Deep Learning Basics

While reading Hung Lee’s Recruiting Brainfood, I stumbled upon this deep learning primer:

The Simple Guide to Deep Learning

The primer is great, and a quick read. Here is my quick summary below:

The basics of deep learning is to think about how the brain breaks up a specific task. For example, let’s say you are hiking the Appalachian Trail, and you see something in the distance running towards you. First, you might notice it is moving. Then, you might notice what shape it is. Then, you might notice how fast it is going. Then, you might notice a big snout. Then, your brain will determine that this is an animal.

The process would continue until your brain evaluated, classified and predicts what object it is seeing. The joy of the mental exercise (for me) is to understand how the human mind works to break down ideas.

Inputs > Algorithm > Prediction > Training: 

The following are the key concepts for thinking about deep learning concepts. Yes, this is overly simplified, but it is still a helpful start. 

  • Inputs: Labels/Images
  • Algorithm: 
    • Levels of Abstraction 1: Is this a shape?
    • Level of Abstraction 2: Is this shape an ear?
    • Level of Abstraction 3: Is this a cat?
  • Prediction = Yes or No. Is this prediction correct?

Current-State of Deep Learning:

  • Supervised Deep Learning: In effect, this is attempting to clone human behavior via labeled images, video, text or speech. 
  • Reinforcement Learning: This is where the model attempts to “learn” behaviors, codify those behaviors (i.e. what does that mean), and then implement strategies to optimize based on those strategies. As the article suggests, the following are some examples:
    • E-Commerce: model learns customer behaviors and tailors service to suit customer interests. 
    • Finance: model learns market behavior and generates trading strategies. 
    • Robots: model learns how physical world behaves (through video) and then navigates that world.

Network Architecture to Detect Objects in Images:

  • Input: Image
  • Extract Feature: Extract the specific features
  • Classification: Classify based on the probability of those features
  • Output: Image prediction

Enjoy your deep learning explorations!

AI & Machine Learning and Knowledge Work

AI and Machine Learning will slowly arrive make its way into all of our apps in both seen and unseen ways.

How do we embrace this? How do we plan for this? How will this change the work all of us do?

Here’s Microsoft’s “Design Ideas” in PowerPoint:

The Importance of Guiding Principles (or, Key Decision Making Criteria)

At the heart of getting something done is getting everyone on the same page to move the project forward. This is especially relevant in an environment of a variety of individuals from department teams, and, ultimately, different walks life.

The diversity of views creates a challenge: Do we argue about the rightness or wrongness of ideas or a decision? Or, do we agree on a set of “guiding principles” and make decisions?

What is a guiding principle?

A guiding principle is a statement that summarizes a criteria or value-based mechanism. Let’s take the following situation in pricing strategy:

  1. Situation: Company X has developed a patented Water Retention System that helps trees grow faster, while using 80% less water.
  2. Complication: The founder and owner of Company X wants to target low-income farmers that cannot afford an expensive system. The venture capitalists wants the founder to charge higher rates to ensure maximum distribution, and ultimately a strong return on their investment.
  3. Question: How should Company X price the Water Retention System?

If you were in this situation, how would you facilitate a decision? Clearly, both the owner and the venture capitalists have a strong case to make regarding the rightness of their decision.

Option A: Conduct a pricing analysis, and present different prices and see if there is a price that meets both needs.

Option B: Develop a core set of guiding principles around making key business decisions, and then conduct a pricing analysis, and evaluate the options based on the guiding principles.

In Option A, there is an implicit debate about what the Owner and the Venture Capitalists value. In Option B, there is an explicit debate about what the Owner and the Venture Capitalists value.

Why is this important?

The point of this example is codifying the unsaid in guiding principles, each individual can evaluate what they value and see whether it resonates.

If there is resonance, then decision making and team dynamics can be more fluid (or, at the leaser — easier).

If there is not resonance, then decision making will be stalled and inauthentic — team members may grudgingly go along, but there will continue to be dissension as increasingly complex decisions are made, and the team will need to decide whether to continue together or not.

Originally posted on Karma Advisory’s medium page here.

What is the role of AI in radiology?

In the article, “New AI Can Diagnose Pneumonia Better Than Doctors (https://www.fastcodesign.com/90152230/new-ai-can-diagnose-pneumonia-better-than-doctors) we begin to see a glimpse of the possibilities:

“In the case of CheXnet, the research team led by Stanford adjunct professor Andrew Ng, started by training the neural network with 112,120 chest X-ray images that were previously manually labeled with up to 14 different diseases. One of them was pneumonia. After training it for a month, the software beat previous computer-based methods to detect this type of infection. The Stanford Machine Learning Group team pitted its software against four Stanford radiologists, giving each of them 420 X-ray images. This graphic shows how the radiologists–represented by the orange Xs–did compared to the program–represented by the blue curve.”

[Article: FastCoDesign.com]

[Image: Stanford Machine Learning Group]

5G is not the “Silver Bullet” for the Digital Divide

From the beginning of the discussion around 5G, I have not been sold on the prospect of 5G networks being the panacea to solve connectivity challenges in the United States (and, in other parts of the world).

In 2007, I remember my enthusiasm for LTE, and the prospect of bridging the digital divide using this technology.  Now, 10 years later the same enthusiasm is there for 5G.

There are two reasons I don’t trust this trend:

  1. If it sounds too good to be true, it probably is. The on-the-ground realities always difer signficantly from the labs. The telecom industry (and, especially players marketing the potential of 5G) have an incentive for the markets to believe this idea.
  2. The middle mile and the last mile still require significant investment. The middle mile will still need fiber, and the last mile is still complicated — i.e. line-of-sight and obstructions will still be there.

This article in FierceWireless “Editor’s Corner—Fixed 5G was tested by the cable industry, and it came up a bit short” describes some key areas to consider:

I’m going to dig into some of the more interesting findings, including a cost comparison between fiber deployments and fixed 5G deployments, a little later, but first let’s cut to the chase: “We have come a long way in the drive to 5G—but as the saying goes—there is still a long way to go,” concludes the report. “As cable operators move Fiber Deeper going to an all passive coax network, the ability to deliver multiple Gbps of capacity to a single home, seems an easier path than building out a FWA [fixed wireless access] millimeter wave architecture.”

Basically, the report concludes that fixed 5G can deliver pretty fast speeds, but that it’s significantly hampered by interference issues, coverage challenges and backhaul and deployment obstacles. It predicts that fixed 5G services might initially be used to deliver services into apartments and other so-called MDUs (multi-dwelling units), and that cable operators might consider using it to reach specific locations more quickly while they build out fiber connections. But Arris and CableLabs definitely don’t present fixed 5G as the panacea that some in the wireless industry have—and they’re not recommending that cable operators immediately switch over to 5G.

So, what do we do?

We do the hardwork and make the long-term descisions to put fiber in the ground. The last mile will be a hybrid solution of fiber, cable, and wireless for years to come. Let’s embrace it fully.

Machine Learning vs. IoT Sensors for Finding Parking Spots

Some of my favorite work with cities and companies is around the Internet of Things. There are many modern day challenges that are best solved with technology. However, the question that we face is how best to solve those challenges. 

For example, some cities, are exploring adding sensors to parking spaces to help inform motorists about where parking spots are available. This type of endeavor will be expensive in so many different ways. Any city looking to do this, will likely need to architect some type of system that looks like the following: 

On the other hand, Google, is leveraging its maps data too help predict where you are most likely to find parking. This is an impressive use of machine learning instead of on-the-ground sensors for smart parking.

Using Machine Learning to Predict Parking Difficulty