Thursday 21 Jan 2021 – SPR
11:00 – 12:00 (CET): Keynote (Nicholas Carlini)
Deep Learning: (still) Not Robust
One of the key limitations of deep learning is its inability to generalize to new domains. This talk studies recent attempts at increasing neural network robustness to both natural and adversarial distribution shifts.
Robustness to adversarial examples, inputs crafted specifically to fool machine learning models, are arguably the most difficult type of domain shift. We study 13 recently proposed defenses at ICLR, ICML, and NeurIPS and find that all can be evaded and offer nearly no improvement on top of the undefended baselines. Worryingly, we are able to break these defenses without any new attack techniques.
It’s not just adversarially-constructed distribution shifts that cause neural networks to suffer: they also don’t generalize across natural distribution shifts that occur in completely benign settings—like re-sampling a new test set. Despite the many proposed techniques to increase synthetic robustness, almost none improve robustness across four natural distribution shifts.
Robustness is still a challenge for deep learning, and one that will require extensive work to solve.
12:30 – 13:30 (CET): Papers discussion session
15:00 – 16:00 (CET): Keynote (Michael Bronstein)
Geometric Deep Learning: Past, Present, And Future
Geometric deep learning has recently become one of the hottest topics in machine learning, with its particular instance, graph neural networks, being used in a broad spectrum of applications ranging from 3D computer vision and graphics to high energy physics and drug design. Despite the promise and a series of success stories of geometric deep learning methods, we have not witnessed so far anything close to the smashing success convolutional networks have had in computer vision. In this talk, I will outline my views on the possible reasons and how the field could progress in the next few years.
- (5) Jakub Nalepa, Krzysztof Hrynczenko and Michal Kawulok. Multiple-Image Super-Resolution Using Deep Learning and Statistical Features
- (9) Juan Maroñas Molano, Daniel Ramos and Roberto Paredes. On Calibration of Mixup Training for Deep Neural Networks
- (15) Wouter Kouw and Marco Loog. Target Robust Discriminant Analysis
- (16) Imran Imran Razzak and Abdul Qayyum. Progressive Light Residual Neural Network for Child’s Spontaneous Facial Expressions Recognition
- (18) Mubarakah Alotaibi and Richard Wilson. Multi-Layer PCA Network For Image Classification
- (19) Yajie Li, Yiqiang Chen, Yang Gu and Jianquan Ouyang. A multimodal fusion model based on hybrid attention mechanism for gesture recognition
- (23) Maximilian Münch, Michiel Straat, Michael Biehl and Frank-Michael Schleif. Complex-valued embeddings of generic proximity data
- (31) Shreyank N Gowda and Chun Yuan. StegColNet: Steganalysis based on an ensemble colorspace approach
- (32) Michal Myller, Michal Kawulok and Jakub Nalepa. Selecting Features from Time Series Using Attention-based Recurrent Neural Networks
- (35) Antonella Mensi, Alessio Franzoni, David M. J. Tax and Manuele Bicego. An Alternative Exploitation of Isolation Forests for Outlier Detection
- (36) Tariq Mahmood, Antonio Robles-Kelly, Syed S. Naqvi and Muhammad Arsalan. Residual Multiscale Full Convolutional Network (RM-FCN) For High Resolution Semantic Segmentation of Retinal Vasculature
- (44) Shashank Gupta, Antonio Robels-Kelly and Mohamed Reda Bouadjenek. Feature Extraction Functions for Neural Logic RuleLearning
- (45) Sinan Kaplan, Joni Juvonen and Lasse Lensu. A Practical Hybrid Active Learning Approach for Human Pose Estimation
- (46) Danil Galeev, Konstantin Sofiiuk, Danila Rukhovich, Mikhail Romanov, Olga Barinova and Anton Konushin. Learning High-Resolution Domain-Specific Representations with a GAN Generator
- (48) Danila Rukhovich, Konstantin Sofiiuk, Danil Galeev, Olga Barinova and Anton Konushin. IterDet: Iterative Scheme for Object Detection in Crowded Environments
- (54) Alexander Welsing, Andreas Nienkötter and Xiaoyi Jiang. Exponential weighted moving average of time series in arbitrary spaces
- (60) Konstantinos Peppas, Konstantinos Tsiolis, Ioannis Mariolis, Angeliki Topalidou Kyniazopoulou and Dimitrios Tzovaras. Multi-Modal 3D Human Pose Estimation for Human-Robot Collaborative Applications
- (69) Peter Bellmann, Ludwig Lausser, Hans Kestler and Friedhelm Schwenker. Experimental Analysis of Bidirectional Pairwise Ordinal Classifier Cascades
Friday 22 Jan 2021 – SSPR
11:00 – 12:00 (CET): Keynote (Max Welling)
On Bayesian bits (babits), Quantum bits (qubits) and Quantum Bayesian bits (qubabits) for Deep Learning.
The bit is the fundamental unit of information. Its value is exclusively x=0 or x=1. However, under uncertainty we can assign a probability to a bit-value p(x=1). We take a Bayesian point of view w.r.t. the bits that encode the parameters of a neural network. Each parameter is written as a finite bitstring. More significant bits are expected to be less uncertain in the posterior than less significant bits, given data. We formulate a posterior model p(x(i)|x(<i)) for each bit in each parameter and infer from data if it should be included. When no Bayesian bits (babits) survive the culling, the parameter is pruned. A direct generalization of babits are qubits where each bit is generalized to a point on the 2-sphere. These are the fundamental units of quantum information. We show how to build a neural architecture from qubits by first designing a model for a classical binary net on a quantum computer, and subsequently deforming that model to entangle the qubits. Since we represent our parameters also as learnable qubits one could say that we have defined a Quantum Bayesian bit, or qubabit.
12:30 – 13:30 (CET): Papers discussion section
15:00 – 16:00 (CET): Pierre Devijver lecture (Fabio Roli)
From known knowns to unknown unknowns: a few reflections on my journey into pattern recognition
AI has been originally developed for closed-world, and noise-free, problems where the possible states of natures and actions that a rationale agent could implement were perfectly known. Using the words of a famous speech by Donald Rumsfeld, one could argue that, at that time, AI dealt with known knowns. I started my journey into pattern recognition dealing with known knowns, realizing a few years later that I was tackling problems involving known unknowns, noisy data which demand probability theory to model uncertainty and decision theory to minimize the risk of wrong actions.
During my journey I lived the rise of benchmark data sets, larger and larger year after year, and the belief that real world problems can be solved collecting enough training data. I had first-hand experience of what David Hand called the “illusion of progress”, and how available data sets have often a limited utility when used to train pattern recognition algorithms that will be deployed in the real world. I met unknown unknowns, called adversarial examples in machine learning, working on practical cybersecurity problems, and this brought me back to the eighties, to the fundamental limitations of learning from data that can contain spurious correlations and exhibit distribution shifts.
In this talk, I want to share a few reflections on my 32-year modest journey, placing it within the evolution of modern pattern recognition.
- (1) Kaspar Riesen, Hans Friedrich Witschel and Loris Grether. A Novel Data Set for Information Retrieval on the Basis of Subgraph Matching
- (13) Linlin Jia, Benoit Gaüzère and Paul Honeine. A Graph Pre-image Method Based on Graph Edit Distances
- (14) Francesco Pelosin, Andrea Gasparetto, Andrea Albarelli and Andrea Torsello. Unsupervised semantic discovery through visual patterns detection
- (25) Haoran Zhu, Hui Wu, Jianjia Wang and Edwin Hancock. Weighted Network Analysis using the Debye Model
- (27) Kieu Diem Ho, Jean-Yves Ramel and Nicolas Monmarché. Multivalent Graph Matching Problem Solved By Max-Min Ant System
- (28) Fatemeh Ansarizadeh, David B. Tay, Dhananjay Thiruvady and Antonio Robles-Kelly. Augmenting Graph Convolutional Neural Networks with Highpass Filters
- (30) Marco Brighi, Annalisa Franco and Dario Mario. Metric Learning for Multi-label Classification
- (38) Linlin Jia, Benoit Gaüzère, Florian Yger and Paul Honeine. A Metric Learning Approach to Graph Edit Costs for Regression
- (51) Luca Rossi and Andrea Torsello. Estimating the Manifold Dimension of a Complex Network Using Weyl’s Law
- (58) Keith Dillon. Efficient Partitioning of Partial Correlation Networks
- (59) Vincenzo Carletti, Pasquale Foggia, Antonio Greco and Mario Vento. Parallel Subgraph Isomorphism on multi-core architectures: a comparison of four strategies based on tree search
- (63) Vincenzo Carletti, Pasquale Foggia, Antonio Greco, Antonio Roberto and Mario Vento. Predicting Polypharmacy Side Effects through a relation-wise Graph Attention Network
- (66) Lixin Cui, Lichi Zhang, Lu Bai, Yue Wang and Edwin Hancock. Alzheimer’s Brain Network Analysis Using Sparse Learning Feature Selection
- (67) Huan Li, Boyuan Wang, Lixin Cui, Lu Bai and Edwin Hancock. LGL-GNN: Learning Global and Local Information for Graph Neural Network
- (68) Boyuan Wang, Lixin Cui, Lu Bai and Edwin Hancock. Graph Transformer: Learning Better Representations for Graph Neural Network
- (70) Darshan Batavia, Rocio Gonzalez-Diaz and Walter G. Kropatsch. Image = Structure + Few Colors
- (75) Francisco Escolano, Miguel Angel Lozano and Edwin Hancock. The Entropy of Graph Embeddings: A proxy of Potential Mobility in Covid19 Outbreaks