Day 1 - 30th SeptemberWednesday, 30 September 2020

Digital innovation is not possible without the tech skills to drive it. Most leaders today don’t have visibility into the technical abilities of their teams, making it impossible to build skills at scale.

Join longtime AI leader and head of learner experiences at Pluralsight, James Aylward for an in-depth conversation about his experience launching and scaling AI products, upskilling for modern technology roles, and staying adaptable to innovate faster.

Intended Audience:

This talk is relevant for any technology leader who is looking to develop or scale AI skills in their organization.

Takeaways:

• Technologists need modern skills as roles and technologies change over time
• AI/ML and adjacent cloud and infrastructure skills evolve quickly; so should your organization
• If you can become adaptable and keep up with the pace of change, you can build transformative products

By the end of this decade, AI will be practicing law and medicine at levels beyond human ability. It will be deeply and irrevocably involved in the production and distribution of our food supply, keeping order in society and running world commerce. It will be the reason we prevent pandemics and make progress on climate change. Our future and that of AI are inseparable. But the shape of this future and our path toward it are not yet set in stone. The default path of leaving it to the whim of Big Tech, like we did with software, is a dangerous option. AI is imminently more powerful and harder to control than software, which is widely recognized for exacerbating inequality, increasing extremism and sewing societal discord. If there is a way to do better with AI, we have a moral obligation to find it because once AI has its grip on humanity, it will be hard to make it let go. This talk is about what we can do to make the world better with AI.

Inversion problems arise in many petro-physical, seismic or more generally oil and gas applications where a physics based forward equation models the behavior of a system. In many cases, especially when the system’s output can be directly observed, the forward model simply exist for the sake of inversion; a technique for estimating a set of parameters, for which the physics based forward model can replicate the system’s observations.

Inverse problems is solved in a number of different ways, among which is the quasi-Newton’s method. The quasi-Newton’s method admits numerical approximations to allow for practical considerations especially in the computation of the Hessian, which is obtained and used along with the Jacobian and step length to find the minimum of a loss function for which the estimated set of parameters is sufficiently close to

However, in spite of the use of heuristics like the Hessian approximation or the estimated step length from an inexact line search formulation or a relaxation of strong Wolfe conditions during the inexact computation of the step length, the computational burden for real-time waveform inversion for either velocity or pressure reconstruction still poses a problem. Compounding the issue, is the reluctance of organizations to migrate from numerical solutions to data science models, at least not entirely.

To address organizational concerns about the fidelity of a complete data science solution, as well as the need for simplified online computational complexity, a hybrid Physics-Data science approach is introduced. This hybrid approach focuses on improving the time complexity of the forward model. The forward model (typically a partial differential equation) has an asymptotic run time of O(n3) (n correlates with the resolution of the solution). For example, a finite element analysis involving 100,000 grid points would invert a 100,000 X 100,000 matrix for each solution of the forward model in each iteration of the inversion algorithm. Since the quasi Newton’s method admits a lot of approximations for the inverse model, we introduce a data science proxy forward model, that can be trained offline to learn the forward equation, and apply it very quickly during online inversion. Consequently, iterations of the inverse model, compute faster, and converges to a solution in a fraction of the original time.

Driving in San Francisco is hard. Teaching an AV to drive the city's chaotic streets is even harder. In fact, an AV driving the streets of SF experiences a challenging situation 46x more often than in suburban areas. One of the roots of this challenge is predicting what other agents on the road are going to do next.

In particular, the long tail of scenarios the car encounters infrequently but are often the most dangerous: U-turns or K-turns, bicycles swerving into traffic, or pedestrians jaywalking unexpectedly.

This session will cover the inner workings of Cruise's Continuous Learning Machine and explore Cruise's approach to building ML solutions that generalize across the long tail of uncommon scenarios:

• Understanding the AV space: what is the prediction problem and why do we use ML to solve it
• Using self-supervised learning to "auto-label" data at massive scale with high throughput
• Addressing dataset imbalance and model generalization challenges by automatically mining generic error situations
• Tying it all together into a continuous learning machine!

Mystique – The Fight Against The Registration Plate Cloner - Deploying and Scaling an Object Detection / Image Classification Model to Process Half a Million Photos a Day

Mystique detects and blurs rego plate of cars on our portal, protecting seller’s privacy against rego plate cloner who would have been able to falsify the rego plate photo and attach it to another vehicle to conduct criminal activities.

The talk will cover:

• How do we build the rego detection/image classification model
• Clever technique to significantly reduce the false positive
• Deploying the tech at scale to process half a million photos a day

• Introduction
• Need of AI for B2B Media & Telecom companies
• Innovations in Content Viewing Experience
• Q & A

• Streaming video quality of experience measurement
• Optimal encoding, delivery and rendering of graphics and animation
• Virtual humans
• Actor placement and compositing (especially motivated by the pandemic)
• Virtual audiences