neural coding. about the author △ Jonathan w. pillowadam J. Calhoun, research assistant, Institute of neuroscience, Princeton University. focuses on how the brain makes decisions and chooses behaviors. the tools and techniques developed can accurately quantify animal behavior through machine learning. △ Adam J. Calhoun Jonathan W. pillow, Professor of psychology and neuroscience, Princeton University. research areas include neuroscience, statistics and machine learning. developed statistical methods for characterizing high-dimensional data structures and worked closely with the experimental team to study the nervous system and the calculations performed. interested in the ability of the brain to perform statistical inference in natural tasks and the theoretical principles for controlling the function and design of sensory system. △ Mala Murthy, Professor of Neuroscience at Princeton University. the research focus is on the neural mechanism of Drosophila communication. portal paper: end - AI internal reference | seize the new opportunities of AI development, expand high-quality contacts, obtain the latest AI Information & amp; thesis tutorial, welcome to join the AI internal reference community to learn together ~ communicate with big names | enter the AI community quantum bit վ'ᴗ the contract author վ'ᴗ ի to track the new trends of AI technology and products, please click "watching"!