“Allegro AI offers the first true end-to-end ML / DL product life-cycle management solution with a focus on deep learning applied to unstructured data.”
Machine learning projects involve iterative and recursive R&D process of data gathering, data annotation, research, QA, deployment, additional data gathering from deployed units and back again. The effectiveness of a machine learning product depends on how intact the synergies are between data, model and various teams across the organisation.
In this informative session at CVDC 2020, a 2 day event organised by ADaSci, Dan Malowany of Allegro.AI presented the attendees with the best practices to imbibe during the lifecycle of an ML product—from inception to production.
Dan Malowany is currently the head of deep learning research at allegro.ai. His Ph.D. research at the Laboratory of Autonomous Robotics (LAR) was focused on integrating mechanisms of the human visual system with deep convolutional neural networks. His research interests include computer vision, convolutional neural networks, reinforcement learning, the visual cortex and robotics.
Art Of Research Management
Dan spoke about the features required to boost productivity in the different R&D stages. This talk specifically focused on the following:
- Data management,
- Experiment management,
- Resource management,
- ML pipelines and Auto ML.
Dan, who has worked for 15 years at the Directorate for Defense Research & Development and led various R&D programs, briefed the attendees about various complexities involved in developing deep learning applications. He shed light on the unattractive and often overlooked aspects of research. He explained the trade offs between effort and accuracy through concepts of diminishing returns in the case of increased inputs.
When your model is as good as your data then the role of data management becomes crucial. Organisations are often in the pursuit of achieving better results with less data. Practices such as mixing and matching data sets with detailed control and creating optimised synthetic data come in handy.
Underlining the importance of data and experiment management, Dan advised the attendees to track the various versions of data and treat it as a hyperparameter. Dan also highlighted the various risk factors involved in improper data management. He took the example of developing a deep learning solution for diagnosis of diabetic retinopathy. He followed this up with an overview of the benefits of resource management.
Unstructured data management is only a part of the solution. There are other challenges, which Allegro AI claims to solve. In this talk Dan introduced the audience to their customised solutions.
Towards the end of the talk, Dan gave a glimpse about the various tools integrated with allegro.ai’s services. Allegro AI’s products are market proven and have partnered with leading global brands, such as Intel, NVIDIA, NetApp, IBM and Microsoft. Allegro AI is backed by world-class firms including household name strategic investors: Samsung, Bosch and Hyundai.
What Allegro AI Has To Offer
Allegro AI helps companies develop, deploy and manage machine & deep learning solutions. The company’s products are based on the Allegro Trains open source ML & DL experiment manager and ML-Ops package. Here are a few features:
Unstructured Data Management
- Allegro AI offers comprehensive probing tools, enabling users to explore datasets in ever finer detail, allowing you to zero-in on points of interest.
- Users can leverage and fuse a multitude of data sources from different sensors – from images and videos to LiDAR, audio, IoT sensors.
Resource Management & ML-Ops
- Users can automate re-use of resources in the cloud or on-prem with a simple API to manage your cluster. They can also easily integrate Allegro AI with leading cluster and orchestration tools: Kubernetes, Slurm, AirFlow, KubeFlow.
- And, also scale their hyperparameter optimization efforts
Know more here.
Stay tuned to AIM for more updates on CVDC2020.