Semiotic Aggregation in Deep Learning

Musat, Bogdan; Andonie, Răzvan

doi:10.3390/e22121365

Cited by 7 publications

(24 citation statements)

References 31 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…These heat maps can help in the interpretation of the algorithm’s classification results, e.g., as shown in Lee et al [ 47 ]. An example of a heat map (Grad-CAM method) is shown in Figure 5 [ 48 , 49 ]. By using these maps, it can be assessed which scans the algorithm classifies correctly as positives for the wrong reasons, and which scans are missed by the algorithm, which can also be used as indications for possible biases [ 50 ].…”

Section: Resultsmentioning

confidence: 99%

Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification

Sourlos

Wang

Nagaraj

et al. 2022

Cancers

View full text Add to dashboard Cite

Artificial Intelligence (AI) algorithms for automatic lung nodule detection and classification can assist radiologists in their daily routine of chest CT evaluation. Even though many AI algorithms for these tasks have already been developed, their implementation in the clinical workflow is still largely lacking. Apart from the significant number of false-positive findings, one of the reasons for that is the bias that these algorithms may contain. In this review, different types of biases that may exist in chest CT AI nodule detection and classification algorithms are listed and discussed. Examples from the literature in which each type of bias occurs are presented, along with ways to mitigate these biases. Different types of biases can occur in chest CT AI algorithms for lung nodule detection and classification. Mitigation of them can be very difficult, if not impossible to achieve completely.

show abstract

Section: Resultsmentioning

confidence: 99%

Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification

Sourlos

Wang

Nagaraj

et al. 2022

Cancers

View full text Add to dashboard Cite

show abstract

“…Therefore it is interesting to observe if any form of superization is present in the training process of a CNN. In [19] we applied the above theorem to the neural layers of CNNs. We computed superization with respect to the spatial entropy variations of the saliency maps.…”

Section: Semiotic Superization In Cnnsmentioning

confidence: 99%

“…The field of interpretability/explainability in deep learning has witnessed an explosion of published papers in recent years. Even if there is no fundamental theory that can elucidate all underlying mechanisms present in those networks, multiple works tried to deal with this issue by coming up with partial solutions, either by visual explanations [22], [18] or theoretical insights [14], [9], [19]. Therefore, we can say that the black box interpretation of deep learning is not true anymore, and what we need are better techniques to interpret these models.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Information Bottleneck in Deep Learning - A Semiotic Approach

Musat

Andonie

2022

INT J COMPUT COMMUN, Int. J. Comput. Commun. Control

View full text Add to dashboard Cite

The information bottleneck principle was recently proposed as a theory meant to explain some of the training dynamics of deep neural architectures. Via information plane analysis, patterns start to emerge in this framework, where two phases can be distinguished: fitting and compression. We take a step further and study the behaviour of the spatial entropy characterizing the layers of convolutional neural networks (CNNs), in relation to the information bottleneck theory. We observe pattern formations which resemble the information bottleneck fitting and compression phases. From the perspective of semiotics, also known as the study of signs and sign-using behavior, the saliency maps of CNN’s layers exhibit aggregations: signs are aggregated into supersigns and this process is called semiotic superization. Superization can be characterized by a decrease of entropy and interpreted as information concentration. We discuss the information bottleneck principle from the perspective of semiotic superization and discover very interesting analogies related to the informational adaptation of the model. In a practical application, we introduce a modification of the CNN training process: we progressively freeze the layers with small entropy variation of their saliency map representation. Such layers can be stopped earlier from training without a significant impact on the performance (the accuracy) of the network, connecting the entropy evolution through time with the training dynamics of a network.

show abstract

“…Despite the ability of generating, especially in image recognition tasks, human-alike predictions, CNNs still lack a major component: interpretability [ 34 ]. Neural networks in general are known for their black-box type of behavior, hiding the inner working mechanisms of reasoning.…”

Section: Computing the Te Feedback In A Cnnmentioning

confidence: 99%

Learning in Convolutional Neural Networks Accelerated by Transfer Entropy

Moldovan

Caţaron

Andonie

2021

Entropy

Self Cite

View full text Add to dashboard Cite

Recently, there is a growing interest in applying Transfer Entropy (TE) in quantifying the effective connectivity between artificial neurons. In a feedforward network, the TE can be used to quantify the relationships between neuron output pairs located in different layers. Our focus is on how to include the TE in the learning mechanisms of a Convolutional Neural Network (CNN) architecture. We introduce a novel training mechanism for CNN architectures which integrates the TE feedback connections. Adding the TE feedback parameter accelerates the training process, as fewer epochs are needed. On the flip side, it adds computational overhead to each epoch. According to our experiments on CNN classifiers, to achieve a reasonable computational overhead–accuracy trade-off, it is efficient to consider only the inter-neural information transfer of the neuron pairs between the last two fully connected layers. The TE acts as a smoothing factor, generating stability and becoming active only periodically, not after processing each input sample. Therefore, we can consider the TE is in our model a slowly changing meta-parameter.

show abstract

Semiotic Aggregation in Deep Learning

Cited by 7 publications

References 31 publications

Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification

Possible Bias in Supervised Deep Learning Algorithms for CT Lung Nodule Detection and Classification

Information Bottleneck in Deep Learning - A Semiotic Approach

Learning in Convolutional Neural Networks Accelerated by Transfer Entropy

Contact Info

Product

Resources

About