Facebook AI Research (FAIR) today unveiled PyTorch3D, a library that enables researchers and developers to combine deep learning and 3D objects. As part of the release, Facebook is also open-sourcing Mesh R-CNN, a model introduced last year capable of rendering 3D objects from 2D shapes in images of interior spaces. PyTorch3D was inspired by Mesh R-CNN and recent 3D work by Facebook AI Research, FAIR engineer Nikhila Ravi said. Working in 3D is important for rendering 3D objects or scenes that appear in mixed reality or virtual reality. It can also be used to tackle AI challenges like robotic grasping or helping autonomous vehicles understand the position of nearby objects.

Rendering a simple shape into a proper object with geometry, texture, and other material properties is a painstakingly long process; however, with AI, researchers can now do this rendering ten times faster than the real-time. A machine learning model is trained on images that are closer to the target. When it is presented with a shape and matching properties, it would recommend a photorealistic image. This opened a whole new field altogether -- differentiable programming. Traditional rendering engines are not differentiable, so they can't be incorporated into deep learning pipelines.

Researchers have developed an artificial neural network using deep learning to identify genes that are related to disease. An artificial neural network has revealed patterns in huge amounts of gene expression data and discovered groups of disease-related genes. The developers, from Linköping University, Sweden, hope that the method can eventually be applied within precision medicine and individualised treatment. The scientists created maps of biological systems based on how different proteins or genes interact with each other. Using artificial intelligence (AI), they investigated whether it is possible to discover biological networks with deep learning, in which entities known as artificial neural networks are trained by experimental data.

Sometimes these notebooks find their way into production, but their code and structure are often far from ideal. In this session, we cover some best practices around creating and operationalising notebooks. We will talk about structure, code style, refactoring in notebooks, unit testing, reproducibility and more. Nikolay Manchev is a machine learning enthusiast and speaker. His area of expertise is Machine Learning and Data Science, and his research interests are in neural networks with emphasis on biological plausibility. Nikolay was a Senior Data Scientist and Developer Advocate at IBM [masked]) and currently acts as the Principal Data Scientist for EMEA at Domino Data Lab. Talk 3: Generative Deep Learning - The Key To Unlocking Artificial General Intelligence by David Foster Generative modelling is one of the hottest topics in AI. It's now possible to teach a machine to excel at human endeavours such as painting, writing, and composing music. In this talk, we will cover: - A general introduction to Generative Modelling - A walkthrough of one of the most utilised generative deep learning models - the Variational Autoencoder (VAE) - Examples of state-of-the-art output from Generative Adversarial Networks (GANs) and Transformer based architectures.

Hopsworks 1.x series brings many new features and improvements, ranging from services such as the Feature Store and Experiments, to enhanced support for distributed stream processing and analytics with Apache Flink and Apache Beam, to building Deep Learning pipelines with TensorFlow Extended (TFX), to code versioning support for Jupyter notebooks with Git, to all-new provenance/lineage of data across all steps of a data engineering and data science. We are also excited that Hopsworks 1.x is the back-bone of the all new Managed Hopsworks platform for AWS, Hopsworks.ai Hopsworks 1.x brings significant Feature Store improvements ranging from updated UI components to connectivity with external systems and feature discovery. Users of Hopsworks Enterprise can now easily connect to the Feature Store from their Databricks notebooks and Amazon Sagemaker. Documentation for connecting with these two platforms can be found at hopsworks.readthedocs.io

Recently smoothing deep neural network based classifiers via isotropic Gaussian perturbation is shown to be an effective and scalable way to provide state-of-the-art probabilistic robustness guarantee against $\ell_2$ norm bounded adversarial perturbations. However, how to train a good base classifier that is accurate and robust when smoothed has not been fully investigated. In this work, we derive a new regularized risk, in which the regularizer can adaptively encourage the accuracy and robustness of the smoothed counterpart when training the base classifier. It is computationally efficient and can be implemented in parallel with other empirical defense methods. We discuss how to implement it under both standard (non-adversarial) and adversarial training scheme. At the same time, we also design a new certification algorithm, which can leverage the regularization effect to provide tighter robustness lower bound that holds with high probability. Our extensive experimentation demonstrates the effectiveness of the proposed training and certification approaches on CIFAR-10 and ImageNet datasets.

The current understanding of deep neural networks can only partially explain how input structure, network parameters and optimization algorithms jointly contribute to achieve the strong generalization power that is typically observed in many real-world applications. In order to improve the comprehension and interpretability of deep neural networks, we here introduce a novel theoretical framework based on the compositional structure of piecewise linear activation functions. By defining a direct acyclic graph representing the composition of activation patterns through the network layers, it is possible to characterize the instances of the input data with respect to both the predicted label and the specific (linear) transformation used to perform predictions. Preliminary tests on the MNIST dataset show that our method can group input instances with regard to their similarity in the internal representation of the neural network, providing an intuitive measure of input complexity.

Deep reinforcement learning has the potential to train robots to perform complex tasks in the real world without requiring accurate models of the robot or its environment. A practical approach is to train agents in simulation, and then transfer them to the real world. One popular method for achieving transferability is to use domain randomisation, which involves randomly perturbing various aspects of a simulated environment in order to make trained agents robust to the reality gap. However, less work has gone into understanding such agents - which are deployed in the real world - beyond task performance. In this work we examine such agents, through qualitative and quantitative comparisons between agents trained with and without visual domain randomisation. We train agents for Fetch and Jaco robots on a visuomotor control task and evaluate how well they generalise using different testing conditions. Finally, we investigate the internals of the trained agents by using a suite of interpretability techniques. Our results show that the primary outcome of domain randomisation is more robust, entangled representations, accompanied with larger weights with greater spatial structure; moreover, the types of changes are heavily influenced by the task setup and presence of additional proprioceptive inputs. Additionally, we demonstrate that our domain randomised agents require higher sample complexity, can overfit and more heavily rely on recurrent processing. Furthermore, even with an improved saliency method introduced in this work, we show that qualitative studies may not always correspond with quantitative measures, necessitating the combination of inspection tools in order to provide sufficient insights into the behaviour of trained agents.

Click-through rate (CTR) prediction is a crucial task in online display advertising and the key part is to learn important feature interactions. The mainstream models are embedding-based neural networks that provide end-to-end training by incorporating hybrid components to model both low-order and high-order feature interactions. These models, however, slow down the prediction inference by at least hundreds of times due to the deep neural network (DNN) component. Considering the challenge of deploying embedding-based neural networks for online advertising, we propose to prune the redundant parameters for the first time to accelerate the inference and reduce the run-time memory usage. Most notably, we can accelerate the inference by 46X on Criteo dataset and 27X on Avazu dataset without loss on the prediction accuracy. In addition, the deep model acceleration makes an efficient model ensemble possible with low latency and significant gains on the performance.

We introduce a novel approach to optimize the architecture of deep neural networks by identifying critical neurons and removing non-critical ones. The proposed approach utilizes a mixed integer programming (MIP) formulation of neural models which includes a continuous importance score computed for each neuron in the network. The optimization in MIP solver minimizes the number of critical neurons (i.e., with high importance score) that need to be kept for maintaining the overall accuracy of the model. Further, the proposed formulation generalizes the recently considered lottery ticket optimization by identifying multiple "lucky" sub-networks resulting in optimized architecture that not only perform well on a single dataset, but also generalize across multiple ones upon retraining of network weights. Finally, the proposed framework provides significant improvement in scalability of automatic sparsification of deep network architectures compared to previous attempts. We validate the performance and generalizability of our approach on MNIST, Fashion-MNIST, and CIFAR-10 datasets, using three different neural networks: LeNet 5 and two ReLU fully connected models.