Titulo Estágio
Deep Learning Scalability
Áreas de especialidade
Sistemas Inteligentes
Sistemas Inteligentes
Local do Estágio
Coimbra
Enquadramento
Deep Learning has been very successful in the last decade, especially on unstructured data. In Feedzai's use-case, these models typically update embeddings representing the past transactional behavior of respecitve entities (hence, they are stateful models). This project would involve benchmarking latest methods that enable serving of stateful deep learning models, ensuring it (1) integrates with current infrastructure, (2) maintains low latencies required in a fraud detection setting (3) is scalable to high throughputs of events.
Objetivo
Deep Learning has been very successful in the last decade, especially on unstructured data. In Feedzai's use-case, these models typically update embeddings representing the past transactional behavior of respecitve entities (hence, they are stateful models). This project would involve benchmarking latest methods that enable serving of stateful deep learning models, ensuring it (1) integrates with current infrastructure, (2) maintains low latencies required in a fraud detection setting (3) is scalable to high throughputs of events.
Plano de Trabalhos - Semestre 1
"0. Onboarding on Feedzai's Model Serving
1. Literature review on Deep Learning frameworks inference frameworks capable of providing high-throughput and low-latency for Recurrent Neural Networks (that require tracking and reliably updating its state on a stream of input events). For instance, Pytorch, Tensorflow and Kubeflow might not have a reliable enough production-ready mechanism for this.
2. Propose architecture design for a system that efficiently performs RNN inference while updating the internal model state for many different entitites.
Expected results:
- Literature review
- Plan approach for second semester based on those findings"
Plano de Trabalhos - Semestre 2
"1. Benchmark existing open-source framework (if there is one)
2. Develop runtime Proof-of-Concept RNN inference service that updates the internal model state under Feedzai's ML Infra ecosystem.
3. Benchmark system regarding throughput and latency.
Expected results
- Implement self-hosted Proof-of-Concept deployed on kubernetes within Feedzai's ML Infra ecosystem
- Assess its performance in throughput and latency
- Documentation, final report and presentation"
Condições
Remunerated
Orientador
Alberto Ferreira
alberto.ferreira@feedzai.com 📩