Azure Cognitive Services, which Microsoft calls the “full family of artificial intelligence and cognitive API services”, is getting new features. The first method is called spatial analysis and completes the Computer Vision subset of this catalog. Spatial analysis actually includes several computer vision models that can be used to analyze video streams from surveillance cameras or CCTV systems (the input stream must provide at least 15 frames per second at 1080p) to count the number of images people in a room and measure distances between people.
Continuation of the article below
It is also possible to demarcate interaction areas and trigger events. The use cases are numerous: respect for social distance, managing a customer line in a store, optimizing space management.
Microsoft rhymes computer vision with caution
Since the product is only shown in the public preview (and Microsoft, in turn, only selects long-term or motivated customers), the cloud giant accompanies the service with a series of warnings. He would like to remind that the service provides probability analysis and that it is not recommended to measure a worker’s performance or be used in making medical decisions. With regard to the protection of privacy, Microsoft advises against the use in protected or public spaces and draws attention to the provisions in this area.
The main focus of the documentation is to highlight the technical limitations inherent in the training sets and the performance of the service. Spatial analysis was not designed to be used in real time, but for people between the ages of 18 and 65. Therefore, the error rates in pictures of people who do not enter are higher in this age group. Finally, public lighting or the weather can have a huge impact on performance: improper outdoor use is subtly discouraged.
Technically, Spatial Analysis is installed in an on-premise container that resides in an Azure IoT Edge runtime that is embedded in an Azure Stack Edge appliance that captures the video streams and runs the facial recognition, zone delimitation, or distance compliance models . close. This set communicates with Azure IoT Hub, which hooks up mobile applications and starts FaaS Azure functions to analyze the results with Power BI or logical flows and to retrieve this information in Office 365. This edge cloud connection mainly explains why Microsoft does not offer real-time processing.
Standard and “neural” NLP
In terms of natural language processing, the cloud giant is expanding the capabilities of the in-speech service, which is dedicated to translating speech into text and text into speech and speech. When the Redmond company communicates less through its Cortana virtual assistant, it enriches its solutions, which it divides into categories.
“With the Neural Text to Speech feature, 18 new languages and 32 new voices have been added for a total of 49 languages and 68 voices. We now offer 151 voices in the Standard and Neural Text to Speech versions. In addition, Speech to Text added 11 new local EU languages, which now cover the 24 official languages set by the European Parliament, ”a Microsoft spokesperson wrote in a press release.
Metrics Advisor, the Azure counterpart to Dynatrace AI
More useful to IT managers, Metrics Advisor, also in preview, is an analytics service based on machine learning to monitor metrics and make it easier to diagnose Azure, Office 365 and Bing services.
Metrics Advisor is a managed version of Azure Anomaly Detector, a REST API that connects to time series data from dedicated databases (InfluxDB, ElasticSearch), but also more general databases or data sources such as MongoDB, MySQL, PostgreSQL, Azure SQL, CosmosDB or Azure Blob Storage. Metrics Advisor offers a multidimensional vision, meaning that the service can analyze the data coming from several of these databases at the same time.
Only one parameter is required: one or more columns must contain numeric values. Microsoft also recommends introducing a column for the time stamp (DateTime or String) and considering the dimensionality (number of columns and different values) for a large volume of data.
While it is possible to analyze the data in real time, the documentation warns of the possibility of obtaining partial data, which limits the capabilities of the tool. Rather, it is recommended that you collect the data once a day. With this established connection method, the SRE or its team selects the metrics and properties of the metrics on the basis of which anomalies are to be detected.
Metrics Advisor automatically selects a recognition algorithm developed in Python from its library (Fourier transform, Z-score recognition, dynamic threshold values, Spectral Residual CNN – a specially developed deep learning algorithm – seasonal analysis through local and extreme regression studentized deviation). When either of these models encounters an error, it sends an alert to Azure DevOps webhooks or mail hooks. The connection to teams has not yet been established.
It is possible to adjust the alarm threshold and then benefit from a root cause analysis and diagnostic trees. The tool offers a dashboard and graphics for this purpose. In the latter case, there is mainly an Incident tab that lists all the events and a diagnosis that can be used to analyze their causes. The Incident tab also provides automated tips for identifying likely causes. Finally, a metric chart can be configured to show the relationships between them.
Contrary to what Dynatrace offers with the Davis AI engine, it is currently not possible to get a full tree structure automatically. Metrics Advisor starts the incident node and shows its dependencies. You can use the metric chart to manually complete relationships between data. The difference is that there is no agent to be installed to return the key data.
Surprisingly, Microsoft Azure already offers a monitoring service with the humble title Azure Monitor. Why not embed Metrics Advisor in the latter? We are awaiting a response from Microsoft on this matter and will update this article accordingly.