Chinese scholars put forward new ideas that are stimulating noise

Figure 1 The image recognition accuracy "counterintuitively" shows the phenomenon of "first increase and then decrease" with the increase of image noise intensity

Figure 2 Application example of positive excitation noise in cross-domain remote sensing target detection task

With the support of the National Natural Science Foundation of China (grant number: 61871470), Professor Li Xuelong of Northwestern Polytechnical University proposed a new idea
of defining noise with "task entropy" and using and even designing "useful noise" 。 The research results were published on December 29, 2022 in the journal IEEE Transactions on Neural Networks and Learning Systems under the title "Positive-incentive Noise
".
Link to paper: https://ieeexplore.
ieee.
org/document/10003114
.

In engineering scientific research, it is almost inevitable that there will be noise at multiple levels such as data, features, samples, and decisions, for example, speckled noise will inevitably occur due to the path difference of coherent sound waves in ultrasound images, and unpredictable noise
will occur due to special environments such as glare during remote sensing imaging.
Therefore, in the current research work, researchers always try to minimize the noise content and think that this will minimize its impact
.
The negative impact of noise on tasks has become a well-established assumption
in the field of signal processing theory and engineering tasks.
However, Professor Li Xuelong found in scientific research such as signal classification that when the right amount of noise is added to the data, the accuracy of artificial intelligence algorithms for signal recognition does not decrease but increases
.
Overall, classification accuracy increases and decreases as noise content increases (Figure 1), suggesting that noise can also be useful
.

In view of this phenomenon, Professor Li Xuelong believed that "the generation of certain noises and signals is homologous" after in-depth thinking, and further realized that the systematic study of noise is inseparable from the quantitative description
of the task.
Therefore, he proposed a theoretical system for quantitatively analyzing task complexity with Task Entropy, and established a noise analysis and calculation framework
in information processing.
According to this computational framework, under the given task, the noise is classified as "positive-incentive noise" (Pi/π-noise) and "pure noise"
by calculating the Mutual Information between arbitrary noise and the task.
Among them, positive excitation noise refers to the noise that "the mutual information with the task is greater than 0, which can reduce the task entropy"; Pure noise refers to noise that "has zero mutual information and cannot reduce task entropy"
.
In the case of a cross-domain remote sensing target detection task (Figure 2), the background is often considered to be the "noise" that needs to be rejected, and the bounding box of the object (in this case, the aircraft) is generally required to be as compact as possible
.
According to the theory of positive excitation noise, reasonable use of background noise can help simplify the detection task and improve the detection accuracy
.
Therefore, when training the model, the size of the bounding box can be adjusted adaptively according to the criterion of maximizing the mutual information between noise and the task, so that it contains an appropriate amount of background with strong coupling with the detection target, such as runway and apron
.
Professor Li Xuelong also pointed out that the core of the calculation framework is to find the "task probability distribution" corresponding to the task and determine the boundary
of the positive excitation noise.

The research work of the project inspires people to re-understand noise in information processing, and provides new ideas
for information processing and noise utilization in the fields of cross-domain remote sensing, multimodal cognitive computing, wading optics, and stability detection.