- Статьи
- Internet and technology
- A mistake has been made: new software will help neural networks fix their own bugs
A mistake has been made: new software will help neural networks fix their own bugs
Russian scientists have created software that allows users without programming skills to develop machine learning models capable of detecting their own failures and errors. The key feature of the technology is predictive mechanisms: neural networks not only add the necessary functions at the user's request, but also evaluate their effectiveness. The models actually created with the help of the new software determine their own level of accuracy and, if necessary, correct the detected flaws. For more information about the development, see the Izvestia article.
Neural networks evaluate themselves
The NTI Center for Digital Materials Science: New Materials and Substances at Bauman Moscow State Technical University has developed a program that can automatically train machine learning models and simultaneously show their confidence in their own forecasts. This will allow both researchers and users with basic computer skills to create modern ML models and understand how reliable their results are.
— For example, it is required to determine from an X-ray image whether the formation in the lung is malignant or benign. The machine learning model analyzes thousands of images and predicts whether it is cancer or not for a new image. Let's say that in the first case, the confidence score is 51%, in the second — 49%. If the doctor takes this uncertainty into account — and 49% is still a significant proportion — the patient will not be immediately prescribed a biopsy or expensive tests, but will be referred for additional diagnostics, such as CT or PET-CT. Or a specialist will upload earlier images to the system, and the model will compare them, which will allow for a more balanced and safe decision," explained Ivan Bespalov, a laboratory assistant at the center.
Another field of application is pharmacology. When developing new drugs, it is necessary to assess their toxicity, and even with limited data, AI is able to predict which drug is potentially more dangerous. However, standard models give only the toxicity value itself, without specifying the scale of possible error. For example, the algorithm can determine a dose of 30 mg for one patient with an error of ± 2 mg, and 20 mg for another with an error of ± 18 mg. Formally, the second value is lower, but due to the high uncertainty, the conclusion may be incorrect. The new algorithm allows such errors to be taken into account, which gives the user the opportunity to interpret the results more accurately.
— In this example, the predicted uncertainty is almost 100% of the value itself (20 mg with an error of 18 mg), — explained Ivan Bespalov. — This means that the model does not understand her own result well and honestly reports it. Accordingly, it is impossible to follow such a forecast. Due to this, a more balanced and safe decision is made, and the patient avoids unnecessary risk.
He called the main advantage that foreign analogues are not able to assess the confidence of their own predictions, and therefore they are more difficult to apply in critical areas where the cost of error can be high. It will be enough to upload an Excel file with the data to the new system, after which it will independently analyze them using machine learning methods. According to him, even users without special training will be able to work with AI, which evaluates the accuracy of its conclusions and indicates a possible error.
By the end of 2026, the NTI center plans to register intellectual property rights.
How to improve the effectiveness of AI training
The new development of the Kryptonite company is moving in the same direction as the NTI center. Experts have found a way to train neural networks so that they can recognize millions of different objects, while remaining compact and not requiring large computing resources. This significantly reduces the risks of freezes, hallucinations, and other errors. The user sets a special hidden space (LSC) structure in advance, so that the neural network can be scaled even in cases where classical learning with a teacher does not work well or is completely impossible. At the same time, the model, in the process of retraining with new objects, continues to recognize those already known to it. This approach improves the quality of image analysis and can be used in a variety of fields, from searching for pathologies in medical images and paying for biometrics to recognizing goods and determining the composition of materials based on photographs.
"This method paves the way for the development of more efficient AI systems capable of working with gigantic sets of categories, which will be especially important with the dynamic development of the market," Nikita Gabdullin, Candidate of Technical Sciences, expert at the Department of Advanced Research and author of this work, told Izvestia.
In traditional recognition methods, the more objects need to be classified, the larger the model itself becomes, which actually limits its scalability. The new approach avoids this: the number of neural network parameters does not grow with the number of classes, while maintaining high accuracy — 87.1% based on 1.28 million images. In addition, the model requires less video memory to work, the company's press service said.
Despite the fact that there are already similar Russian solutions on the market, the MSTU approach stands out for its deep integration of confidence assessment. This is critical for tasks with a high cost of error — in medicine, pharmacy, and materials science. The ability to see where a model is "in doubt" turns AI from a "black box" into an assistant for making informed decisions, says Denis Bokov, head of the Milestone creative group.
— The prospects for the direction are very strong. Such a tool lowers the entry threshold, allowing specialized experts (doctors, chemists, engineers) to independently test hypotheses without waiting for the help of IT departments and without wasting time communicating with developers," he said.
The field of automated machine learning (AutoML) itself, as in the case of the MSTU product, is very promising. But its real value is revealed only with deep integration into specific business processes. Pavel Boyuka, Product Director at Gazprom ID, believes that the key success factor is not the model itself, but the expertise of the team, which is able to take into account the specifics of the data and business objectives.
— This area enjoys serious support at the state level, it is a priority within the framework of the national AI strategy, and a strong educational base of leading universities helps to develop it. If AutoML's effectiveness is confirmed, the next step is to move from automating model development to creating autonomous, self—managed AI ecosystems. Today we are helping the user to build a model, and in the future the system will be able to take over its entire lifecycle," the expert noted.
In the long term, autonomous AI agents will be able to independently set goals and manage complex business processes. Neural networks that can identify their own mistakes and correct forecasts will usher in an era of "intellectual sovereignty" when the capabilities of AI will become a truly accessible tool for everyone.
Переведено сервисом «Яндекс Переводчик»
