The Foundation of Understanding: Teaching CNNS to see the connection

The warm discussions around artificial intelligence: What features of one's own personal learning are?

Many authors suggest that artificial intelligence models do not have the same skills as people, especially when it comes to plastic, flexibility and adaptability.

One of the models that models can be trapped for several causal relationships regarding the outside world.

This scripture discusses these problems:

• Similarities in the same between the CONVOONAL Neural Networks (CNN) and a human cortex
• Limitations of CNNs with understanding Causal relationships and learning unseen concepts
• How can CNN make track of simple relationships they are afraid

Is it the same? Is it different?

ConvelcalCalula networks (CNN) [2] Neural networks installed neural take pictures as installing and can be used for multiple tasks. One of the most interesting CNNS stuffs are their inspiration from the human cortex [1]:

• Hierarchical processing. The visual process of the Cortex processes higher images, when early views hold simple features (such as the edges, lines, and colors) and deep areas hold the complex features such as circumstances, items and scenes. CNN, due to its specified structure, absorb the edges and the entry and the original layer, while layers down and include parts or all things.
• The fields receiving. Neurons in a visual cortex responds to the rehabilitation of a visible field of visible field (called the fields called receptives. As we move deep, accepting fields of expansion are exposed, allowing many local information to be compiled. Due to steps to add, the same happened to CNNs.
• Feature distribution. Although biological neurons are different, the same features are seen in different parts of the visible field. In CNNS, different filters scan the whole picture, allowing patterns to be recognized regardless of the area.
• Cleaning. People can see things or are distributed, enlarged, or rotated. CNNs have this place.

These features have made CNNs efficient in visual activity on Superhuman function:

Russiaakovsky et al. [22] The most reported that one's performance reflects a 5.1% of the 5th Dataset dataset. This number is available by a well-trained man in verification photos to better know the existence of appropriate categories. […] Our result (4.94%) crossed the performance of people. Source [3]

Although CNNs are more efficient than people of several jobs, there are still situations where they fail. For example, in a study of 2024 [4]AI models failed to make photography separation. State-of-The-Art models are doing better than people are distant ones but fails when things are on the odds.

In conclusion, our results are more powerful than more powerful in seeing things about illegal materials, (2) time for the duration of the deep networks. Source [4]

In the study [4]They noticed that people need time to succeed. Other functions do not require only physical recognition but also an illegal understanding, which requires time.

General skills make people able to view the rules ruling relationship between things. People see things by ruling and introducing these rules to adapt to new circumstances. Some of the simplest rules “different relationships”: The ability to explain that two things are the same or different. This ability grows fast during infantry and compliance with language development [5-7]. In addition, some animals such as duck and chimpanzees also have [8]. In contrast, learning a different relationship is very difficult for neural networks [9-10].

Convelcalual networks showed hardship in reading this relationship. Similarly, they fail to read other types of easy causes in humans. Therefore, many researchers concluded that CNNS has collapsed illegal choices is necessary for learning this relationship.

These side effects do not mean neural networks cannot read different relationships. Hundreds of trained large models can learn this relationship. For example, Vision-Transformer models are first trained to ImaginNet for different reading can indicate this ability [12].

CNNs can learn different relationships?

The fact that the manufacturers can learn these types of relationships open the desire of CNNS. The opposite relationship is considered between basic logical functions that form the basics of the highest understanding of the order and consultation. To demonstrate that the shallow CNNs can read this idea that you can allow to examine other relationships. In addition, it will allow models to learn the complex complex relationship. This is an important step in developing regular AI skills.

The past work suggests that CNNs have irregular discrimination of buildings so that they can read the relationships. Some authors think this problem is on the training. Generally, the Classical Gradient Forece is used to read one work or a set of tasks. Given a task t or a set of tasks t, the work of losses l is used to increase weights Φ

This can be viewed as part of the loss of different functions (if we have more than one job). Instead, model-Agnostic Meta-learning (MAML) algorithm [13] designed to search the correct point in the area of ​​the network set of related activities. MAML wants to find the first set of instruments θ that reduces the work of losses in all functions, facilitating a quick conversion:

The difference can seem less, but honestly, this method is directed at the issuing and normal performance. If there are many tasks, traditional training is trying to make good work weights. Mammi is trying to identify the set of well-effective weights but at the same time equal in space. This beginnings point θ allows the model to make good use of different functions.

Now we have a racist manner in regular disfellowshipping, we can examine whether we can make CNNs learn a different relationship.

In this study [11]They compare the shallow CNN trained in classic Gradient and Meta-Learn the data designed for this report. The data contains 10 different functions that test different relationships.

The authors [11] Compare CNNS 2, 4, or 6 of the traditional layers are trained in a traditional or meta-learning, indicating several exciting results:

1. Traditional CNN performance indicates the same behavior of random speculation.
2. Meta-Learnient is very enhances working, suggesting that the model can learn a different relationship. The CNN of the 2 key works better than the opportunity, but by increasing the network depth, working improves complete accuracy.

One of the most interesting results for [11] Does the model be trained in a break (use 9 jobs and leave one) and demonstrate the general division force. Therefore, the model has learned behavior can be seen as a small model (6 layer).

Conclusions

Although Convenal networks are inspired that the human brain processes the visual encouragement, they do not include other basic skills. This is especially true when it comes to causal relationship or invisible concepts. Some related relationships can be read from only large models of several training. This has led to thinking that small CNNs cannot read these relationships due to lack of construction of buildings. In recent years, efforts were made to build new buildings for the benefit of learning related thinking. However, many of these projects failed to read these types of relationships. By crossing, this can be succumbed to meta-learning use.

Meta-Learning Profits includes a Brion's learning power. The pressure of the meta reading about the usual look, trying to do good at all times. To do this, reading more than the Abstract features are popular (low features, such as an angles of a particular condition, do not work in normal and suffer). The meta-learning allows the shallow CNN to learn the mysterious behavior that was not otherwise and many parameters and training.

Double CNNs and different relationships is an example of higher understanding activities. Meta-learning and various training methods can help improve the models of models.

One thing!

You can look at other my articles Mediumand you can also connect or get to me LinkedIn or in Sinful. Check This last placecontaining News of Mr & Ai News, or here in some tutorials and here AI reviews. Open to work together and projects, And you can reach me on LinkedIn.

Indication

Here is a list of key indicators I will have to write this article, only the first name of the article cited.

1. Lindsay, 2020, neural networks and neural neural as model to view: the past, which is, and the future, link
2. Li, 2020, survey of NEURAL NEURAL NEURAL networks: Analysis, Programs, and Hopes, Link
3. He, 2015, drinking in the integerance of repairs: Through the performance of people in the separation of ImagetNet, link
4. Ollikka, 2024, comparisons between humans and AI in seeing things with unusual things, link
5. Promark, 1981, Personal Codes and Rights, Connection
6. Blog, 1999, strategies for the organization of the Children's Foreign Work: MicroGenetic Study and Training Research, Link
7. Lupker, 2015, is there a nipple based on a good situation in a different work? Evidence from Babyluntuals-English-Zulu Zulu
8. Tegerent, 2021, studying every including different Relations: Comparison of types of species, link
9. Kim, 2018, not CLEVR: learning a variety of different relationships to shop network networks, link
10. Puebla, 2021, is the deeper neural neural networks that support the opposite work-related consultation? a member of something
11. Gupta, 2025, Neural Neural Neural Networks can (meta-) learn different relationships, link
12. Tartagnange, 2023, deep neural networks can learn a different relationship to view the same, link
13. Finn, 2017, Model-Agnostic Meta-learning Meta Quick conversion of deep networks, link