| By Stefan Enev, HeleCloud Managing Consultant – Software Development |
Dr. Swami Sivasubramanian, VP Amazon Machine Learning delivered the Machine Learning (ML) keynote at AWS re:Invent 2020.
My key takeaway?
ML solutions just got easier, quicker, more accessible and a lot more cost effective. Here’s how.
More computing power for the most advanced users
EC2 instances upgrades and Custom Chip announcements are great news or those AWS customers not using the SageMaker stack and opting in for their own ML (and using AWS for just the VM infrastructure).
Building on last year’s announcement of the Custom Chip Inferentia, designed and created to improve model inference speeds, this year AWS announce model training improvements with EC2 instances powered by Gaudi accelerators. Forecasted for general availability in the first half of 2021, the new Gaudi EC2 instances claim to achieve up to 40% better price performance than current GPU-based alternatives for training deep-learning models.
It is clear Intel (owner of Habana, the company producing the Gaudi accelerators) is competing with NVidia (the most prominent GPU manufacturer) here which presents good news for all consumers. Training expensive models will now cost less. Artificial Intelligence (AI) is centered on vector and matrix data, therefore having a more cost-effective alternative to the GPU is fantastic news.
In addition, the second half of 2021 will see another Custom Chip launched that it purposefully built for ML. Whilst precise numbers were not confirmed, you can expect this to be even more cost-efficient than the Gaudi instances. It should be easy to switch from one to the other, as both will support all the major frameworks – TensorFlow, PyTorch and MXnet.
More options for developers and data scientists
If you don’t want to go ‘hard-core’ or want to spend a small fortune to build a single small ML-based feature to compliment your business, have no fear as the upgrades for SageMaker are here, including almost everything needed to support an ML solution.
- Data Wrangler, a tool to help you prepare data for ML faster
- Feature Store, to reuse model features across projects
- Clarify, to detect potential data bias across ML workflows
- Deep Profiling for SageMaker Debugger for NN training, to highlight bottlenecks and help maximize resource utilisation
- SageMaker pipelines, CI/CD workflows for ML
- SageMaker Neo Edge, to deploy and manage models on edge devices
The largest expense when problem solving in ML is obtaining quality data and preparing it in a format that’s suitable for training. Data Wrangler provides a User Interface to make this task easier. Clarify helps discover biases in the data, a process thus far would have been performed by data scientists manually by indexing and querying the data in order to understand how it’s distributed.
Business solutions for retailers
The AWS Forecast managed service upgrade is announced. AWS Forecast is a tool that takes your historical sales, traffic, inventory, cashflow data and uses ML to accurately forecast future business outcomes such as product demand, resource needs, or financial performance.
On reflection, I’ve seen this completed on spreadsheets in small-medium sized businesses, so it will be interesting to see how those internal forecasting processes compare to AWS Forecasts’ AI. Typically, smaller businesses don’t have enough data points meaning their performance is susceptible to all sorts of internal and/or external influences. As ML requires large amounts of data to train itself, I believe AWS Forecast will make the biggest impact for larger businesses. Typically, such organisations would involve mathematicians to build rule-based models or statistical models that can predict when inventory needs to be replenished, forecast quantity of sales to expect, or to calculate risk. AI replaces all this effort, time and expertise.
With these announcements AWS made ML/AI much less niche and I would expect more and more businesses to implement ML driven features in their core businesses.
The tool availability allows companies with smaller teams to focus on the actual problems they’re trying to solve without having to build their own vast ML infrastructure.
With advances in computing options the model development and training is also much faster, so the whole cycle of testing a hypothesis has been brought down from months to weeks and from weeks to days.
As an AWS Premier Consulting Partner, HeleCloud help customers unlock the full benefits of Cloud technology, from enablement to modernisation and security through to operational governance we have every step of the cloud journey covered.
If you want to discuss this and other announcements from this week’s AWS re:Invent join HeleClouds panel of experts for our live event tomorrow.