Lets decode “AI in Networking”!

Before I say anything about “AI in Networking” (“AI for Networking” is something else), lets ask basic question, What is AI, by the way?

So, Consider DATA = TEXT (LANGUAGE)/ IMAGE/ AUDIO(LANGUAGE)/VIDEO(LANGUAGE)/SENSOR SUPPLY STREAM etc. 

AI software is a computer algorithm/application that consumes some $DATA. It converts that $DATA into some mathematical formula (vector representation of the data) and when something related to that data questioned/thought off, it correlates and answers. When an AI algorithm learns new set-off $DATA, it re-calculate the mathematical formula, re-compute the data and corrects/updates self-learned/built knowledge. The algorithm which has the ability to interpret/understand text language (i.e English) is called Large Language Models (LLM).Note that, LLM has just learned the language, it is not yet intuitive or cognitive. For example, it generates ansible playbooks code but it does not know whether it is right or wrong.

Now every use case has a different usage of AI software like the doctor’s need from AI tool is different than advocates. So, we need to model (Feed the specific domain data so an AI algorithm can build/learn meaning of that) AI algorithm accordingly, so it helps doctors the way doctors need it and to the advocates the way advocates need it. And the way network engineer needs it!!

Ok, So what if i write some Smart program who does the same like give suggestions to doctors based on symptoms or suggest medicines etc? Can i call it AI program? No, There is a difference. See below.

Traditional software program	AI software program
fixed logic to process the data i.e if wrong or out of context data is feed, it will throw the error	Re-Learns every time given data i.e If wrong data feed, it may give wrong answer
No need to model it	Need to model with pre-data
Deterministic nature	probabilistic nature

AI’s end goal is a human level of cognitive and intuitiveness. It has many phases to pass through before it reaches there, Generative AI (generates the content), Predictive AI (predict based on past data pattern) etc are its evaluation phases. I hope, basics of AI are clear now.

Now, Before going to “AI in Networking”, lets see what is “Not AI in Networking”! I see many companies/vendors are re-branding some old stuff again as “AI in Networking”. So, it is necessary to understand the difference between the two.

Traditional NMS (Network Monitoring Systems)/Observibility tools/Zero touch provisions/Smart Automation/Alerting system etc. are not an “AI in Networking” ! (or call it traditional software program) So, 

• Smart config management/Intent-based networking is not an AI
• Managing upgrades/patches is not an AI
• Graphical presentation, topology discovery is not an AI
• HW usages/matrices, raise alarms are not an AI
• Telling the Impacted applications due to network failure is not an AI
• etc…

All these are Smart/Intelligent Automation But an AI! So, What is “AI in networking”, Let me tell that with the example. 

If my AI tool can differentiate BGP related config (which it learned via modeling process) from the given config, it is an AI. If my AI tool can understand the traffic pattern and identify Dos, configure ACLs accordingly, it is an AI. If my AI tool can answer me if upgrading router now is a good time, it is an AI. If my AI tool can help me in finding potential failed component, it is an AI.If my AI tool can detect congestion and configure queue/qos based on past learning or data priority, it is an AI.

Alright, now we know what is “AI in Networking” then who are the vendors today shipping it? There are many and I read through few, like, Juniper Mist AI platform, Cisco’s DNA center, VMware Nyansa Voyage, Fortinet FortiGuard Labs, IBM Security QRadar Suite, Aruba AI Networking. From the reading these solutions’s description they can figure out traffic pattern, false positives, SLAs verification etc but primarily are examples of Smart/Intelligent automation.

Remember, Networking is always a rule based actions technology and needs 100% accuracy. AI may have limited predictability and accuracy i.e Differentiate Dos attack vs Legitimate high volume of traffic, Which makes AI prediction little unfit (at least for near future) for Networking. However it can be a great support to a admin(human) in assisting a large and scaled network environments as virtual assistant or aid. But I don’t see “AI in Networking” doing so called “self-learning” or “self-healing” networks soon. I did not find any open source project too working in this direction. I don’t see any models available to be trained with Networking data like Log, statistics, CLIs/APIs etc. We need these networking semantics understanding algorithm in order to process them and make some thing out of those input/feed data.

What are the challenges of “AI in Networking” have today? We are still in early stage of “AI in Networking” so they are few obvious challenges like vendor locking, skills set (To develop/train the networking models), cyber security, cost, hosting infrastructure (I dont think cloud style hosting will be helpful for networking where problems/support are needed more in real-time and at other than hosting place) and top of all, Accuracy.

So, Who is target end user of “AI in Networking” then? vendor/operator/admin? Its a Network engineer or person who is in charge of running daily affairs of Networks who can take support of such virtual/generative assistants in order to perform his/her daily routine jobs. Thus, Whatever the routing jobs network engineer performs are the “AI in Networking” use cases as well in the real life.

Alright, in summary, “AI in Networking” today is mostly a Smart automation and going to be Smart Aid tomorrow, so “self-healing network” or similar marketing words would take much more time till accuracy of today’s network is not achieved!!