In the field of machine learning, there is a black box where researchers usually do n’t know how the model works and make decisions. This has also become one of the topics they have always wanted to understand.

On April 14, local time, the non-profit organization OpenAI launched a neuron visualization tool called “Microscope” (Microscope), which contains 9 popular current Neural network. As its name suggests, this new set of tools can help AI researchers better understand the structure and behavior of neural networks with thousands of neurons, like microscopes in the laboratory.


Neuron visualization

OpenAI was established at the end of 2015 and is a non-profit The artificial intelligence research organization aims to benefit humanity as a whole by promoting and developing friendly artificial intelligence. Among the founders are Silicon Valley “Iron Man” Musk.

OpenAI stated on its official blog that “microscope” visualizes every important layer and neuron in several common models, which can be used for neural networks. Explanatory research was conducted. The “microscope” makes it easier to analyze the internal features of these neural networks, and it can also help researchers progress toward discovering these complex systems.


Legend visualization legend

According to OpenAI, the “microscope” modelThe type consists of “node” graphs of the neural network layer, and these graphs are connected to each other by “edges”. Each operation contains hundreds of “units”, roughly similar to neurons. In the field of machine learning and cognitive science, a neural network is a mathematical model that mimics the structure and function of biological neural networks. The neural network is calculated by connecting a large number of artificial neurons. Our neuron can give some input and then output a result. In machine learning, the neuron also plays the same role. It is a placeholder for a mathematical function. Its job is to use a function for the input, and then give An output result.

The initial model included in the “microscope” includes the computer vision model that is important and widely studied in the field of machine learning. For example, the 2012 ImageNet Challenge champion AlexNet, AlexNet has been cited more than 50,000 times in the study; there is the 2014 ImageNet winner GoogleNet, and ResNet v2. OpenAI said that each model’s visualization comes with some scenes, and the images are available in the OpenAI Lucid library, and can also be reused under the Creative Commons license.

“We hope that anyone who wants to know how a neural network works can use this tool. The main value of this tool is that it can provide lasting shared results, thereby promoting Research on these models. We also hope that researchers with relevant professional knowledge, such as neuroscientists, can easily use these models and at the same time find value from the models. “OpenAI said.

In addition to the “microscope” neuron visualization, in recent years, there have also been attempts in the field of machine learning to visualize the activities of machine learning models. Facebook launched Captum last fall, using visualization technology to explain the decisions made by machine learning models. In March 2019, OpenAI and Google also released activation atlas technology to visualize the decisions made by machine learning algorithms.