Patrick Haffner

Also known as Patrick G. Haffner

Patrick Haffner is a machine learning scientist whose career spans speech recognition, document imaging, and human-in-the-loop AI. He earned an engineering mathematics degree at École Polytechnique and a PhD in computer science and signal processing in Paris, then built speech networks and a bank-check reader. At AT&T and Interactions, he worked on DjVu, check-recognition networks, conversational systems, and sequence models; at AWS, he coauthored vision-language and annotation research. As of October 2026, he is a principal applied scientist at AWS.

Works: 83
Citations: 63,287
h-index: 23
Badges: Bridge
Loyal

Publication numbers from OpenAlex, as of 22 October 2026.

Timeline

1987–1989

Engineering mathematics and a Paris PhD

He earns an engineering mathematics degree at École Polytechnique in 1987. In 1989, he earns a PhD in computer science and signal processing at École Nationale Supérieure des Télécommunications in Paris.
1990–1995

Speech networks at France Telecom

From 1990 to 1995, he works as a research scientist at France Telecom Research Laboratories. In 1991, he and Alex Waibel apply Multi-State Time Delay Neural Networks to speaker-independent, telephone-quality isolated digit recognition and report a 1.6% error rate. In 1993, he presents a connectionist speech architecture with Forward-Backward alignment and a global one-pass training algorithm based on Maximum Mutual Information.
1995

The Courtesy Amount Reader

In 1995, he leads Bell Labs' Courtesy Amount Reader project for bank-check recognition.
1996–2001

Developing DjVu at AT&T Labs

From 1996 to 2001, he develops DjVu, a format for distributing scanned documents on the web, with Yann LeCun, Léon Bottou, Paul G. Howard, Patrice Simard, and Yoshua Bengio at AT&T Labs in Red Bank, New Jersey. He works in LeCun's Image Processing Research Department and presents “Color Documents on the Web with DjVu” at the International Conference on Image Processing in Kobe in October 1999.
1998

LeNet-5 and check reading

In 1998, he coauthors “Gradient-based learning applied to document recognition” in Proceedings of the IEEE with Yann LeCun, Léon Bottou, and Yoshua Bengio. The paper describes LeNet-5 and systems that read millions of checks per day. With LeCun, he leads the first industrial deployment of convolutional neural networks for check recognition.
2002–2014

Lead expert at AT&T Labs

In 2002, he becomes a lead expert at AT&T Labs Research and applies machine learning to natural-language and sequence processing, network data, and text analytics. He remains at AT&T through 2014. In 2005, he coauthors “ACAS: automated construction of application signatures.”
2014–2021

Lead inventive scientist at Interactions

In 2014, he becomes a lead inventive scientist at Interactions and leads machine-learning research on conversational systems, including algorithms that can say “I don't know” and call on humans for help. In 2017, he coauthors “Learning to Adapt by Minimizing Discrepancy,” which introduces the Temporal Neural Coding Network and a sequence-learning algorithm trained without back-propagation through time. In 2021, he coauthors “Deep Clustering with Measure Propagation,” which proposes DECAMP for short-text clustering.
2021–2026

Principal applied scientist at AWS

In 2021, he joins Amazon Web Services as a principal applied scientist focused on human-in-the-loop machine learning with the SageMaker Ground Truth team through 2026. In 2023, he coauthors a study of linguistic priors and compositional generalization in vision-language models; in 2024, he coauthors Snapper, a find-and-snap tool for bounding-box annotation in object detection, and Annorama, an at-desk virtual-reality annotation experience for 3D point clouds. As of October 2026, he is a principal applied scientist at AWS.