Self-supervised depth estimators have shown results comparable to the supervised methods on the challenging Single Image Depth Estimation task by exploiting relations between target and reference views in the data. However, previous works have not effectively neglected occlusions between the target and the reference images and rely on rigid photometric assumptions or on the SIDE network to infer depth and occlusions, resulting in limited performance. In this paper, we propose a method to “Forget About the LiDAR” (FAL), with Mirrored Exponential Disparity (MED) probability volumes for the training of monocular depth estimators from stereo images. Our MED representation allows us to obtain geometrically inspired occlusion maps with our novel Mirrored Occlusion Module (MOM), which does not impose a learning burden on our FAL-net. Our FAL-net is remarkably light-weight and outperforms the previous state-of-the-art methods with 8x fewer parameters and 3x faster inference speeds on KITTI.

"Data programming has proven to be an attractive alternative to costly hand-labeling of data. In this paradigm, users encode domain knowledge into \emph{labeling functions}, heuristics that label a subset of the data noisily and may have complex dependencies. The effects on test set performance of a downstream classifier caused by label model misspecification are understudied--presenting a serious knowledge gap to practitioners, in particular since LF dependencies are frequently ignored. In this paper, we focus on modeling errors due to structure over-specification. Based on novel theoretical bounds on the modeling error, we empirically show that this error can be substantial, even when modeling a seemingly sensible structure."

We present a novel metric for generative modeling evaluation that uses divergence between dendrograms computed from training and generated data. Our approach, which borrows theoretical foundations from clustering analysis, is validated by sampling from real datasets and also on samples generated by a GAN during training, with results comparable to state-of-the-art metrics.

"Generative Adversarial networks (GANs) have become predominant in image generation tasks. Their success is attributed to the training regime which employs two models: a generator G and discriminator D that compete in a minimax zero sum game. Nonetheless, GANs are difficult to train due to their sensitivity to hyperparameter and parameter initialisation which often leads to vanishing gradients, non-convergence, or mode collapse where the generator is unable to create samples with different variations.

In this work, we propose a novel Generative Adversarial Stacked Convolutional Autoencoder (GASCA) model and a generative adversarial gradual greedy layer-wise learning algorithm designed to train Adversarial Autoencoders in an efficient and incremental manner. Our training approach produces images with significantly lower reconstruction error than vanilla joint training."

Presenters in this session: Model Misspecification in Multiple Weak Supervision Forget About the LiDAR: Self-Supervised Depth Estimators with MED Probability Volumes An evaluation metric for generative models using hierarchical clustering Generative Adversarial Stacked Autoencoders

"Evaluating difficulty and biases in machine learning models has become of ex- treme importance as current models are now being applied in real-world situations. In this paper we present a simple method for calculating a difficulty score based on the accumulation of losses for each sample during training. Our proposed method does not require any modification of the model neither any external supervision. We test and analyze our approach in two different settings that provide empirical evidence of the applicability of our method."

The interpretation of seismic data is vital for characterizing sediments' shape in areas of geological study. In seismic interpretation, deep learning becomes useful for reducing the dependence on handcrafted facies segmentation geometry and the time required to study geological areas. This work presents a Deep Neural Network for Facies Segmentation (DNFS) to obtain state-of-the-art results for seismic facies segmentation. Our network is trained using a combination of cross-entropy and Jaccard loss functions. The results show that DNFS with fewer parameters than StNet and U-Net, trained with a composite loss function and a dataset for binary-segmentation in a few minutes, can offer high detailed predictions for seismic facies segmentation.

Education in Latin America is a key factor for social development and poverty, but there are several barriers that stand in the way of a better education system. Build tools that empower students to improve their performance in the educational system is an important task to provide a better future for all of us.

Zero-shot classification (ZSC) is the task of learning predictors for classes not seen during training. The different proposals in the literature are commonly evaluated using standard category splits but little attention has been paid to the impact on performance under different class partitions. In this work we show experimentally that ZSC perform with strong variability with respect to the class partitions. We propose an ensemble learning method to attempt to mitigate it.

The authors for this QA session are: Unsupervised Difficulty Estimation Binary Segmentation of Seismic Facies Using Encoder-Decoder Neural Networks Study, Attend and Predict: Academic Performance Prediction using Transformers Performance Variability in Zero-Shot Classification

Key note speakers in this session: Aline Villavicencio

The poster session will be in GatherTown please follow the link below

Q& A for the key note and introduction to paper session III Fernanda Viegas

"Motivated by high-stakes decision-making domains like personalized medicine where user information is inherently sensitive, we design privacy preserving exploration policies for episodic reinforcement learning (RL). We first provide a meaningful privacy formulation using the notion of joint differential privacy (JDP)--a strong variant of differential privacy for settings where each user receives their own sets of output (e.g., policy recommendations). We then develop a private optimism-based learning algorithm that simultaneously achieves strong PAC and regret bounds, and enjoys a JDP guarantee. Our algorithm only pays for a moderate privacy cost on exploration: in comparison to the non-private bounds, the privacy parameter only appears in lower-order terms. Finally, we present lower bounds on sample complexity and regret for reinforcement learning subject to JDP."

Is there an effective Spanish Sentiment Analysis algorithm? The aim of this paper is to answer this question. The task is challenging because there are several dialects for the Spanish Language. Thus, identically written words could have several meanings and polarities regarding Spanish speaking countries. To tackle this multidialect issue we rely on a transfer learning approach. To do so, we train a BERT language model to ``transfer'' general features of the Spanish language. Then, we fine-tune the language model to specific dialects. BERT is also used to generate contextual data augmentation aimed to prevent overfitting. Finally, we build the polarity classifier and propose a fine-tuning step using groups of layers. Our design choices allow us to achieve state-of-the-art results regarding multidialect benchmark datasets.

Large scale natural disasters involve budgetary problems for governments. Prioritiz-1ing investment requires near real time information about the impact of the hazard in2different locations. However, such information is not available through sensors or3other devices specially in developing countries that do not have such infrastructure.4In this work, we use mobile phone activity data to infer the affected zones in the5Ecuadorian province of Manabí, after the 2016 earthquake, with epicenter in the6same province. We calculate a series of features to train a classifier based on the7K-Nearest Neighbors algorithm to detect affected zones with a 75% of precision.8We compare our results with official reports published two months after the disaster

COVID-19 has brought about a significant challenge to the whole of humanity, but mainly to the medical community. Clinicians must keep updated continuously about symptoms, diagnoses, and effectiveness of emergent treatments under a never-ending flood of scientific literature. In this context, the role of evidence-based medicine (EBM) for curating the most substantial evidence to support public health and clinical practice turns especially essential but is being challenged as never before. Artificial Intelligence can have a crucial role in this situation. In this article, we report the results of an applied research project to classify scientific articles to support Epistemonikos, one of the essential foundations worldwide conducting EBM. We test several methods, and the best one, based on XLNet, improves the current approach by 93% on average F1-score, saving valuable time from physicians who volunteer to curate COVID-19 research articles manually.

The papers for this session are:

Private Reinforcement Learning with PAC and Regret Guarantees Overcoming Transformer Fine-Tuning process to improve Twitter Sentiment Analysis for Spanish Dialects Detecting Damaged Regions after Natural Disasters using Mobile Phone Data: The Case of Ecuador Neural language models for text classification in evidence-based medicine

Keynote speakers: Nayat Sanchez-Pi

"Increases in flash rates detected in ground-based lightning data can be a precursor to severe weather hazards. Lightning data from the Geostationary Lightning Mapper (GLM) aboard the GOES-16 satellite is not currently a systematic part of an operational model used by forecasters and is underutilized in severe storm research. We harness the spatial and temporal advantages of geostationary satellite data to create a machine learning model that augments the current meteorological practices and capabilities of forecasters during extreme weather. Our results suggest that false alarms for warned thunderstorms could be decreased by half, and that tornadoes and severe hail could be correctly identified 8 out of 10 times using GLM data.

Random Graphs are random objects which take its values in the space of graphs. We take advantage of the expresibility of graphs in order to model the uncertainty about the existence of causal relationships within a given set of variables. We adopt a Bayesian point of view in order to capture a causal structure via interaction and learning with a random environment. We test our method on a simple scenario, and the experiment confirms that our technique can learn a causal structure. Furthermore, the experiment presented demonstrate the usefulness of our method to learn an optimal action.

We present a new accelerated distributed algorithm for the robust solution of convex optimization problems over networks. We propose a novel distributed restarting mechanism for accelerated optimization dynamics with individual asynchronous time-varying coefficients. Graph-dependent restarting conditions are derived to establish suitable stability, convergence, and robustness properties for problems characterized by strongly convex smooth and non-smooth primal functions. Since the algorithm combines continuous-time dynamics and discrete-time dynamics, we model the complete system as a hybrid dynamical system. Numerical results illustrate our results.

The papers for this session are:

Severe Weather Prediction Using Lightning Data Learning a causal structure: a Bayesian Random Graph approach Robust Optimization over Networks Using Distributed Restarting of Accelerated Dynamics

Searching for new biometric traits is currently a necessity because traditional ones such as fingerprint, voice, or face are highly prone to forgery. A motivation for using electroencephalogram signals is that they are unique to each person and are much more difficult to replicate than conventional biometrics. This study aims to analyze the factors that influence the performance of a biometric system based on electroencephalogram signals. This work uses six different classifiers to compare several decomposition levels of the discrete wavelet transform as a preprocessing technique and also explores the importance of the recording time. This work proves that the decomposition level does not have a high impact on the system's overall result. On the other hand, the recording time of electroencephalograms has a significant impact on the classifiers' performance.

Early wildfire detection is of vital importance to prevent the damage caused by wildfires to both the environment and human beings. We propose a Deep Learning (DL) model that leverages the information fusion of visible and infrared images for accurate wildfire detection in controlled datasets; we expect the said model to display a lower rate of false-positives in comparison with current techniques. To this end, it is necessary to first investigate, analyze, and test existing early wildfire detection and image fusion methods. Additionally, we will create a dataset comprised of fused visible-infrared images. In the present paper, we introduce the proposed approach and some preliminary results regarding the evaluation of two state-of-the-art image fusion techniques on the Corsican Fire Dataset, as well as advances towards the fused image dataset generation.

Wildfires have been on the rise during the past five years. For this reason, the management of wildfires has become critical for the future. This paper shows the comparison between a Deep Learning model and other image processing methods for segmenting infrared images of fire into three zones delimited by temperature. The goal of the test results presented in this paper, and the subsequent tests that will follow it, is to provide insight into the usage of Deep Learning for this new and specific segmentation task. We will ultimately use the knowledge obtained from the tests while building a wildfire detection system that will provide fire engineers with important information for the management of forest fires.

As reinforcement learning agents become increasingly integrated into complex, real-world environments, designing for safety becomes a critical consideration. We introduce Safety Aware Reinforcement Learning (SARL) - a framework where a virtual safe agent modulates the actions of a main reward-based agent to minimize side effects. Here, a safe agent learns a task-independent notion of safety for a given environment. The main agent is then trained with a regularization loss given by the distance between the native action probabilities of the two agents, allowing us to learn a task-independent notion of safety. This notion can then be ported to modulate multiple policies solving different tasks within the given environment without further training. We contrast this with solutions that rely on task-specific regularization metrics and test our framework on the SafeLife Suite, based on Conway's Game of Life, comprising a number of complex tasks in dynamic environments.

This study aims to exploit driving data collected from sensors embedded in our own smartphones. In this way, the smartphone can be made to act as our copilot and make us aware of risky decisions that we make behind the driving wheel. By using smartphones rather than, for example, a) Aftermarket systems, b) On Board Diagnostics (OBD) systems and, c) Vehicle-mounted video cameras, we achieve broad applicability of this study by allowing practically any driver in any car to be supported by this technology. As we know, smartphones come integrated with a variety of sensors, some of which (accelerometer, gyroscope, magnetometer) can be used to measure accelerations and forces that the smartphone itself experiences. We exploit this sensor data by means of a deep architecture composed of Inception layers and gated units that together learns the optimal amount of source to improve classification performance on driving maneuvers.

We consider the model of cooperative learning via distributed non-Bayesian learning, where a network of agents tries to jointly agree on a hypothesis that best described a sequence of locally available observations. Building upon recently proposed weak communication network models, we propose a robust cooperative learning rule that allows asynchronous communications, message delays, unpredictable message losses, and directed communication among nodes.

Since the introduction of DQN, a vast majority of reinforcement learning research has focused on reinforcement learning with deep neural networks as function approximators. New methods are typically evaluated on a set of environments that have now become standard, such as Atari 2600 games. While these benchmarks help standardize evaluation, their computational cost has the unfortunate side effect of widening the gap between those with ample access to computational resources, and those without. In this work we argue that, despite the community's emphasis on large-scale environments, the traditional small-scale environments can still yield valuable scientific insights and can help reduce the barriers to entry for underprivileged communities. To substantiate our claims, we empirically revisit paper which introduced the Rainbow algorithm [Hessel et al., 2018] and present some new insights into the algorithms used by Rainbow.

We explore Machine Learning techniques and Quantum Computing concepts that could be applied in High Energy Physics considering a phenomenological and theoretical view. In this framework, we show the main tools to explore the Standard Model extensions, decay process and the parameter space. With this set of tools, we want to explore the bounds and define exclusion regions, this results might be interesting for the next generation of colliders and could prove to be useful in the understanding of phenomena.

Several deep learning architectures have been proposed over the last years to deal with the problem of generating a written report given an imaging exam as input. Most works evaluate the generated reports using standard Natural Language Processing (NLP) metrics (e.g. BLEU, ROUGE), reporting significant progress. In this article, we contrast this progress by comparing state of the art (SOTA) models against weak baselines. We show that simple and even naive approaches yield near SOTA performance on most traditional NLP metrics. We conclude that evaluation methods in this task should be further studied towards correctly measuring clinical accuracy, ideally involving physicians to contribute to this end.

"We present a novel method for the automatic georeferencing of heterogeneous map images based on the analysis of the spatial relationships between their lines of text and the geographical locations they depict. Our approach differs from previous work in that the only input provided is the raster image, it does not require additional hint or metadata. The method is also designed to be highly tolerant to maps with different art styles, scales, orientations and cartographic projections. To accomplish this task we leverage the power of modern OCR (Optical Character Recognition) and geocoding services to generate a series of candidate ground control points (GCP) and then discriminate between them using a combination of clustering algorithms and graph analysis. Experimental results for 359 map images demonstrate the viability proposed method. We achieved a precision ranging from 80% to 97% and a recall higher than 60%."

Conventional CNN cannot guarantee the invariance response even to small changes in input images, for example, with brightness variations. This paper analyzes the performance of a novel convolutional layer in the Fourier domain capable of classifying images even with large alterations of their brightness. The results show that the performance of a given CNN architecture is significantly more resilient to large brightness changes if use the proposed layer.

"Around the globe, ticks are the culprit of transmitting a variety of bacterial, viral and parasitic diseases. The incidence of tick-borne diseases has drastically increased within the last decade, with annual cases of Lyme disease soaring to an estimated 300,000 in the United States alone. As a result, more efforts in improving lesion identification approaches and diagnostics for tick-borne illnesses is critical. The objective for this study is to build upon the approach used by Burlina et al. by using a variety of convolutional neural network models to detect tick-borne skin lesions. We expanded the data inputs by acquiring images from Google in seven different languages to test if this would diversify training data and improve the accuracy of skin lesion detection. The final dataset included nearly 6,080 images and was trained on a combination of architectures. We obtained an accuracy of 80.72% with our model trained on the DenseNet 121 architecture."

In real world applications of supervised learning methods, training and test sets are often sampled from the distinct distributions and we must resort to domain adaptation techniques. One special class of techniques is Covariate Shift Adaptation, which allows practitioners to obtain good generalization performance in the distribution of interest when domains differ only by the marginal distribution of features. Traditionally, Covariate Shift Adaptation is implemented using Importance Weighting which may fail in high-dimensional settings due to small Effective Sample Sizes (ESS). In this paper, we propose (i) a connection between ESS, high-dimensional settings and generalization bounds and (ii) a simple, general and theoretically sound approach to combine feature selection and Covariate Shift Adaptation. The new approach yields good performance with improved ESS.

This work explores the use of a triplet loss deep network setting for the forensic identification of speakers in Spanish. Within the framework we train a convolutional network to produce vector representations of speech spectrogram slices. Then we test how similar are these vectors for a given speaker and how dissimilar are compared with other speakers. Based on these metrics we propose the calculation of the Likelihood Radio which is a cornerstone for forensic identification.

Ambient sensing facilitates non-intrusive data collection within sensitive environments, but also escalates the complexity in associated classification. In this paper, we propose a methodology based on Transformer Neural Networks to classify activities performed by a resident of an ambient sensor based environment. We also propose a methodology to pre-train Transformers in a self-supervised manner, as a hybrid autoencoder-classifier model instead of using contrastive loss. By having consistency in ambient data collection, the quality of data is considerably more reliable, presenting the opportunity to perform classification with enhanced accuracy. Therefore, in this research we look to find an optimal way of using deep learning to classify human activity with ambient sensor data.

In this paper, we compare automated metrical pattern identification systems available for Spanish against extensive experiments done by fine-tuning language models trained on the same task. Despite being initially conceived as a model suitable for semantic tasks, our results suggest that BERT-based models retain enough structural information to perform reasonably well for Spanish scansion.

Social media trends are increasingly taking a significant role for the understanding of modern social dynamics. Platforms like Twitter enable democratic interaction within users, thus, providing an opportunity to inject influential ideas within formed communities. In this work, we take a look at how the Twitter landscape gets constantly shaped by automatically generated content.Twitter bot activity can bet raced via network abstractions which, we hypothesize, can be learned through state-of-the-art graph neural network techniques. We employ a large bot database,continuously updated by Twitter, to learn how likely is that a user is mentioned by a bot, as well as, for a hashtag. Furthermore, we model this likelihood as a link prediction task between the set of users and hashtags. Moreover, we contrast our results by performing similar experiments on a crawled data set of real users.