The machine learning system is built on a layer of artificial neurons called a neural network. These network layers are connected by seemingly random connections between neurons, and the entire system "learns" by fine-tuning these connections.
This has become an important support for the effective operation of artificial intelligence systems today, however, it operates in an extremely "mysterious" way.
For things like "Is this a cat's photo?" "How should the next move go?" "Whether the self-driving car should accelerate when encountering a yellow light?" And so on, the neural network can often give an easy answer. But the key issue is that we don't know how it works. This is the so-called "black box".
In order to make artificial intelligence more trustworthy in specific applications, more and more researchers are trying to open the "black box" and understand the process by which the system draws a specific conclusion.
Recently, Trey Ideker, a professor of bioengineering and medicine at the University of California, San Diego, and his research team developed a "visible" neural network and used it to construct a brewer's yeast cell model called DCell (usually used as a model of basic research).
Specifically, mapping neural networks into simple yeast cells allows researchers to see how AI systems work. In the process, researchers have obtained many analytical conclusions about cell biology, and the resulting technology may also help to develop new cancer drugs and personalized treatment options.
First, let's introduce the basics of neural networks used in today's machine learning systems:
Computer scientists build neural network frameworks by setting up multiple layers, each of which contains thousands of "neurons" responsible for performing tiny computational tasks.
On this basis, trainers enter data sets (such as millions of cats, dogs, millions of Goss, millions of driving operations and results, etc.) that connect the neurons in the layer by the system. Perform a structured sequence calculation on it. The system will process the data through the neural network and then check the actual effect of its task (such as the accuracy of distinguishing the cat from the dog).
Finally, check if the new pattern produces better results by rearranging the connection patterns between neurons and running the data set again. When the neural network is able to complete the task very accurately, the trainer will conclude that the training is successfully completed.
"Although they are called neural networks, the human neural system inspired by these systems is still very basic," Ideker explained.
He pointed out: "Looking at AlphaGo, you can see that the internal workings of this system are completely messy. In fact, it is not like the human brain. It has a new way of thinking, but it just happens to make a good prediction. "
Based on this, Ideker began to make new attempts in the research of artificial intelligence in cell biology. He hopes to use the neural network to show researchers how to draw these conclusions, not just the simple and rude answer.
In an interview, Ideker said: "We have a strong interest in a specific structure that is not optimized by computer scientists but optimized through evolution."
â–² DCell can accurately predict the growth and reproduction of yeast cells as well as laboratory experiments.
This project is feasible because S. cerevisiae is a single-celled organism that has been studied as a basic biological system since the 1950s. Ideker pointed out: "We have a lot of knowledge about cell biology, so this study is very convenient."
Therefore, his team gradually extends up to a larger structure by mapping the various layers in the neural network to the components of the yeast cell, starting with the most microscopic constituent elements (the nucleotides that make up the DNA) - for example Ribosomes (obtain instructions from DNA to make proteins), and finally reach organelles such as mitochondria and nucleus (responsible for performing cellular activities). Overall, the DCell neural network will utilize a total of 2,526 subsystems in yeast cells.
â–² DCell is available as an online application for researchers
DCell allows researchers to alter the DNA of cells (ie, genetic code) and observe how these changes spread upward to alter their biological characteristics, which in turn affects subsequent cell growth and reproduction. Its training data set consists of gene mutation instances from millions of real yeast cells and matches the corresponding mutation result information.
The researchers found that DCell can accurately predict cell growth by simulating yeast. Because this is a "visible" neural network, researchers can see the changes in cellular mechanisms that occur when DNA is confused.
This visibility means that DCell can potentially be applied to computerized research in cells without the need to spend a lot of time investing in laboratory experiments. If researchers can figure out the actual modeling process -- not simple yeast cells -- they can further simulate more complex human cells. "If you can build and simulate the overall working model of a single human cell, this will revolutionize the direction of precision medicine and drug development," Ideker said.
Cancer is currently the most concerned disease research direction because each cancer patient's tumor cells contain a unique combination of mutations. And Ideker and his team are using models of the patient's genome and mutation conditions to observe the actual growth rate of the cells and the aggressive properties of the cancer.
More importantly, pharmaceutical companies that are looking for new cancer drugs will be able to use cell growth as a criterion for success or failure. They will observe a number of different genetic molecules that can be turned on and off, and then consider whether a potential drug can stop the proliferation of tumor cells. Considering the multi-billion dollar investment in anticancer drug research and development in the past, this more convenient research method is now significantly more attractive.
Of course, upgrading from yeast cells to human cells is no easy task. Researchers need to gather enough information about human patients to build the training data set necessary for neural networks -- at least millions of records, including the patient's genetic map and health outcomes. Ideker predicts that the data will accumulate quite quickly. In his view, sequencing the patient's genome will be highly regarded.
The more difficult part is the accumulation of knowledge of the mechanisms of human cancer cell activity, in order to map neural networks to various parts of the cell. Ideker himself is a member of the Cancer Mapping Program and they hope to solve this challenge as soon as possible. At present, it is a very difficult task to summarize the biological activities of cancer cells, because these mutations can not only open and close the cell function, but also affect the cell function to varying degrees, and lead to coordination in an extremely complicated way. Variety.
However, Ideker is still optimistic about the use of migration learning technology to transform machine learning programs from simulated yeast cells to neural networks that mimic human cells. He concluded: "As long as you have a system that recognizes cats, you don't need to retrain them completely, you can teach them how to recognize squirrels."
Greenhouse Automatic Proportional Pump
Fertilizer pump mounted directly on the pipe, the kinetic energy of pipe flow to drives the pump to work at a set ratio to suck high erconcentration drug or quantitative fertilizer in to the pump. After mixed with the water and delivered to downstream. No matter How to change the water pressure, Drug or fertilizer will be mixed and discharged according to a certain proportion.
Greenhouse Automatic Proportional Pump,Greenhouse Automatic Drip Irrigation,Greenhouse Automatic Fertilizer Injector
JIANGSU SKYPLAN GREENHOUSE TECHNOLOGY CO.,LTD , https://www.alibabagreenhouse.com