Practical Broadcast Storage - Part 2

Broadcast systems are renowned for their high speed and high capacity data demands. Up to recently, they relied on bespoke hardware solutions to deliver the infrastructure required for live real-time uncompressed video. But new advances in IT data storage have now opened the doors for broadcasters to take advantage of this state-of-the-art IT innovation.

High-speed data and capacity requirements are now no longer exclusive to broadcasting, the proliferation of web sites and phone-apps has driven IT speed, capacity, reliability, and resilience way beyond the levels demanded by broadcasters. Big-data processing has further driven innovation and data storage systems capable of processing and streaming real-time video and audio is easily available using high-end IT solutions.

SDI broadcast infrastructures are highly tuned and optimized facilities that are difficult and expensive to maintain. Years of bolted-on tweaks and changes to facilitate new workflows add to the risk of outage and loss of transmission. Engineers are constantly chasing their tails trying to keep up with the changes to the system while maintaining high levels of reliability.

Web-site and phone-app service providers also encounter similar challenges to broadcasters, especially when trying to optimize data workflows, balance workloads, fine tune systems, and reduce the need for their team to constantly firefight.

IT Innovators Solutions

But IT innovators have been working hard to deliver high levels of intelligent automation and provide highly optimized infrastructures and workflows.

Although the first experimental AI systems go back to the 1940’s, recent advances in virtualized computation, networking, and storage has witnessed an explosion in the number of AI systems now available.

But what makes AI so important? Seemingly, we just analyze data collected from sensors and data-sets to help understand systems and provide a level of prediction. This method of statistical analysis has been used by engineers and scientists for hundreds of years. 

Diagram 1 – Although this statistical chart implies a high correlation between humidity and disc parity error, AI provides much greater levels of verification and prediction by parsing many more data sets

Diagram 1 – Although this statistical chart implies a high correlation between humidity and disc parity error, AI provides much greater levels of verification and prediction by parsing many more data sets

By looking at the number of users logging onto a web-site, the designers can set a pre-programmed service to spin up new virtualized servers to meet the increased demand. If analysis indicates the number of users double for four hours on a Friday at 6:00pm, pre-emptive action can be taken to spin up the servers in anticipation of the increased demand.

Systems are Dynamic

For as long as businesses have been able to harness information, managers have analyzed data to help predict the future using ever increasingly complex statistical methods. But this form of analysis assumes a static or very slow-moving system.

Statisticians and Business Intelligence Analysts spend a great deal of time trying to find correlation between different data sets. In the example above, analysis indicated web-site users doubled at 6:00pm on a Friday. Although this is a simple example, it’s difficult to establish with much certainty whether this is a trend or an anomaly. The data used for analysis is out of date before the information can be adequately scrutinized.

Context Aware

To improve the prediction, more data-sets would be needed so better correlations can be established. The prediction can never be 100% accurate, but accuracy is directly proportional to the number of correlated data-sets that can be established. AI comes into its own when we look at context-aware dynamic and fast-changing systems.

AI is a generic term that covers four different disciplines; reasoning, natural language processing, planning, and machine learning. Split into four more categories, machine learning includes supervised learning, unsupervised learning, reinforcement learning, and deep learning (or neural learning).

Traditional computer programming requires a set of logical rules to be provided for the data to be analyzed. Usually providing a series of decision trees, the programmer will parse the data and then provide binary outcomes. For example, if it’s 6:00pm on a Friday, then spin up two more virtual web-servers.

Programming is Too Slow

Changing the data to variables, such as “6:00pm” and “Friday”, allows generic rules to be established so that the time and day can be referenced from a database and easily changed. However, this method soon becomes complex and difficult to administer when the rules need to be adapted. 

Diagram 2 – AI encapsulates four different disciplines and machine learning is further classified into four more classes

Diagram 2 – AI encapsulates four different disciplines and machine learning is further classified into four more classes

It might become clear that Tuesdays also need an extra server at 6:00pm. This would require the programmer to add an addition to the rule or add a completely new rule. The example is trivial, but complexity of the program soon escalates, and efficient software maintenance is further compromised.

The method is slow and potentially problematic as the rules must be constantly adjusted to improve the efficiency of the system.

AI Generates Rules

Instead of manually programming rules for the data, AI takes the opposite approach and the data is used to automatically generate its own rules within the context of the data it’s working with in real-time. And that is the true power of AI.

Algorithms continuously monitor historic as well as real-time data and find accurate correlation between the data sets to help improve the future predictions for the system. This approach further improves efficiencies to take advantage of dynamic systems to constantly balance resource.

In the previous article in this series we looked at how advanced IT storage systems balance the user experience against the different storage resource available. An edit system would require instant low latency access to their edit masters and rushes, and this would be provided by the Solid-State-Device (SSD) and memory storage during the edit session.

Improving Efficiency

It would be uneconomical to continue to keep the media files in the high-speed storage when the edit session had finished so the files would need to be copied back to spinning disks. Although manual rules could be established to copy the files between the two storage mediums, AI provides a far more efficient method of providing this facility.

Algorithms can establish which files to copy based on previous sessions and the editor’s behavior. If the editor was previewing off-line rushes earlier in the day, the AI algorithm would associate these files with them and automatically copy the media files to SSD and memory storage associated with the edit suite for their edit session.

Real-Time Analysis

AI not only provides ever increasing efficient levels of automation, it does it without any human analysis, programming, or intervention. The algorithms learn the behavior of the storage dynamics and accurately predict which files need to be transferred between the different storage medium to improve the user experience and make the most of the resource available.

All this happens in near real-time making the traditional record-analyze-program methods now obsolete. In the next article in this series, we look at how AI has influenced and greatly improved the reliability and resilience of storage. 

Comments:

Mr. Orme tells us that AI encapsulates four different disciplines and machine learning is further classified into four more classes. Most textbooks and papers I’ve read ML concur with your assertion that ML be subdivided into the four categories shown in diagram 2. 

However, this is the first time I’ve seen AI branched directly to NLP. 

Trained with ML and DL algorithms NLP can perform:

Text Classification
Language Modeling
Speech Recognition
Caption Generation
Translation

Other ML/DL applications include

Metadata extraction
image, audio, and video recognition
image, audio, and video search
image, audio, and video classification
image, audio, and video QC and anomaly identification
viewer recommendation engine

We’ve seen promising developments in content-generation and post-production

Can you please elaborate on how reasoning and planning applications fit into M&E?  I would imagine that AI planning apps could serve the business and project development.  But where does reasoning fit into the equation?  Why would there be a direct subset of AI rather than a subset of ML/DL? 

Rick Singer
https://www.linkedin.com/in/ricsinger/

January 31st 2019 @ 21:03 by Rick Singer

Hello Rick,

Thank you for taking the time to read my article and provide such detailed questions.

In my research I have found NLP has two distinct categories, rule-based and statistical. From your question it appears the examples you’ve given refer to statistical NLP, of which the core components are ML and DL.

Misunderstandings in human communication and language are a regular consequence or our interaction. This leads onto the suggestion that language processing isn’t deterministic, that is the same language doesn’t always have the same interpretation.

The following quote demonstrates this, “Always remember that you’re unique. Just like everyone else”. Is this humour? Satire? Or meaningless rubbish? Many people reading the quote will have their own interpretation and may well have another answer to the three examples I have given.

Now contrast this to the Peano axioms of mathematical logic. Under these rules, 1+1 is always 2. This isn’t open to interpretation. It’s a statement of fact (based on the axioms). Similarly, a disc drive either fails or not (based on its parity errors), and the temperature either increases to 28C or it doesn’t. Therefore, a machine can easily judge the outcome of many objective measurements and correlate other deterministic data-sets to achieve greater predictive accuracy.

From my research, the non-determinism of human communication makes it challenging to train statistical NLP models. It’s difficult to measure the truth as it’s based on human subjective judgements, which vary and are dependent on many environmental, social, and economic factors.

And for me, this is where the whole concept of AI becomes fascinating. The accuracy of ML is limited by the data-sets available, so for subjective systems, such as human communication, we must have an external adjudicator deciding the rules. For NLP systems, who decides who the rule-maker should be? Democratically elected representatives who will form committees of committees? Scholars who agree? Or tyrants who benefit from the manipulation of the masses? We soon disappear down a philosophical rabbit hole and this is before we consider how languages throughout the ages change over time as cultures develop.

I completely agree that we regularly use statistical NLP with the methods you describe to provide the results you demonstrate, but I believe this only works within the limits of certain rules. I argue, that due to the interactive nature of rule-based and statistical NLP, the inherent inaccuracy of human subjective decision making, and the inability for humans to agree on the truth, the whole complex subject becomes a subset of AI directly. But I also accept that this may change as AI develops further, as a science.

I should also add that this article was laying the foundation for how AI is used to improve real-time storage and make it more reliable in the context of broadcast television, and it wasn’t my intention to discuss NLP in-depth. The second diagram points to the greater efficiency storage systems deliver when using AI. I agree it needs further clarification and I’ll update the description to make this clearer.

As an aside, I’m currently studying Gödels incompleteness theorems as I’m investigating the limitations of AI (if there are any). If you know anybody who has an interest in this then please would you let me have their details so we can share ideas?

Regards

Tony

February 5th 2019 @ 09:04 by Tony Orme Editor
Let us know what you think…

Log-in or Register for free to post comments…

You might also like...

Server-Based “At Home” Workflows Provide Efficiency For NASCAR Productions

NASCAR Productions, based in Charlotte NC, prides itself on maintaining one of the most technically advanced content creation organizations in the country. It’s responsible for providing content, graphics and other show elements to broadcasters (mainly Fox and NBC), as w…

Essential Guide:  Practical Broadcast Storage

Ground breaking advances in storage technology are paving the way to empower broadcasters to fully utilize IT storage systems. Taking advantage of state-of-the-art machine learning techniques, IT innovators now deliver storage systems that are more resilient, flexible, and reliable than…

Practical Broadcast Storage - Part 3

Artificial Intelligence (AI) has made its mark on IT and is rapidly advancing into mainstream broadcasting. By employing AI methodologies, specifically machine learning, broadcasters can benefit greatly from the advances in IT infrastructure innovation and advanced storage designs.

Practical Broadcast Storage - Part 1

Broadcast and IT technology collaboration is continuing its journey with the cross-over between them becoming ever clearer. Storage is a major growth area with advanced methods achieving greater efficiencies and resilience. And as broadcast repeatedly demands more and more capacity,…

Essential Guide: Intelligent Storage for Broadcast Workflows

​Major advances in seemingly unrelated industries such as Deep Learning and High Performance Computing are empowering broadcasters to deliver greater efficiencies and improved performance. IT has perfected the art of optimizing costs against resource through scalability and flexibility. And broadcasters c…