Review – Agenda 2019

Deep learning certainly has roots in the autonomous vehicle space. However, most trucking companies have a substantial investment in existing class 8 semi-trailer trucks that are not going to be replaced overnight. Trimble Transportation Mobility is using deep learning technologies, in conjunction with other advanced analytic techniques and state of the art DevOps approaches to help ensure the safe operation of trucking fleets. While it may be premature for many trucking fleets to embrace autonomous vehicles, TTM has made it possible for those same companies to leverage deep learning as a way to reduce costs and improve safety.

Field issue (malfunction) incidents are costly for the manufacturer’s service department. A normal telematics system has difficulty in capturing useful information even with pre-set triggers. In this session, Yong Sun will discuss how a machine learning, deep learning based predictive software/hardware system has been implemented to solve these challenges by 1) identifying when a fault will happen 2) diagnosing the root cause on the spot based on time series data analysis. Yong Sun will cover a novel technique for addressing a lack of training data for the neural network based root cause analysis.

Machine learning has been sweeping our industry, and the creativity it is already enabling is incredible. On the flip side there has also been the emergence of technology like Deep Fakes with the possibility to spread disinformation. As a tool maker, is our technology neutral, or are we responsible for creating technology for good? How should we be thinking about biases of multiple forms when training AI? What can go wrong when learning is applied to indiscriminate user data? 

 At Adobe we look at this problem from multiple angles, from weighing the positives of technology against their possible misuses, researching detection technology for manipulated images, assembling diverse teams of experts, and having internal training and reviews of technology around Artificial Intelligence.

At PayPal achieving four nines of availability is the norm. In the pursuit of exponentially complex additional nines we have recently embarked on applying deep learning to forecasting datacenter metrics. With little shared to the open community, this talk will shine light on how we apply Seq2Seq networks to forecasting CPU and Memory metrics at scale. In sharing ideas around building deep networks for forecasting; the talk will highlight how life of Data Scientists at PayPal has been greatly simplified by the use of Template Notebooks stitched into stateful and stateless pipelines using PayPal’s open source PPExtensions.

How much data is enough to build an accurate deep learning model? This one of the first and most difficult questions to answer early in any machine learning project. However, the quality and applicability of your data are more important considerations than quantity alone. This talk presents some insights and lessons learned for gauging the suitability of electronic health record (EHR) training data for a life underwriting project. You will see how to determine if more data might increase accuracy and how to identify any weaknesses a deep neural network might have as a result of your current training data.

On the forefront of deep learning research is a technique called reinforcement learning, which bridges the gap between academic deep learning problems and ways in which learning occurs in nature in weakly supervised environments. This technique is heavily used when researching areas like learning how to walk, chase prey, navigate complex environments, and even play Go. In this session, Martin Görner will detail how a neural network can be taught to play the video game Pong from just the pixels on the screen. No rules, no strategy coaching, and no PhD required. Martin will build on this application to show how the approach can be generalized to other problems involving a non-differentiable steps that cannot be trained using traditional supervised learning techniques.

This is a prelude to Martin’s full day workshop on Thursday, June 20th: Hands-On Deep Learning in the Cloud: Fast and Lean Data Science with Tensorflow, Keras, and TPUs.

Application of advanced analytics techniques like machine learning and deep learning is in the initial stages in airline industry. In this presentation, we outline how this technology is being applied and GE Aviation's experience in advancing this for the last 7 years.

In this talk, Chandra Kahtri, Senior AI Scientist at Uber AI, formerly at Alexa AI, will detail various problems associated with Conversational AI such as speech recognition, language understanding, dialog management, language generation, sensitive content detection and evaluation and the advancements brought by deep learning in addressing each one of these problems. He will also present on the applied research work he has done at Alexa and Uber for the problems mentioned above.

In applications like fraud and abuse protection, it is imperative to use progressive learning and fast retraining to combat emerging fraud vectors. However, somewhat unfortunately, these scenarios also suffer from the problem of late-coming supervision (such as late chargebacks), which makes the problem even more challenging! If we use a direct supervised approach, a lot of the valuable sparse supervision signal gets wasted on figuring out the manifold structure of data before the model actually starts discriminating newly emerging fraud. At Microsoft we are investigating unsupervised learning, especially auto encoding with deep networks, as a preprocessor that can help tackle this problem. An auto-encoding network, which is trained to reconstruct (in some sense) the input features through a constriction, learns to encode the manifold structure of the data into a small set of latent variables, similar to how PCA encodes the dominant linear eigen spaces. The key point is that the training of this auto-encoder happens with the abundant unlabeled data – it does not need any supervision. Once trained, we then use the auto-encoder as a featurizer that feeds into the supervised model proper. Because the manifold structure is already encoded in the auto-encoded bits, the supervised model can immediately start learning to discriminate between good and bad manifolds using the precious training signal that flows in about newly emerging fraud patterns. This effectively improves the temporal tracking capability of the fraud protection system and significantly reduces fraud losses. We will share some promising early results we have achieved by using this approach.