Picture source: quanta
Author: Wu Wenhao, Ringo
Deep neural networks (DNNs) are an important achievement in the AI field, but its "sense of existence" is no longer limited to this field.
Some cutting-edge biomedical research is also being attracted by this particular concept. Especially computational neuroscientists.
After revolutionizing computer vision with unprecedented task performance, the corresponding DNNs network was soon used to try to explain the brain's information processing capabilities, and increasingly used as a modeling framework for primate brain neural computing. The task-optimized deep neural network has become one of the best model types for predicting the activities of multiple regions in the visual cortex of primates.
Simulating the brain with neural networks or trying to make neural networks more like the brain is becoming the mainstream. However, some research groups have chosen to use neurobiological methods to re-examine the DNNs invented by the computer science community.
And they found that conditions such as changing the initial weights can change the final training results of the network. This puts forward new requirements for the common practice of using a single network to gain a glimpse of the biological neural information processing mechanism: If the differences of deep neural networks with the same functions are not taken into account, use such networks to build biological brain operating mechanisms. There will be some random effects on the modulus. To avoid this phenomenon as much as possible, computational neuroscientists engaged in DNNs research may need to base their inferences on multiple network instance groups, that is, try to study the centroids of multiple neural networks with the same function to overcome Random influence.
For researchers in the AI field, the team also hopes that this concept of representational consistency can help machine learning researchers understand the differences between deep neural networks operating at different task performance levels.
How deep neural networks explain the brain
Artificial neural networks are built by interconnected units called "perceptrons". Perceptrons are simplified digital models of biological neurons. An artificial neural network has at least two layers of perceptrons, one for the input layer and the other for the output layer. Put one or more "hidden" layers between the input and output, and you get a "deep" neural network. The more these layers, the deeper the network.
Deep neural networks can be trained to recognize features in data, such as features representing cat or dog images. Training involves using an algorithm to iteratively adjust the strength of the connection between the perceptrons (weight coefficients) so that the network learns to associate a given input (pixels of the image) with the correct label (cat or dog). Ideally, once trained, a deep neural network should be able to classify the same type of input it has not seen before.
But in terms of overall structure and function, deep neural networks cannot be said to strictly imitate the human brain. The adjustment of the strength of connections between neurons reflects the associations in the learning process.
Some neuroscientists often point out the limitations of deep neural networks compared with the human brain: a single neuron may process information in a wider range than "failed" perceptrons. For example, deep neural networks often rely on perceptrons. Back-propagation communication method, and this communication method does not seem to exist in the human brain nervous system.
However, computational neuroscientists think differently. Sometimes, deep neural networks seem to be the best choice for modeling the brain.
For example, the existing computer vision system has been affected by what we know as the primate vision system, especially in the path responsible for recognizing people, positions and things, borrowing a mechanism called ventral visual flow.
For humans, the ventral nerve pathway starts from the eyes and then enters the lateral geniculate body of the thalamus, which is a relay station for sensory information. The lateral geniculate body is connected to an area called V1 in the primary visual cortex. Downstream of V1 and V4 are areas V2 and V4, which eventually lead to the lower temporal cortex. The brains of non-human primates have a similar structure (corresponding to the back visual flow is a largely independent channel used to process information about seeing movement and the location of objects).
The neuroscientific insight embodied here is that visual information processing is layered and advanced in stages: the low-level features in the field of view (such as edges, contours, colors, and shapes) are processed in the early stage, and complex representations, such as entire objects and The face will later be taken over by the temporal cortex.
Just like the human brain, each DNN has unique connectivity and representation characteristics. Since the human brain may have better memory or mathematics due to differences in internal structure, the initial settings before training are different. Will the neural network also show a difference in performance during the training process?
The difference between neural networks with the same function
In other words, are there any differences between neural networks with the same functions but different starting conditions?
The key to this question is that it determines how scientists should use deep neural networks in their research.
In a previous paper published in the Nature Newsletter, it was composed of scientists from the MRC Cognitive and Brain Research Group at Cambridge University, the Zuckerman Institute at Columbia University, and the Donders Brain Science and Cognitive and Behavioral Research Center at Rudburg University in the Netherlands. A scientific research team from is trying to answer this question. The title of the paper is "Individual differences among deep neural network models".
According to this paper, deep neural networks with different initial conditions will indeed show increasing individual differences in representation as training progresses.
Previous research mainly used linear canonical correlation analysis (CCA) and centered-kernel alignment (CKA) to compare the internal network representation differences between neural networks.
This time, the team's research also used a common analysis method in the field-representational similarity analysis (RSA, representational similarity analysis).
This analysis method is derived from the multivariate analysis method of neuroscience. It is often used to compare the data produced by the calculation model with the real brain data. It is based on the principle of expressing the system by using "double (or'pair')" feedback difference. RDMs (representational dissimilarity matrices) of Inner stimulus representation (Inner stimulus representation), and the geometry composed of all double feedback groups can be used to represent the geometric arrangement of high-dimensional stimulus space.
If two systems have the same characteristics in the stimulus representation (that is, the similarity of the representation difference matrix is as high as a certain value), they are considered to have similar system representations.
The similarity calculation of the characterizing difference matrix is performed in source spaces with different dimensions and sources, so as to avoid the definition of "mapping network between systems". One of the characteristics of this research in this respect is to compare the representational similarity between networks using the network example comparison analysis method commonly used in neuroscience research, which allows the research results to be directly used in commonly used models in neuroscience research.
In the end, the comparison results show that there are obvious individual differences between different neural networks only on the starting random seed.
This result is valid when different network architectures, different training sets and distance measurements are used. The team analyzed that the degree of this difference is comparable to the difference produced by "training the neural network with different inputs".
Picture source: Nature
As shown in the figure above, the research team compares the representation geometry of all-CNN-C in all network instances and layers by calculating all pairwise distances between corresponding RDMs.
Picture source: paper
Then use MDS to project the data points in a (each point corresponds to a layer and instance) to two dimensions. The layers of each network instance are connected by gray lines. Although the early representative geometric figures are highly similar, as the depth of the network increases, individual differences gradually appear.
After proving the significant individual differences in deep neural networks, the team continued to explore the explanations for these differences.
Internal representation of two deep neural network instances. Picture source: paper
Subsequently, the researchers investigated the influence of network regularization (network regularization) on the results by using the Bernoulli dropout method in the training and testing phases, but found that although regularization can improve "the use of different starting random seeds" The identity of the network is consistent, but it cannot correct the individual differences between these networks.
Finally, by analyzing the training trajectory of the network and the process of individual differences and visualizing this process, the team stated in the paper that there is a strong negative correlation between the performance of the neural network and the consistency of the representation, that is, the individual differences between the networks will be Intensified during training.
All in all, this study mainly investigates whether there are individual differences between multiple neural networks under the minimum experimental intervention conditions, that is, setting random seeds with different weights for the network before the start of training, but keeping other conditions consistent, and expand the previous Research related to "correlation between neural networks".
In addition to this research, Hinton, one of the "Big Three in Deep Learning" and a well-known AI scholar, also has related research. The paper is titled "Similarity of Neural Network Representations Revisited", and the article explores measuring the similarity of deep neural network representations. For sexual issues, interested readers can read them together.
Refrence：
[1]https://www.nature.com/articles/s41467-020-19632-w#citeas
[2]https://www.quantamagazine.org/deep-neural-networks-help-to-explain-living-brains-20201028/
About data combat faction
The data practitioner hopes to use real data and industry practical cases to help readers improve their business capabilities and build an interesting big data community.
