Extracting non visual attributes from images using Deep Learning.
Deep learning is a set of techniques for training neural networks with many layers. It’s a subfield of machine learning, which considers other models of data besides neural networks.
The increasing performance of visual recognition tasks provided by deep learning methods has open the way to new
computer vision applications that rely on rich descriptions of image content. Today, we can tag an image with labels related to thousands of visual concepts such as objects, scenes, or even short captions describing the image content. But even if some specific image classification problems remain unsolved, state-of-the-art end-to-end learning currently allows to go one step further: we can learn non visual concepts from images that are highly relevant for data science applications. Extracting non visual attributes from images was considered very hard because of the lack of good image descriptors. For example, what is a good descriptor for labeling an image with an index of “happiness”, or a city image with an index of “livability”? There were good image descriptors for visual attribute extraction (HoG, SIFT, Daisy, etc.) because one of the main tasks of researchers was the design of handcraftted features. This process was heavily dependent on researchers’ intuition about how to represent objects and scenes. Deep learning automatically discovers the features that best represent the problem, rather than just a way to combine them. For example, in object recognition, shallow learning starts with handcrafted features of the image, but deep learning starts with the raw pixels. For this reason, deep learning frees the researcher from handcrafting features and is well suited to extract useful attributes from images that are not realted to
researcher’s intuition. It paves the way to answer questions such as ‘Can computer vision be used to automatically select pictures that will make your apartment listing successful on Airbnb?’.
The main objectives of the project are:
1. To build (design, train and deploy) a deep learning system that is able of predicting a set of non visual attributes from images.
2. To use the system to measure the value of images in some online markets (f.e. peer-to-peer accomodation).
3. To propose and validate a set of tools based on the developed system to show its potential added value for real
problems.
International Workshop on Social Signal Processing and Beyond
**********************************************************************
http://www.ub.edu/cvub/SSPandBE/index.html
September 11, 2017, Catania, Italy
in association with ICIAP 2017 (http://www.iciap2017.com/)
Given the increasing quantities of personal data being gathered by individuals, the concept of a digital library of rich multimedia and sensory content for every individual is becoming a reality and fast becoming a mainstream topic for multimedia research. This is often referred to as lifelogging and there are significant challenges to be addressed in the area, concerning the gathering, enriching, searching and accessing of lifelog data.
In recent years, deep learning has achieved remarkable results in fields such as: computer vision, speech recognition and natural language processing. This DL revolution is slowly reaching the challenging problems of the medical domain, opening the doors for personalized medicine. Medical domain is characterized by high variability of data including text, imaging, and genomic data. In this talk, we will present recent advances in two domains of medical data: imaging and genomics.
First, we will introduce a simple, yet powerful pipeline for medical image segmentation that combines Fully Convolutional Networks (FCNs) with Fully Convolutional Residual Networks (FC-ResNets). We propose and examine a design that takes particular advantage of recent advances in the understanding of both Convolutional Neural Networks as well as ResNets. Our approach focuses upon the importance of a trainable pre-processing when using FC-ResNets and we show that a low-capacity FCN model can serve as a pre-processor to normalize medical input data. We show that using this pipeline, we exhibit state-of-the-art performance on the challenging Electron Microscopy benchmark, when compared to other 2D methods. We improve segmentation results on CT images of liver lesions, when contrasting with standard FCN methods. Moreover, when applying our 2D pipeline on a challenging 3D MRI prostate segmentation challenge we reach results that are competitive even when compared to 3D methods. The obtained results illustrate the strong potential and versatility of the pipeline by achieving highly accurate results on multi-modality images from different anatomical regions and organs.
Second, we will introduce a novel deep learning architecture to address the challenges posed by genomic data, where the number of input features can be orders of magnitude larger than the number of training examples, making it difficult to avoid overfitting, even when using the known regularization techniques. Improving the ability of deep learning to handle such datasets could have an important impact in medical research, more specifically in precision medicine, where high-dimensional data regarding a particular patient is used to make predictions of interest. We propose a novel neural network parameterization, that we call Diet Networks, which considerably reduces the number of free parameters in the model. The Diet Networks parametrization is based on the idea that we can first learn or provide an embedding for each input feature and then learn how to map a feature's representation to the parameters linking the value of the feature to each of the hidden units of the classifier network. We experiment on a population stratification task of interest to medical studies and show that the proposed approach can significantly reduce both the number of parameters and the error rate of the classifier. This work was accepted at ICLR 2017.
Open Postdoctoral fellowship/s in the Dept. de Matemàtiques i Informàtica at Universitat de Barcelona.
The Dept. de Matemàtiques i Informàtica (Mathematics and Computer Science) is looking for and willing to support excellent postdoctoral researchers in the fields of Machine Learning, Computer Vision and Human-Computer Interaction who are interested in applying for a Beatriu de Pinós (BP) 2016 fellowship so as to conduct a two-year postdoc at Universitat de Barcelona.
The purpose of the Beatriu de Pinós programme is to award 60 individuals grants for the hiring and incorporation of postdoctoral research staff into the Catalan science and technology system. These grants are designed for the incorporation of young researchers (who obtained their PhD between 2007 and 2014 and have not resided or worked in Spain for more than 12 months in the three years prior to date of submission of the application), so that they can improve their professional prospects and obtain an independent research position. Candidates must carry out a research and training project for the entire period of the grant, one that will allow them to progress in the development of their professional careers. Please check the website of the BP programme* for further information about this fellowship.
Some of the specific projects, we are working, include:
+ Machine Learning: Deep Learning for time series analysis, Supervised Online Learning Algorithms, Bayesian statistics and deep learning.
+ Computer Vision: Visual Lifelogging and Egocentric Vision, Neuroimage processing, Computer Vision for Food Analysis, Deep learning and Image Aesthetics, Ultrasound image analysis.
+ Human-Computer Interaction: ageing / older people, interfaces for people with mild dementia or with aphasia, universal design of STEM documents
Deadline: 0112/2016.
For further information about this postdoctoral opportunity please feel free to contact us:
Petia Radeva (petia.ivanova@ub.edu or radevap@gmail.com),
www.ub.edu/cvub,
www.cvc.uab.es/people/petia
Petia Radeva has been invited speaker to the workshop “Humanitarian and social science: from the university to the enterprise” 17 of November, 2016, Faculty of Filology, University of Barcelona.
Best paper award at CIAPR'2016 to our paper: "Deep Learning Features for Wireless Capsule Endoscopy Analysis",
by Santi Segui, Michal Drozdzal, Guillem Pascual, Carolina Malagelada, Fernando Azpiroz, Petia
Radeva and Jordi Vitrià, Lima Perú, 2016.
Petia Radeva received the International CIARP Award "Aurora Pons Porrata" in recognition of an outstanding technical contribution to the field of pattern recognition, data mining and related areas, 2016.
We got the Mention prize (II place) in the DKV competence with our App for Automatic Food Recognition for Healthy Habits Promotion! Congratulations, team!!!
The journal Medical Physics chose figures of our work on Stent analysis to use as a cover on their journal. http://www.medphys.org
Petia Radeva gave a plenary talk "Can Deep Learning and Egocentric Vision for Visual Lifelogging help us eat better?" at the CCIA'2016, organized by Xavier Binefa, UPF, Barcelona, Spain.
The prestigous GRADIANT award was assifned to Beatriz Remeseiro for the best PhD thesis applied to the ICT sector 2016.
3 abstracts accepted at the NIPS Workshop WiML'16. Congratulations, Mariella, Maya and Beatriz!
On 8 of February, 2016 
Michal Drozdzal is one of the 5 researchers who received the Pioneer Award for her doctoral thesis "Sequential image analysis for computer-aided endoscopi wireless". This is the second edition of the competition promoted by CERCA. This year a total of nineteen researchers (10 males and 9 females) participated from thirteen centres nearby. By areas, there have been 3 science projects; 7 of medical sciences and health; 3 of 6 engineering and life sciences. The jury appreciated the advanced technology that solves problems linked to the endoscopy with a significant population component and a business link which enables a good transfer of knowledge generated.
Re-memory: Cognitive Training based on autobiograohic records to exercise memory
Exposition “+HUMANS”, 17 December 2015, CCB, Barcelona Spain
The majority of patients with mild cognitive impairment develop dementia and Alzheimer's disease. Re-memory, a project financed by the Foudation “La Marató de TV3”, study a new entrenament cognitiu based on the concept of lifelogging to exrcise memory. Re-memory works in the following way: the patient carries a camera that capture automatically images from all locations visited, events in which the wearer participates, the activities that he/she participated and persons he/she interacted with. Inspired in the use of wearable cameras for first cases of amnesia, els membres of the Re-memory project create a cogntive program for training based on the autobiographic reexperimenting to positively impact the cognition and improve the memory and function of peoplewith mild cognitive impairment.
Our work on using Deep Learning techniques for endoluminal image analysis was presented as example at NVIDIA’s GPU Technical Conference (http://www.gputechconf.com/).