Week #2: Wrapping up Machine Learning (ML) and NAND Learning, Planning Data Gathering

Hey guys! Welcome to my second blog for my project, I’m so excited to share with you guys everything I got to learn this week. I was mainly focused on finalizing my machine learning and NAND Flash knowledge to ensure that I’m well-prepared for the upcoming weeks. 

Last week, I started learning various concepts that would be relevant while I start my project. I studied NAND Flash: a non-volatile memory that’s used a lot in SSDs, phones, and laptops. I also started my journey learning ML, and tried it out on my own as well! 

This week, I started off by wrapping up NAND Flash learning. I read this updated article called NAND 201 from 2011 that discussed density enhancements made to NAND Flash memory. For instance, NAND Flash used to only be made up of 1 die,  but now they can be made up to 4. Some NAND Flash now hold up to double the number of pages as before. These enhancements help each device have a higher-density of memory without utilizing more resources like silicon. 

In addition to NAND 201, I read another paper on PCIe SSDs. PCIe SSDs are a type of SSD that is a lot faster than typical SSDs. Though it is usually more expensive, PCIe SSDs can handle large amounts of data while having high endurance. They’re suitable for enterprise setups and data centers. You might be confused why I had to read this paper, as it doesn’t really discuss NAND Flash. The CSD5000, which is the Scaleflux SSD product I will be collecting data on, is a PCIe SSD, so it’s necessary to know the structure and uses of these SSDs. 

But the main focus this week was machine learning! There are two types of supervised machine learning: classification and regression. Classification is used when you want to predict a value out of a fixed number of outcomes (like classifying a dog breed), while regression is used when you want to predict a continuous value (like the price of a house). I realized that my machine learning project would use regression, so I focused on learning and practicing regression. By watching Stanford Online lectures on Machine Learning, I dove into linear regression, decision tree regression, and support vector regression. Through these Stanford lectures and last week’s machine learning video, I started seeing more connections between machine learning and math. For example, here’s a really intimidating calculus equation used in linear regression. 

This equation is the gradient (derivative) of the cost function. It describes the change (represented by the partial derivative on both sides) in the deviation from the predicted value (hθ(x)) to the actual value (y). The equation is squared to increase the consequence of having a larger difference between the predicted and actual value. Also, there is a ½ fraction at the beginning of the equation to simplify the result after taking the derivative (because ½ * 2 = 1). Oh, and θ represents the importance of the feature x on affecting the predicted value hθ(x). This equation is really useful for finding the best θ value to minimize the differences between y and hθ(x). That way, the prediction can be as close to the actual value as possible!

Here’s the link to the Stanford lecture I referenced: https://www.youtube.com/watch?v=gqKaVgQxEJ0&list=PLoROMvodv4rNyWOpJg_Yh4NSqI4Z4vOYy&index=2

And that’s all for now! I spent a lot of time this week refining my knowledge in NAND Flash, Python coding, and machine learning through reading papers, watching lectures, and trying it out on my own. Stay tuned for next week, where I’ll finally begin gathering physical data from Scaleflux’s SSD product called CSD5000. I’ll also begin coding my first program of the project next week. I hope you guys had a wonderful week, and I will see you soon! 

Sources: 

1. Cooke, Jim, and Micron Technology. “Nand 201: The Continued Evolution of NAND Flash.” EE Times, EE Times, 13 Feb. 2011, www.eetimes.com/nand-201-the-continued-evolution-of-nand-flash/. 
2. Kim, Eden, et al. “PCIe SSD 101.” SNIA, 2013, www.snia.org/sites/default/files/SNIASSSIPCIe101WhitePaper_120314.pdf.