PAPIs 2018

Environment misconfiguration or upstream data pipeline inconsistency can silently kill the model performance.
Common production incidents include:
- Data drifts, new data, wrong features
- Vulnerability issues, adversarial attacks
- Concept drifts, new concepts, expected model degradation
- Dramatic unexpected drifts
- Biased Training set / training issue
- Performance issue
In this talk we'll discuss a solution, tooling and architecture that allows machine learning engineer to be involved in delivery phase and take ownership over deployment and monitoring of machine learning pipelines.

Freebird models US domestic flights in a way that captures uncertainty at every step. We present a case study of using Bayesian modelling and inference to directly model behavior of aircraft arrivals and departures, focusing on the uncertainty in those predictions. Along the way we will discuss theoretical considerations, highlighting what can go wrong, while emphasizing practical implications around scaling to large data sets.

In order to build robust NLP models that can reason from language well, architectures should function more similarly to how our human brains work over pure pattern recognition. Attention is an interpretable type of neural network layer that is loosely based on attention in humans, and it has recently enabled a powerful alternative to RNNs. Attention-based models have produced new techniques and state of the art performances for many language modeling tasks. In this presentation, an introduction to Attention layers will be given along with why and how they have been utilized to revolutionize NLP.

I will present research that I conducted during my Ph.D. at University College London and in collaboration with Google. My primary interest lays in the development of neural architectures for 2D imagery problems in big scale. Will present the recently published analysis of different upsampling methods in the decoder part of visual architectures, together with last week ongoing extension for GANs. Will discuss attention mechanism for text recognition and review for what kind of application it can be useful (automatically updating Google Maps based on Google Street View imagery).

In this talk, we describe recent work on the offline estimation of recommender system A/B tests using counterfactual reasoning techniques. We can determine whether our customers would have clicked on what we would have recommended by adding stochasticity to our recommendations. This ensures non-zero probability of having shown our new recommendations at some point in the past, which can leverage using a technique known as Pareto-smoothed importance sampling. This allows us to create a low-bias, low-variance estimator of how our recommender systems would have performed had they been deployed.