A good example of solving for the right problems can be seen in Formula One World Championship Ltd. The motorsport company was looking for new ways to deliver race metrics that could change the way fans and teams experience racing, but had more than 65 years of historical race data to sift through. After aligning their technical and domain experts to determine what type of untapped data had the most potential to deliver value for its teams and fans, Formula 1 data scientists then used Amazon SageMaker to train deep learning models on this historical data to extract critical performance statistics, make race predictions and relay engaging insights to their fans into the split-second decisions and strategies adopted by teams and drivers.

Our initial attempts at representing wave functions with neural networks were met with frustration when we couldn't reach the accuracy of even the standard Hartree–Fock method of quantum chemistry. Not even on the smallest molecules, where such a calculation takes only a few seconds on a modern computer. This eventually motivated us to build the Hartree–Fock baseline and other components enforcing the correct physics of wavefunctions into our architecture, which we dubbed PauliNet--a neural network that obeys the Pauli exclusion principle.2 We were then quite surprised in September 2019 when we found out that researchers at DeepMind, who pursued the same idea in parallel, were able to reach impressive accuracy without building any physics into their architecture, called FermiNet, although at the cost of using much larger networks and hence requiring more computational resources.3

This will be an interactive post using Google Colab notebooks. If you have not used Google Colab before, there is a quick-start tutorial at tutorialspoint. You can access the notebook at this link: Train your first DL model. First, make a copy and save it into your Drive so that you can access it and make changes. Next, make sure the runtime is set to GPU so you can make use of the free resources provided by Google.

These days we are hearing a lot about AI, but have you ever heard about EDGE AI ..? What does it mean and what is it used for? Network edge or edge, where data resides and collected. Edge computing processes data on local places like computers, IoT devices or Edge servers, here we are doing computation to a network edge which indeed reduces long-distance communication between client and server. Edge AI, where AI algorithms will locally process sensor data or signals that are created on hardware devices in less than a few milliseconds by providing real-time information. Most of the time the AI algorithms are being processed in cloud data centers with deep learning models, which consume heavy compute capacity.

Seven years ago, my student and I at Penn State built a bot to write a Wikipedia article on Bengali Nobel laureate Rabindranath Tagore's play "Chitra." First, it culled information about "Chitra" from the internet. Then it looked at existing Wikipedia entries to learn the structure for a standard Wikipedia article. Finally, it summarised the information it had retrieved from the internet to write and publish the first version of the entry. However, our bot did not "know" anything about "Chitra" or Tagore. It did not generate fundamentally new ideas or sentences.

Image matting is an essential technique to estimate the foreground objects in images and videos for editing and composition. The conventional deep learning approach takes the input image and associated trimap to get the alpha matte using convolution neural networks. But since the real-world input images for matting are mostly of very high resolution, such approaches efficiency suffers in real-world matting applications due to hardware limitations. To address the issue mentioned above, HD-Matt, the first deep learning-based image matting approach for high-resolution image inputs, is proposed by a group of researchers from UIUC (University of Illinois, Urbana Champaign), Adobe Research, and the University of Oregon. HD-Matt works on the'divide-and-conquer' principle.

Time series forecasting is an important area of machine learning. It is important because there are so many prediction problems that involve a time component. However, while the time component adds additional information, it also makes time series problems more difficult to handle compared to many other prediction tasks. Time series data, as the name indicates, differ from other types of data in the sense that the temporal aspect is important. On a positive note, this gives us additional information that can be used when building our machine learning model -- that not only the input features contain useful information, but also the changes in input/output over time.

We develop a recurrent gamma belief network (rGBN) for radar automatic target recognition (RATR) based on high-resolution range profile (HRRP), which characterizes the temporal dependence across the range cells of HRRP. The proposed rGBN adopts a hierarchy of gamma distributions to build its temporal deep generative model. For scalable training and fast out-of-sample prediction, we propose the hybrid of a stochastic-gradient Markov chain Monte Carlo (MCMC) and a recurrent variational inference model to perform posterior inference. To utilize the label information to extract more discriminative latent representations, we further propose supervised rGBN to jointly model the HRRP samples and their corresponding labels. Experimental results on synthetic and measured HRRP data show that the proposed models are efficient in computation, have good classification accuracy and generalization ability, and provide highly interpretable multi-stochastic-layer latent structure.

We study the problem of batch learning from bandit feedback in the setting of extremely large action spaces. Learning from extreme bandit feedback is ubiquitous in recommendation systems, in which billions of decisions are made over sets consisting of millions of choices in a single day, yielding massive observational data. In these large-scale real-world applications, supervised learning frameworks such as eXtreme Multi-label Classification (XMC) are widely used despite the fact that they incur significant biases due to the mismatch between bandit feedback and supervised labels. Such biases can be mitigated by importance sampling techniques, but these techniques suffer from impractical variance when dealing with a large number of actions. In this paper, we introduce a selective importance sampling estimator (sIS) that operates in a significantly more favorable bias-variance regime. The sIS estimator is obtained by performing importance sampling on the conditional expectation of the reward with respect to a small subset of actions for each instance (a form of Rao-Blackwellization). We employ this estimator in a novel algorithmic procedure---named Policy Optimization for eXtreme Models (POXM)---for learning from bandit feedback on XMC tasks. In POXM, the selected actions for the sIS estimator are the top-p actions of the logging policy, where p is adjusted from the data and is significantly smaller than the size of the action space. We use a supervised-to-bandit conversion on three XMC datasets to benchmark our POXM method against three competing methods: BanditNet, a previously applied partial matching pruning strategy, and a supervised learning baseline. Whereas BanditNet sometimes improves marginally over the logging policy, our experiments show that POXM systematically and significantly improves over all baselines.

Multicasting in wireless systems is a natural way to exploit the redundancy in user requests in a Content Centric Network. Power control and optimal scheduling can significantly improve the wireless multicast network's performance under fading. However, the model based approaches for power control and scheduling studied earlier are not scalable to large state space or changing system dynamics. In this paper, we use deep reinforcement learning where we use function approximation of the Q-function via a deep neural network to obtain a power control policy that matches the optimal policy for a small network. We show that power control policy can be learnt for reasonably large systems via this approach. Further we use multi-timescale stochastic optimization to maintain the average power constraint. We demonstrate that a slight modification of the learning algorithm allows tracking of time varying system statistics. Finally, we extend the multi-timescale approach to simultaneously learn the optimal queueing strategy along with power control. We demonstrate scalability, tracking and cross layer optimization capabilities of our algorithms via simulations. The proposed multi-timescale approach can be used in general large state space dynamical systems with multiple objectives and constraints, and may be of independent interest.