Master 2 Internship ∗
subject: Deep Network Training under Uncertrainty
Ghazaleh Khodabandelou, Amir Nakib
Laboratoire Images, Signaux et Systèmes Intelligents (LISSI)
Domaine Chérioux, 122 rue Paul Armangot, 94400, Vitry sur Sein
Misclassification is an important issue with many potential negative consequences. It occurs
when learning algorithm attributes a wrong label to a given class. It may decrease the accuracy
of predictions while increasing the complexity of inferred models and the number of necessary
training samples. Understanding the underlying causes of the label errors is one of main challenges
of classification methods. This can help to prioritize certain directions to others for handling the
misclassification and then resolve the model under-performance problem. Generally, classification
algorithms perform well when the positive and negative data points are clearly separated.
Otherwise, the classification becomes naturally error prone. In such a case, a solution to alleviate this is
to train the algorithm with more labeled datasets around the demarcation line area. Overfitting
or underfitting of a given dimension can also be the sources of misclassification. These happen
when a given classifier cannot perfectly distinguish between classes. This means, the model fails
to learn a dimension from the available data. It is possible to overcome this issue by training the
classifier to classify data based on the main differentiator more than other parameters. Although
training the algorithms with more labeled datasets may help to overcome these issues, providing
labels for large amounts of unlabeled data is challenging because labeling the data is expensive
and time-consuming. Moreover, it is an error prone process since it requires human-in-the-loop
interactions that produces naturally mislabeled dataset causing a misclassification.
Many attempts have been done to make algorithms in machine learning label noise-robust (see
a detailed survey on this subject in ). Mislabeling can also have a negative impact on the
training of a convolutional neural network for image classification. Consequently, the learning
of these label errors can lead to a misclassification and a decrease in overall performance with
lower than expected image detection rates. An approach proposed to add an extra noise layer
into the network which adapts the network outputs to match the noisy label distribution . A
classification method has been proposed in  which was designed, especially for the cases where
multiple objects are present in the scene and when the context of the scene impacts the classification
result. Although, the method showed better results using a smaller amount of data in comparison
with the state-of-the-art methods, its classification computation time remains high. This could
be problematic for sensitive classification tasks. Therefore, there is the need for a robust method
tackling the misclassification problem by producing very high accuracy (≥99.97) along with very
low computation time (≤32.99 ms).
The goal of this internship is first to develop an algorithm allowing to detect label error that
the distance between the image and the center of its label. This may be useful to improve the
detection of label errors especially in the case where two images output have a similar distance of
their expected label but one of them has a label error. In other words, it aims at distinguishing
the correct classification from the errors. A misclassified point with a small distance has a higher
probability to be wrongly labeled than one with a larger distance. Although regular convolutional
neural networks are powerful tools for image classification, their performance is hampered by their
vulnerability to adversarial attacks. This is due to the fact that they are assigning high confidence
to regions with few or even no feature points (i.e. a nonlinear transformation of the input space
extracting a meaningful representation of the input data). A Radial Basis Function (RBF) network
potentially could be used in this project to develop a non-linear classifier. A RBF network assigns
high confidence exclusively to the regions containing enough feature points. As it is claimed in ,
RBF networks are naturally immune to adversarial and rubbish-class examples in the sense that
they give low confidence to such examples. RBF units are activated within a well-circumscribed
region of their input-space so that they can make the regions of each class in the feature space
finite and narrow.
∗The opportunity to pursue the subject of internship in Ph.D.
 B. Frénay, M. Verleysen, Classification in the presence of label noise: a survey, IEEE transactions on neural networks and learning systems 25 (5) (2013) 845–869.
 S. Sukhbaatar, J. Bruna, M. Paluri, L. Bourdev, R. Fergus, Training convolutional networks
with noisy labels, arXiv preprint arXiv:1406.2080.
 M. Khata, N. Shvai, A. Hasnat, A. Llanza, A. Sanogo, A. Meicler, A. Nakib, Novel contextaware classification for highly accurate automatic toll collection, in: 2019 IEEE Intelligent
Vehicles Symposium (IV), IEEE, 2019, pp. 1105–1110.
 M. D. Buhmann, Radial basis functions: theory and implementations, Vol. 12, Cambridge
university press, 2003.
The opportunity to pursue the subject of internship in Ph.D.
(c) GdR 720 ISIS - CNRS - 2011-2020.