The Top 10 Misconceptions in Artificial Intelligence and Machine Learning

10 Misconceptions in AI and ML in 2018

8/15/2018

Artificial Intelligence and Machine Learning (AI/ML) are often regarded to be "mythical topics" with data scientists working their magic and half million dollar nVidia rigs making deep neural networks "come to live".

But then there is much sobering news
As outlined in my article on "applied AI/ML" the discipline is far from mature or even well understood by practitioners. This has led to a big divide between the understanding of experts, such as data scientists, and practitioners in DevOps, business, and IT ops groups

DevOps and AI
I have published numerous articles on the topic of how we can enable DevOps teams to incorporate AI/ML capabilities into their code. This is a difficult problem due to the experimental character of AI/ML and the very real release deadlines DevOps groups must adhere to today. Please dig into this article over at DevOpsAgenda.com to read my "Seven steps to move a DevOps team into the ML and AI world."

The EMA Top 3 Enterprise Decision Guide for Artificial Intelligence and Machine Learning
This EMA Top 3 guide helps enterprises understand and plan their AI/ML strategy choices and product selections. The EMA Top 3 for AI/ML aims at demystifying the 10 key misconceptions of AI/ML.
10 Common Misconceptions

I need specialized GPUs as I cannot train neural networks on simple GPUs.
Once I provide the AI/ML algorithm with training data, it will do its magic all by itself.
I can train my AI/ML algorithm by looking over my IT or DevOps staff's shoulder.
AI/ML can radically reduce the number of alerts in DevOps and IT Op.
AI/ML is only for data scientists. Business users and developers need the help of these data scientists for their projects to be successful.
The limit of AI/ML is in the algorithm and infrastructure performance.
Today's AI/ML is lightyears ahead of where we were 10 or even 50 years ago.
Self-driving cars can make truly autonomous decisions, similar to human drivers.
AI/ML is coming closer and closer to how humans think.
The differentiation between AI/ML products lies in their underlying learning algorithms.

8 Comments

Profile

This Blog is all about demystifying artificial intelligence and machine learning (AI/ML) for enterprise use. The EMA team and outside experts will offer pragmatic advice to help you plan, prepare, and execute your AI/ML projects. Without becoming overly technical, this blog will provide perspective and a clear understanding of how ML/AI works and what results we can and cannot expect today.

Lloris M

8/15/2018 08:47:41 am

Interesting point about the GPUs. How do I figure out if, which, and how many GPUs I need for my projects? How do I get to a future proof strategy?

Btw, can I get the EMA Top 3 report? :-)

Torsten Volk (EMA)

8/15/2018 09:22:36 am

Great question. If and when to buy dedicated hardware such as GPUs for AI/ML depends on a few factors: 1) how large is your training set, 2) which algorithms are you using, 3) how quickly do you need to train your neural nets, e.g. for experimentation, staging, or production, 4) are you running automatic hyperparameterization or comparing multiple algorithms to find out which one minimizes the total error. In short, there is no global answer as this needs use-case specific investigation and experimentation.

The same can be said about your question in regard to future proofing your strategy. Today, most people's focus is on the different types of neural networks, especially deep learning and reinforcement learning. Due to the fundamental shortcomings of these approaches I'm hesitant to commit to declaring them the future of AI/ML. However, they are the best thing we have today, so when you make your purchasing decision, just be aware of which AI/ML tech you are betting on.

Matt Bartlett

8/15/2018 11:07:18 am

What's the #1 use case for AI in IT today? Where do you see the biggest future potential?

Torsten Volk

8/16/2018 06:46:18 am

Great question, as I believe most of us think about AI/ML within IT and DevOps in a much too glamorous manner. What I mean by "glamorous" is that we tend to believe that AI/ML models are able to make meaningful decisions based on numerous complex considerations.

In reality, AI/ML is more of a development tool that can be used wherever instruction based development is too complex, labor-intense, or simply unreliable. This means that AI/ML models are mostly used as tools/artifacts to solve programming problems where it is easier to provide examples than creating concise rules.The "original example" is Yann LeCuns first convolutional network from 1993 that recognizes diverse and imperfect handwriting. Now take this example and transfer it to any other domain, such as IT operations monitoring. Monitoring tools do not have a global intelligence that decides that something is a true problem versus just a meaningless anomaly. In reality, developers use AI/ML models to recognize individual decision relevant aspects, such as:
1. Which applications can be affected?
2. What groups of users can be affected?
3. Has a similar trend/anomaly caused impact, such as help desk tickets or application downtime in the past?
4. Does it fit the general criteria of issues that typically cause downtime or performance problems.
5. Is this an issue that has been manually addressed in the past.

You can see that the actual intelligence is still on the side of the programmers creating this "checklist", while the AI/ML uses its ability to ingest and analyst huge amounts of abstract data provides the low level answer to each individual question.

Simon K.

8/21/2018 09:14:47 am

How close are we to having fully self driving cars without a human driver? What is still missing?

9/7/2018 11:29:05 am

The problem is that we are using a large number of micromodels to enable the self driving car. This means, that human engineers need to anticipate, simulate, and test for any kind of situation that could possibly occur. This can only be done through millions of miles of test driving and it is still not fail safe as traffic situations are complex and the so-called AI doesn't understand these situation in a human sense, so that it cannot evaluate its decision options based on ethics or legal liabilities to the driver or owner of the car. This is a gap that cannot be closed by throwing more hardware at the problem, but we need to figure out how to model human-type concept learning with AI.

Finley

8/21/2018 09:17:07 am

Why can't the AI just observe what humans do and imitate that? Doesn't a self-driving car just learn from human drivers?

9/7/2018 11:32:41 am

Yes, the AI learns from the human driver. However, there are many things that you cannot learn just by observing. For example, the AI could not tell if certain bad driving habits lead to accidents, without getting into a large number of accidents first, so that it can identify the relevant input variables.

Demystifying Artificial Intelligence, Machine Learning, Containers and Serverless

10 Misconceptions in AI and ML in 2018

Leave a Reply.

Profile

Archives

Categories

Visit our website: Enterprise Management Associates

Twitter

Telephone

Email