A look into the black box - the new system from MIT will show how the machine learning algorithms work
The MIT presented an interactive tool that makes it clear why the intelligent system makes this or that decision. This article is about how it works.
/ Unsplash / Randy Fath
Automated machine learning systems (AutoML) repeatedly test and modify algorithms and their parameters. Using the method of learning with reinforcement , such systems choose AI-models that are more suitable for solving one or another task. For example, to automate technical support . But AutoML systems act like black boxes, that is, their methods are hidden from users.
')
This feature greatly complicates the debugging of machine learning algorithms. And, for example, in the case of autopilot systems, the consequences can be fatal. In 2016, Tesla on autopilot for the first time became a participant in a fatal accident, faced with a large truck. The cause of the accident is not known. The experts have only a guess - the algorithm confused a tall truck with a road sign installed on the lower edge of the overpass. And the error has not yet been eliminated - in the beginning of March a similar accident occurred again in the USA.
The situation was decided by engineers from MIT, the Hong Kong University of Science and Technology and Zhejiang University. They presented a tool for visualizing the processes occurring inside the black box. He was called ATMSeer.
ATMSeer is based on Auto-Tuned Models (ATM). This is an automated machine learning system that searches for the most effective models for solving specific tasks (for example, searching for objects). The system arbitrarily chooses the type of algorithm — neural network, decision trees, “ random forest ” or logistic regression. In the same way, it determines the hyperparameters of the model - the tree size or the number of layers of the neural network.
ATM conducts a series of experiments with test data, automatically adjusting the hyperparameters and evaluating the performance. Based on this information, she chooses the following model, which can show the best results.
Each model is represented as a kind of “unit of information” with variables: algorithm, hyperparameters, performance. Variables are displayed on the corresponding graphs and charts. Further, engineers can manually edit these parameters and monitor changes in the intelligent system in real time.
The tool interface MIT engineers showed in the following video . In it, they disassembled several yuzkeys.
The ATMSeer control panel allows you to manage the learning process and download new data sets. The performance indicators of all models are also displayed on a scale from zero to ten.
Engineers say that the new tool will contribute to the development of the field of machine learning, making the work with intelligent algorithms more transparent. A number of specialists in MO already noted that with ATMSeer they are more confident in the correctness of their models generated by AutoML.
/ Unsplash / Esther Jiao
In the future, you can expect more tools to look into the black box. For example, MIT engineers are already working on another solution. It will help medical students train their history.
In addition to MIT, IBM operates in this area. Together with colleagues from Harvard, they presented the Seq2Seq-Vis tool. He visualizes the decision-making process in machine translation from one language to another. The system shows how each word in the source and destination text is associated with examples in which the neural network was trained. So, it is easier to determine if an error occurred due to incorrect source data or a search algorithm.
Tools that make machine learning algorithms more transparent will also find use in ITSM when implementing the Service Desk. Systems will help with the training of intelligent chat bots, and will avoid situations where they behave differently than intended .
Materials from our corporate blog:
And the blog on Habré:
/ Unsplash / Randy Fath
Black box problem
Automated machine learning systems (AutoML) repeatedly test and modify algorithms and their parameters. Using the method of learning with reinforcement , such systems choose AI-models that are more suitable for solving one or another task. For example, to automate technical support . But AutoML systems act like black boxes, that is, their methods are hidden from users.
')
This feature greatly complicates the debugging of machine learning algorithms. And, for example, in the case of autopilot systems, the consequences can be fatal. In 2016, Tesla on autopilot for the first time became a participant in a fatal accident, faced with a large truck. The cause of the accident is not known. The experts have only a guess - the algorithm confused a tall truck with a road sign installed on the lower edge of the overpass. And the error has not yet been eliminated - in the beginning of March a similar accident occurred again in the USA.
To explain how a machine algorithm came to a conclusion, engineers use a posteriori techniques or interpreted models like decision trees . In the first case, the input and output data is used to approximate the “thinking process” of the algorithm. The accuracy of such techniques leaves much to be desired.
Decision trees are a more accurate approach, but work only with categorized data . Therefore, for complex problems of computer vision is inconvenient.
The situation was decided by engineers from MIT, the Hong Kong University of Science and Technology and Zhejiang University. They presented a tool for visualizing the processes occurring inside the black box. He was called ATMSeer.
How the system works
ATMSeer is based on Auto-Tuned Models (ATM). This is an automated machine learning system that searches for the most effective models for solving specific tasks (for example, searching for objects). The system arbitrarily chooses the type of algorithm — neural network, decision trees, “ random forest ” or logistic regression. In the same way, it determines the hyperparameters of the model - the tree size or the number of layers of the neural network.
ATM conducts a series of experiments with test data, automatically adjusting the hyperparameters and evaluating the performance. Based on this information, she chooses the following model, which can show the best results.
Each model is represented as a kind of “unit of information” with variables: algorithm, hyperparameters, performance. Variables are displayed on the corresponding graphs and charts. Further, engineers can manually edit these parameters and monitor changes in the intelligent system in real time.
The tool interface MIT engineers showed in the following video . In it, they disassembled several yuzkeys.
The ATMSeer control panel allows you to manage the learning process and download new data sets. The performance indicators of all models are also displayed on a scale from zero to ten.
Perspectives
Engineers say that the new tool will contribute to the development of the field of machine learning, making the work with intelligent algorithms more transparent. A number of specialists in MO already noted that with ATMSeer they are more confident in the correctness of their models generated by AutoML.
Also, the new system will help companies meet the requirements of GDPR. General data protection regulations require transparency from machine learning algorithms. The developers of an intelligent system should be able to explain the decisions made by the algorithms. This is necessary so that users can fully understand how the system processes their personal data.
/ Unsplash / Esther Jiao
In the future, you can expect more tools to look into the black box. For example, MIT engineers are already working on another solution. It will help medical students train their history.
In addition to MIT, IBM operates in this area. Together with colleagues from Harvard, they presented the Seq2Seq-Vis tool. He visualizes the decision-making process in machine translation from one language to another. The system shows how each word in the source and destination text is associated with examples in which the neural network was trained. So, it is easier to determine if an error occurred due to incorrect source data or a search algorithm.
Tools that make machine learning algorithms more transparent will also find use in ITSM when implementing the Service Desk. Systems will help with the training of intelligent chat bots, and will avoid situations where they behave differently than intended .
Materials from our corporate blog:
- IT project management - 5 challenges and overcoming them
- Artificial Intelligence classifies requests in ServiceNow
- IT Operations Management: how to manage your IT infrastructure
And the blog on Habré:
Source: https://habr.com/ru/post/459030/
All Articles