Deep Learning Under Privileged Information Using Heteroscedastic Dropout

Lambert, John; Şener, Ozan; Savarese, Silvio

doi:10.1109/cvpr.2018.00926

Cited by 68 publications

(55 citation statements)

References 37 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This guidance led to significantly improved results which is in line with related work using LUPI for image related tasks (Gu et al . (2020); Lambert et al . (2018); Chen et al .…”

Section: Resultsmentioning

confidence: 99%

See 1 more Smart Citation

Learning to see colours: generating biologically relevant fluorescent labels from bright-field images

Wieslander

Gupta

Bergman

et al. 2021

Preprint

View full text Add to dashboard Cite

Fluorescence microscopy, which visualizes cellular components with fluorescent stains, is an invaluable method in image cytometry. From these images various cellular features can be extracted. Together these features form phenotypes that can be used to determine effective drug therapies, such as those based on nanomedicines. Unfortunately, fluorescence microscopy is time-consuming, expensive, labour intensive, and toxic to the cells. Bright-field images lack these downsides but also lack the clear contrast of the cellular components and hence are difficult to use for downstream analysis. Generating the fluorescence images directly from bright-field images would get the best of both worlds, but can be very challenging to do for poorly visible cellular structures in the bright-field images. To tackle this problem deep learning models were explored to learn the mapping between bright-field and fluorescence images to enable virtual staining for adipocyte cell images. The models were tailored for each imaging channel, paying particular attention to the various challenges in each case, and those with the highest fidelity in extracted cell-level features were selected. The solutions included utilizing privileged information for the nuclear channel, and using image gradient information and adversarial training for the lipids channel. The former resulted in better morphological and count features and the latter resulted in more faithfully captured defects in the lipids, which are key features required for downstream analysis of these channels.

show abstract

“…This guidance led to significantly improved results which is in line with related work using LUPI for image related tasks (Gu et al . (2020); Lambert et al . (2018); Chen et al .…”

Section: Resultsmentioning

confidence: 99%

“…The way in which this is done in deep learning is to develop separate network streams, one for the main task and one based on the privileged information, and then to incorporate some means of guiding the main model with insights from the privileged information stream (Gu et al . (2020); Lambert et al . (2018); Chen et al .…”

Section: Introductionmentioning

confidence: 99%

Learning to see colours: generating biologically relevant fluorescent labels from bright-field images

Wieslander

Gupta

Bergman

et al. 2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Hyperbolic tangent as the nonlinear activation function is employed in the experiments. To alleviate the problem of overfitting, dropout [26] is applied and set to 0.5. e number of layers is adjusted to 7, and the number of epochs is set to 100. For deep learning algorithms, CNN, RNN, and LSTM models have 7 layers.…”

Section: Experiments Resultsmentioning

confidence: 99%

“…It is followed by the implementation of RNN with 2 Simple RNN layers each with 32 RNN cells followed by 2 time distribute dense layers for 5 class classification. Dropout [26] rate on the fully connected layers is set to 0.5. Tangent activation function is used at the last layer.…”

Section: Experiments Resultsmentioning

confidence: 99%

“…is study employs images to forecast the features that are extracted from images and calculates probabilities of the final movement. Electroencephalography (EEG) devices are also utilized in some of studies in the literature [21][22][23][24][25][26][27][28][29]. In [21], it is shown that measuring EEG signals gives an opinion about human emotional conditions.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Mood Detection from Physical and Neurophysical Data Using Deep Learning Models

Kilimci

Guven²,

Uysal

et al. 2019

Complexity

View full text Add to dashboard Cite

Nowadays, smart devices as a part of daily life collect data about their users with the help of sensors placed on them. Sensor data are usually physical data but mobile applications collect more than physical data like device usage habits and personal interests. Collected data are usually classified as personal, but they contain valuable information about their users when it is analyzed and interpreted. One of the main purposes of personal data analysis is to make predictions about users. Collected data can be divided into two major categories: physical and behavioral data. Behavioral data are also named as neurophysical data. Physical and neurophysical parameters are collected as a part of this study. Physical data contains measurements of the users like heartbeats, sleep quality, energy, movement/mobility parameters. Neurophysical data contain keystroke patterns like typing speed and typing errors. Users’ emotional/mood statuses are also investigated by asking daily questions. Six questions are asked to the users daily in order to determine the mood of them. These questions are emotion-attached questions, and depending on the answers, users’ emotional states are graded. Our aim is to show that there is a connection between users’ physical/neurophysical parameters and mood/emotional conditions. To prove our hypothesis, we collect and measure physical and neurophysical parameters of 15 users for 1 year. The novelty of this work to the literature is the usage of both combinations of physical and neurophysical parameters. Another novelty is that the emotion classification task is performed by both conventional machine learning algorithms and deep learning models. For this purpose, Feedforward Neural Network (FFNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM) neural network are employed as deep learning methodologies. Multinomial Naïve Bayes (MNB), Support Vector Regression (SVR), Decision Tree (DT), Random Forest (RF), and Decision Integration Strategy (DIS) are evaluated as conventional machine learning algorithms. To the best of our knowledge, this is the very first attempt to analyze the neurophysical conditions of the users by evaluating deep learning models for mood analysis and enriching physical characteristics with neurophysical parameters. Experiment results demonstrate that the utilization of deep learning methodologies and the combination of both physical and neurophysical parameters enhances the classification success of the system to interpret the mood of the users. A wide range of comparative and extensive experiments shows that the proposed model exhibits noteworthy results compared to the state-of-art studies.

show abstract

TRADI: Tracking Deep Neural Network Weight Distributions

Franchi

Bursuc

Aldea

et al. 2020

Lecture Notes in Computer Science

View full text Add to dashboard Cite

During training, the weights of a Deep Neural Network (DNN) are optimized from a random initialization towards a nearly optimum value minimizing a loss function. Only this final state of the weights is typically kept for testing, while the wealth of information on the geometry of the weight space, accumulated over the descent towards the minimum is discarded. In this work we propose to make use of this knowledge and leverage it for computing the distributions of the weights of the DNN. This can be further used for estimating the epistemic uncertainty of the DNN by aggregating predictions from an ensemble of networks sampled from these distributions. To this end we introduce a method for tracking the trajectory of the weights during optimization, that does neither require any change in the architecture, nor in the training procedure. We evaluate our method, TRADI, on standard classification and regression benchmarks, and on out-of-distribution detection for classification and semantic segmentation. We achieve competitive results, while preserving computational efficiency in comparison to ensemble approaches.

show abstract

Deep Learning Under Privileged Information Using Heteroscedastic Dropout

Cited by 68 publications

References 37 publications

Learning to see colours: generating biologically relevant fluorescent labels from bright-field images

Learning to see colours: generating biologically relevant fluorescent labels from bright-field images

Mood Detection from Physical and Neurophysical Data Using Deep Learning Models

TRADI: Tracking Deep Neural Network Weight Distributions

Contact Info

Product

Resources

About