dotAI 2018
as the organisers put it
With world-class experts on stage, we will explore the current state-of-the-art, the latest machine learning frameworks and everything there is to know about building intelligent applications in 2018.
who it's for
machine learning engineers, data engineers, ML platform engineers, research scientists, applied AI researchers, product engineers
topics
distributed training at scale; neural architecture search; billion-image pretraining; identifier embeddings for code; MXNet model serving; explainable AI with LIME; crowdsourced reinforcement learning; network-constrained lasso selection; Unicode-formula model transpilation; scalable storage and I/O for ML; AutoML workflows
| title | author | length |
|---|---|---|
| All AI Roads Lead to Distribution | Jim Dowling | 17m |
Claims that distribution both cuts training time and enables larger experiments with concrete numbers and tools: ImageNet is ~300 GB and 1M images, Google and… | ||
| Processing 100+ million images per month with deep learning | Eliot Andres | 04m |
Eliot explains how to use deep learning models in near real-time with a large amount of data, taking as an example Linkfluence who is processing more than 100… | ||
| Machine Learning on Source Code | Vadim Markovtsev | 16m |
Claims that identifier embeddings capture code naturalness by mapping millions of names to dense vectors, trains them from an explicit PMI co-occurrence matrix… | ||
| Logistics and data: towards a decision support tool | Charlotte Ledoux | 05m |
In logistics services, the only way to grow is to reduce costs, as this is a cost centre for many managers. On average, in France and in all business sectors… | ||
| Software 2.0, a Babel fish for deep learning? | Olivier Wulveryck | 05m |
Claims that mathematical models written as Unicode formulas can be parsed and transpiled into an execution graph that separates the learned weights from the… | ||
| Traps to avoid when working with AI | Virginie Mathivet | 15m |
Three concrete traps and what actually fixes them: the accuracy paradox with 99% accuracy on highly imbalanced defect images and augmentation that accidentally… | ||
| Deep Learning in the Wild | Isabel Schwende | 05m |
Journey from biomedical handcrafted features to production exposes three concrete gaps: models hailed by benchmarks often fail on field data, an mean average… | ||
| Knowledge Graphs & Deep Learning at YouTube | Aurelien Geron | 19m |
Aurelien explains how you can combine Knowledge Graphs and Deep Learning to dramatically improve Search & Discovery systems. By using a combination of signals… | ||
| Crowdsourcing Artificial Intelligence for Science: YOU can do a lot! | Sharada Mohanty | 20m |
A crowdsourced reinforcement learning challenge uses the crowdAI platform and a donated NVIDIA DGX‑1 to teach a simulated skeleton named Bob with 18 muscles to… | ||
| Using structure to select features in high dimension | Chloé-Agathe Azencott | 17m |
Genomics-scale setting with ~1,000 samples and hundreds of thousands to millions of features, claim: using feature-structure improves selection and prediction… | ||
| Model Serving for Deep Learning | Hagay Lupesko | 18m |
Open source Model Serving for Apache MXNet provides a concrete deploy path: pip install MXNet model server installs a CLI that downloads a 5 MB SqueezeNet… | ||
| Building Trust through Explainable AI | Peet Denny | 16m |
Shows a concrete method for post hoc explanations by using LIME (local interpretable model agnostic explainer) to jiggle features and fit a simple local… | ||
| nothing matches | ||