Deep learning (DL) became an overnight "star" when a robot player beat a human player in the famed game of AlphaGo. Deep learning training and learning methods have been widely acknowledged for "humanizing" machines. Many of the advanced automation capabilities now found in enterprise AI platforms are due to the rapid growth of machine learning (ML) and deep learning technologies. A Deep Dive into Deep Learning in 2019 comments on the "ubiquitous" presence of DL in many facets of AI -- be it NLP or computer vision applications. Gradually, AI and DL-enabled automated systems, tools, and solutions are penetrating and taking over all business sectors --from marketing to customer experience, from virtual reality to natural language processing (NLP) -- the digital impact is everywhere.

Homefeed is a discovery platform at Pinterest that helps users find and explore their interests. We work with some of the largest datasets in the world, tailoring over billions of unique content to 300M users. Our content ranges across all categories, from food and fashion to books, art and our dataset is rich with textual and visual content and has nice graph properties -- harnessing these signals at scale is a significant challenge. You'll have the opportunity to work on various machine learning and deep learning challenges, build the systems and machine learning models to improve user experience on Homefeed at scale.

Three years ago, AI became a new selling point for smartphones. For a time, smartphones supporting AI features quickly spread, and AI performance has become the most important mobile processor parameter after the CPU and GPU. Today, demonstrating AI performance has become an indispensable part of mobile phone and mobile processor conferences. When it comes to the AI function on mobile phones, many people can immediately think of it-smart voice assistant and AI camera. There is no doubt that AI photography is the most commonly used function by ordinary consumers.

How can we enable machines to make sense of the world, and become better at learning? To approach this goal, I believe viewing intelligence in terms of many integral aspects, and also a universal two-term tradeoff between task performance and complexity, provides two feasible perspectives. In this thesis, I address several key questions in some aspects of intelligence, and study the phase transitions in the two-term tradeoff, using strategies and tools from physics and information. Firstly, how can we make the learning models more flexible and efficient, so that agents can learn quickly with fewer examples? Inspired by how physicists model the world, we introduce a paradigm and an AI Physicist agent for simultaneously learning many small specialized models (theories) and the domain they are accurate, which can then be simplified, unified and stored, facilitating few-shot learning in a continual way. Secondly, for representation learning, when can we learn a good representation, and how does learning depend on the structure of the dataset? We approach this question by studying phase transitions when tuning the tradeoff hyperparameter. In the information bottleneck, we theoretically show that these phase transitions are predictable and reveal structure in the relationships between the data, the model, the learned representation and the loss landscape. Thirdly, how can agents discover causality from observations? We address part of this question by introducing an algorithm that combines prediction and minimizing information from the input, for exploratory causal discovery from observational time series. Fourthly, to make models more robust to label noise, we introduce Rank Pruning, a robust algorithm for classification with noisy labels. I believe that building on the work of my thesis we will be one step closer to enable more intelligent machines that can make sense of the world.

In this paper we develop a new model for deep image clustering, using convolutional neural networks and tensor kernels. The proposed Deep Tensor Kernel Clustering (DTKC) consists of a convolutional neural network (CNN), which is trained to reflect a common cluster structure at the output of its intermediate layers. Encouraging a consistent cluster structure throughout the network has the potential to guide it towards meaningful clusters, even though these clusters might appear to be nonlinear in the input space. The cluster structure is enforced through the idea of unsupervised companion objectives, where separate loss functions are attached to layers in the network. These unsupervised companion objectives are constructed based on a proposed generalization of the Cauchy-Schwarz (CS) divergence, from vectors to tensors of arbitrary rank. Generalizing the CS divergence to tensor-valued data is a crucial step, due to the tensorial nature of the intermediate representations in the CNN. Several experiments are conducted to thoroughly assess the performance of the proposed DTKC model. The results indicate that the model outperforms, or performs comparable to, a wide range of baseline algorithms. We also empirically demonstrate that our model does not suffer from objective function mismatch, which can be a problematic artifact in autoencoder-based clustering models.

In problem solving, the paths towards solutions can be viewed as a sequence of decisions. The decisions, made by humans or computers, describe a trajectory through a high-dimensional representation space of the problem. Using dimensionality reduction, these trajectories can be visualized in lower dimensional space. Such embedded trajectories have previously been applied to a wide variety of data, but so far, almost exclusively the self-similarity of single trajectories has been analyzed. In contrast, we describe patterns emerging from drawing many trajectories---for different initial conditions, end states, or solution strategies---in the same embedding space. We argue that general statements about the problem solving tasks and solving strategies can be made by interpreting these patterns. We explore and characterize such patterns in trajectories resulting from human and machine-made decisions in a variety of application domains: logic puzzles (Rubik's cube), strategy games (chess), and optimization problems (neural network training). In the context of Rubik's cube, we present a physical interactive demonstrator that uses trajectory visualization to provide immediate feedback to users regarding the consequences of their decisions. We also discuss the importance of suitably chosen representation spaces and similarity metrics for the embedding.

We envision that in the near future, humanoid robots would share home space and assist us in our daily and routine activities through object manipulations. One of the fundamental technologies that need to be developed for robots is to enable them to detect objects and recognize them for effective manipulations and take real-time decisions involving those objects. In this paper, we describe a novel architecture that enables multiple low-compute NAO robots to perform real-time detection, recognition and localization of objects in its camera view and take programmable actions based on the detected objects. The proposed algorithm for object detection and localization is an empirical modification of YOLOv3, based on indoor experiments in multiple scenarios, with a smaller weight size and lesser computational requirements. Quantization of the weights and re-adjusting filter sizes and layer arrangements for convolutions improved the inference time for low-resolution images from the robot s camera feed. YOLOv3 was chosen after a comparative study of bounding box algorithms was performed with an objective to choose one that strikes the perfect balance among information retention, low inference time and high accuracy for real-time object detection and localization. The architecture also comprises of an effective end-to-end pipeline to feed the real-time frames from the camera feed to the neural net and use its results for guiding the robot with customizable actions corresponding to the detected class labels.

Ultrasound (US) is one of the most commonly used imaging modalities in both diagnosis and surgical interventions due to its low-cost, safety, and non-invasive characteristic. US image segmentation is currently a unique challenge because of the presence of speckle noise. As manual segmentation requires considerable efforts and time, the development of automatic segmentation algorithms has attracted researchers attention. Although recent methodologies based on convolutional neural networks have shown promising performances, their success relies on the availability of a large number of training data, which is prohibitively difficult for many applications. Therefore, in this study we propose the use of simulated US images and natural images as auxiliary datasets in order to pre-train our segmentation network, and then to fine-tune with limited in vivo data. We show that with as little as 19 in vivo images, fine-tuning the pre-trained network improves the dice score by 21% compared to training from scratch. We also demonstrate that if the same number of natural and simulation US images is available, pre-training on simulation data is preferable.

The Information Bottleneck (IB) principle offers a unified approach to many learning and prediction problems. Although optimal in an information-theoretic sense, practical applications of IB are hampered by a lack of accurate high-dimensional estimators of mutual information, its main constituent. We propose to combine IB with invertible neural networks (INNs), which for the first time allows exact calculation of the required mutual information. Applied to classification, our proposed method results in a generative classifier we call IB-INN. It accurately models the class conditional likelihoods, generalizes well to unseen data and reliably recognizes out-of-distribution examples. In contrast to existing generative classifiers, these advantages incur only minor reductions in classification accuracy in comparison to corresponding discriminative methods such as feed-forward networks. Furthermore, we provide insight into why IB-INNs are superior to other generative architectures and training procedures and show experimentally that our method outperforms alternative models of comparable complexity.

In recent years, neural networks have been extensively deployed for computer vision tasks, particularly visual classification problems, where new algorithms reported to achieve or even surpass the human performance. Recent studies have shown that they are all vulnerable to the attack of adversarial examples. Small and often imperceptible perturbations to the input images are sufficient to fool the most powerful neural networks. Advbox is a toolbox to generate adversarial examples that fool neural networks in PaddlePaddle, PyTorch, Caffe2, MxNet, Keras, TensorFlow and it can benchmark the robustness of machine learning models. Compared to previous work, our platform supports black box attacks on Machine-Learning-as-a-service, as well as more attack scenarios, such as Face Recognition Attack, Stealth T -shirt, and DeepFake Face Detect.