This thing is like 900 pages long.
You want to take a deep breath and make sure you're committed before you even open it.
I did check the Table of Contents and scrolled quickly through, and I see it's definitely a hands-on applied technology book using the Python programming language.
I will be blogging more about it when I get going.
I haven't been able to understand why the original chip pioneers, like Intel and AMD, have not been able to pivot in order to compete with NVidia (Stock Symbol: NVDA).
I know a few things, like the fact that when gaming became popular, NVidia made the graphics chips that had graphics acceleration and such. Graphics tend to draw polygons, and drawing polygons is geometric and trigonometric - which require floating point arithmetic (non-integer based mathematics). Floating point is difficult for a CPU to do, so much so that classical CPUs either offloaded or employed other tricks to do these kinds of computations.
Now, these graphics chips are the "rave" for AI. And Nvidia stock has gone through the roof while Intel and AMD have been left behind.
So what does an AI chip have, that is different from an older CPU?
CPUs use all three of these in some form or another, but an AI chip has all three of these in a highly optimized and accelerated design. Things like prediction (such as branching prediction), parallelism, etc. They're simply better at running "algorithms".
This link, by the way, from NVidia, discusses the distinction between Training and Inference:
https://blogs.nvidia.com/blog/difference-deep-learning-training-inference-ai/
CPUs, they were so bent on running Microsoft for so long, and emulating continuous revisions of instructions to run Windows (286-->386-->486-->Pentium--> and on and on), that they just never went back and "rearchitected" or came up with new chip architectures. They sat back and collected money, along with Microsoft, to give you incremental versions of the same thing - for YEARS.
When you are doing training for an AI model, and you are running algorithmic loops millions upon millions of times, the efficiency and time start to add up - and make a huge difference in $$$ (MONEY).
So the CPU companies, in order to "catch up", I think, with NVidia, would need to come up with a whole bunch of chip design software. Then there is the software kits necessary to develop to the chips. You also have the foundry (which uses manufacturing equipment, much of it custom per the design), etc. Meanwhile, NVidia has its rocket off the ground, with decreasing G forces (so to speak), which accelerates its orbit. It is easy to see why an increasing gap would occur.
But - when you have everyone (China, Russia, Intel, AMD, ARM, et al) all racing to catch up, they will at some point, catch up. I think. When NVidia slows down. We shall see.
In my last blog post, I posed the question of how many hidden layers should be in a neural network, and how many hidden neurons should be in each hidden layer. This is related to the Neural Network Design, or Neural Network Architecture.
Well, I found the answer, I think, in the book entitled An Introduction to Neural Networks for Java authored by Jeff Heaton. I noticed, incidentally, that Jeff was doing AI and writing about it as early as 2008 - fifteen years ago prior to the current AI firestorm we see today - and possibly before that, using languages like Java, C# (C Sharp), and Encog (which I am unfamiliar with).
In this book, in Table 5.1 (Chapter 5), Jeff states (quoted):
"Problems that require two hidden layers are rarely encountered. However, neural networks with two hidden layers can represent functions with any kind of shape. There is currently no theoretical reason to use neural networks with any more than two hidden layers. In fact, for many practical problems, there is no reason to use any more than one hidden layer. Table 5.1 summarizes the capabilities of neural network architectures with various hidden layers."
Jeff then has the following table...
"There are many rule-of-thumb methods for determining the correct number of neurons to use in the hidden layers, such as the following:
Simple - and useful! Now, this is obviously a general rule of thumb, a starting point.
There is a Goldilocks method for choosing the right sizes of a Neural Network. If the number of neurons is too small, you get higher bias and underfitting. If you choose too many, you get the opposite problem of overfitting - not to mention the issue of wasting precious and expensive computational cycles on floating point processors (GPUs).
In fact, the process of calibrating a Neural Network leads to a concept of Pruning, where you examine which Neurons affect the total output, and prune out those that don't have the measure of contribution that makes a significant difference to the end result.
When training a Neural Net, it is important to have what is referred to as a Key Performance Indicator - a KPI. This is an objective, often numerical, way of "scoring" the aggregate output so that you can actually tell that the model is learning - that it is trained - and that the act of training the model is improving the output. This seems innate almost, but it is important to always step back and keep this in mind.
Chapter 5 discusses the effort that goes into training a Neural Net, but from the perspective of Efficiency. How well, is the Neural Net actually learning as you run through a specified number of Epochs, with whatever batch sizes you choose, etc.?
In this chapter, Michael Nielsen discusses the Vanishing Gradient. He graphs the "speed of learning" on each Hidden Layer, and it is super interesting to notice that these Hidden Layers do not learn at the same rate!
In fact, the Hidden Layer closest to the Output always outperforms the preceding Hidden Layer in terms of speed of learning.
So after reading this, the next questions in my mind - and ones that I don't believe Michael Nielsen addresses head-on in his book, is
I will go back and re-scan, but I don't think there are any Rules of Thumb, or general guidance tossed out in this regard - in either book I have covered thus far. I believe that in the examples chosen in the books, the decisions about how to size (dimension) the Neural Network is more or less arbitrary.
So my next line of inquiry and research will be on the topic of how to "design" a Neural Network, at least from the outset, with respect to the sizing and dimensions. That might well be my next post on this topic.
As if Chapter 2 wasn't heavy enough, I moved onto Chapter 3, which introduced some great concepts, which I will mention here in this blog post. But, I couldn't really follow the detail very well, largely again due to the heavy mathematical expressions and notations used.
But, I will iterate what he covers in this chapter, and I think each one of these will require its own "separate study", preferably in a simpler way and manner.
Chapter 3 discusses more efficient ways to learn.
It starts out discussing the Cost Function.
In previous chapters, Nielsen uses the Quadratic Cost Function. But in Chapter 3, he introduces the Cross-Entropy Cost Function, and discusses how by using this, it avoids learning slowdown. Unfortunately, I can't comment much further on this because frankly, I got completely lost in this discussion.
He spends a GREAT DEAL of text discussing Cross-Entropy, including a discussion about the fact that he uses different Learning Rates on the Quadratic Cost Function (.15), versus the Cross-Entropy Cost Function (.005) - and discusses that the rates being different doesn't matter because it is more about how the speed of learning changes, than the actual speed of learning.
After a while, he mentioned an alternative to Cross-Entropy called Softmax. Now, this term seemed familiar. In doing a backcheck, I found that Softmax was used in the first book I read, called AI Crash Course by Haydelin de Ponteves. I remembered both Softmax and Argmax being mentioned.
Softmax introduces a layer of neurons parallel to the output neurons. So if you had 4 output neurons, you would have 4 Softmax neurons preceding the 4 output neurons. What Softmax does, is return a Probability Distribution. All of the neurons add up to 1, and if one neuron decreases, there must be an alternative increase amongst the other Softmax neurons. This could be useful, for example, in cases where the AI is guessing which animal type it is: Dog, Cat, Parrot, Snake. You might see a higher correlation between Dog and Cat, and a lower one with the Parrot and Snake.
Nielsen then goes on to discuss Overfitting (OverTraining) and Regularization, which is designed to combat Overfitting. He discusses four approaches to Regularization, which I won't echo here, as I clearly will need to consult elsewhere for simpler discussions, definitions and examples of these.
I came across this book entitled Neural Networks and Deep Learning, by Michael Nielsen, and decided to invest some time in reading it. I am hoping to reinforce what I have read thus far (for instance in the book AI Crash Course by Haydelin de Ponteves).
The first chapter in this book covers recognition of hand-written digits. This is a pretty cool use case for AI technology, given that everyone's individual handwriting constitutes a font in and of itself, and even then, there is a certain degree of variation every time an individual hand-writes characters. Who uses technology like this? Well, banks for example might use something like this as part of check processing!
I like the fact that this example is optical in nature - and, it could even be inverted in a Generative AI manner to generate new fonts, or new characters in some context or use case!
One of the first things this book covered was Perceptrons - and explained how Perceptrons are essentially binary (1s and 0s) as result output of action functions. He then moves on to explain that binary results are not always in demand or optimal, and that gradient values (shades of gray) between 0 and 1 are often desired or necessary. And this is why we have Sigmoid Neurons.
This was insightful, because the first book I read (unless I overlooked) never even mentioned the concept of Perceptrons, and jumped right into Sigmoid Neurons.
Nielsen then goes on to explain, briefly, the architecture of Neural Networks and embarks on a practical example using the Handwritten Digits Recognition use case.
One thing that was helpful in this chapter, was the fact that he explains how these hidden layers of a neural network "fit together" by use of a practical example. This helps as far as design and modeling of a Neural Network go.
He goes on to discuss the purpose of using bias along with weights, and then goes into a deeper discussion on gradient descent - and stochastic gradient descent. Stochastic Gradient Descent is critical for reinforcement learning because it minimizes the cost function. Minimizing a cost function does no good though, if you can't "remember" it, so the concept of back propagation is required to instill back into the model, this minimization of cost.
I downloaded the code for his Handwriting Recognition example. I immediately figured out by reading the README.txt file (most people ignore this and waste time learning the harder way) that the code was not written for Python3. Fortunately, someone else had ported the code to Python3 - and this was in another Github repository.
In running the code from the Python3 repository, I immediately ran into an error with a dependent library called Theano. This error was listed here: Fix for No Section Blas error
So I was fortunate. This fix worked for me. And we got a successful run! In this example, the accuracy in recognizing handwritten digits was an astounding 99.1%!
This code uses a data sample from MNIST, comprised of 60K samples. The
code by default breaks this into a 50/10 training and validation sample.
So now that we know that we can run the code successfully, what might we want to do next on this code, in this chapter, before moving on to the next chapter?
Here is a table that shows some results from playing with the epochs and mini batch sizes:
test.py
Run 1: Epochs=30, Mini Batch Size=10, Results: 8700/10,000at its peak in epoch 26
Run 2: Epochs=40, Mini Batch Size=10,Results: 9497/10,000 at its peak in epoch 36
Run 3: Epochs=40, Mini Batch Size=20,Results: 9496/10,000 at its peak in epoch 36
So interestingly, the numbers go up each epoch, until a peak is reached, and then the number settles back down for the last 2-3 epochs in these trial runs. It appears, that the number of epochs is a more important coefficient than the mini batch size.
Another test with adjusted learning rates could be interesting to run, as this impacts Gradient Descent quite a bit. As Gradient Descent is a Calculus concept for finding the minimum, when learning rates are too low or too high, it impacts the time and ability for the "ball to drop to the bottom" of said low point.
He also, at the very end, mentions the use of more/different algorithms, referencing the Support Vector Machine (SVM).
As Part II of my foray into Stock Portfolio Optimization, I decided to implement Python code that I found in a blog: https://www.datacamp.com/tutorial/lstm-python-stock-market
This blog was written by a Data Scientist, Thuresan Ganegedar . And he spent some time on this. I was super excited after I read the blog, to go in and compile the code and run it myself. The first thing I did, was get myself an API Key so that I could fetch data for a stock symbol.
But - I had issue after issue in running this code.
First, the code was sufficiently outdated that many libraries, especially Keras, had "moved on" and completely deprecated certain libraries, calls, etc. I spent considerable time picking through the code, doing web searches, and trying to "just get the code to run".
Second, the code was using hard-coded array values for the number of data points for the stock symbol, and because of this, I ran into null exceptions because the number of data points I had in my stock symbol fetch was over 2K less than the example had. To get around this, I wound up going with 4000 training data points and 600 testing data points. Later, I enhanced the code to dynamically set the training and testing sizes based on how many data points came down in a csv, on a percentage basis.
I also decreased the window size and the batch size as well.
In the end, I was able to produce nice graphs that matched the ones in his blog for the Predictions based on Simple Average and Exponential Moving Average. But these graphs are based on algorithms, not a neural AI network - in other words, they aren't "true AI".
The AI portions of the code did run to completion, but the graphs that came out from implementing the AI, did not look correct. The price predictions were FAR FAR lower than the actuals, as well as the ones from the Simple and Exponential Moving Average graphs.
I may come back and take another look at this, but it was a good exercise, and this exercise made me realize that I need to read and learn more about AI - and what is going on "under the hood".
Okay, last chapter in the book!
In this chapter, you get to "create" (actually, "create" means download and run) some Github hosted code that allows you to train a model to learn how to play the video game "Snake".
Snake is an early video game, probably from the 1970s or 1980s. I don't know the details of it but I am sure there is plenty of history on it. I think you could run it on those Radio Shack Tandem TRS80 computers that had 640K of RAM on them and saved to a magnetic cassette tape (I remember you could play Pong, and I think Snake was one of them also).
The idea was that each time the snake ate an apple (red square) the snake's length would increase (by one square). You could move up, down, left, right constrained by coordinate boundaries, and if the snake overlapped with itself, it died and the game ended.
When I first ran the model training for this, it ran for more than a day - perhaps all weekend, and then died. The command prompt, when I returned to check on progress, had a [ Killed ] message.
I had other models in this book die this way, and decided that I was running out of memory, and my solution to the other models was to edit the source code, and decrease the number of Epochs, and reduce the loop complexity. This made the models a LOT less efficient and reliable, but I still saw beneficial results from running them with this tactic.
In this case, for some reason, I went to Github and looked at the Issues, and I saw a guy complaining about a Memory Leak in the Tensorflow libraries. There was a patch to fix this!
Below is a Unix/Linux "diff" command, which shows this patch:
% diff train.py train.py.memoryleak
5d4
< import tensorflow as tf
12,15d10
< import gc
< import os
< import keras
<
64,67c59
< #qvalues = model.predict(currentState)[0]
< qvalues = model.predict(tf.convert_to_tensor(currentState))[0]
< gc.collect()
< keras.backend.clear_session()
---
> qvalues = model.predict(currentState)[0]
So in summary, the patches are:
Of course some imports are necessary to reference and use these new calls.
This fixes the memory problem. It does not appear that the training will ever end on its own when ou run this code. You have to Ctl-C it to get it to stop, because it just trains and trains, looking for a better score and more apples. I had to learn this the hard way after running train.py for a full weekend.
So this wraps up the book for me. I may do some review on it, and will likely move on to some new code samples and other books.
I just finished Chapter 12 in my AI Crash Course book, from Hadelin de Ponteves.
Chapter 12 is a short chapter actually. It explains, in a refreshing and surprising simple way, the concept of Convolutional Q Learning which pertains to how image recognition/translation is fed into a Deep Q Neural Network (from prior chapters in the book).
The chapter covers four steps:
I probably don't need to go over all of these details in this blog as that would be redundant.
If you have some exposure to Computing and are familiar with Bitmaps, I think this process has shares some conceptual similarity to Bitmaps.
For example, in Step 1 - Convolution - you are essentially sliding a feature detector or "filter" (e.g. 3x3 in the book) over an image - starting on Row 1, and sliding it left to right one one column at a time before dropping down to Row 2 and repeating that process. On each slide interval, you are employing a mapping process of multiplying each square of the image (again using a 3z3 area in the book) to the corresponding value of the map. Each individual iteration of this process creates a single Feature Map.
In sliding this 3x3 map across, you can only go 5 times to the right before you run out of real estate. Then you drop down - and you can only drop down 5 times until you run out of real estate in that direction. So if I am correct in my interpretation about how this works, you would get 5 x 5 = 25 Feature Maps with a 7x7 image and a 3x3 filter.
Pooling is actually similar to the process of filtering to a feature map. The aim of it is to reduce the size and complexity of all of those feature maps. The sliding process is the main difference; instead of going one column / row on each slide, you are sliding (using a 2x2 in the book) over the entire area of the pool size.
Once you get all of the pools, these are flattened out into a single dimensional array, and fed into the Inputs of the standard Q Learning model, with the outputs pertaining to image recognition.
This diagram shows how all of this is done, with a nice comparison between Biological image recognition, with this AI image recognition process.
Image Recognition - Biological vs AI
Source: frontiersin.org
Now in Chapter 12 of the book, the process represents what we see above. There is just a single Convolutional Layer and Pooling Layer before the AI Neural Network (hidden layers) are engaged.
Chapter 12 does not cover the fact that the Convolutional Layer is an iterative process that includes Convolution followed by Sub-Sampling in an iterative fashion.
This diagram below represents this.
In the next chapter, there is a code sample, so I will be able to see whether it includes this sub-sampling process or not.
I now am starting to see why those companies with deep pockets have an unfair advantage in the not-so-level playing field of adopting AI. Resources.
It takes a LOT of energy and computing resources to train these Artificial Intelligence models.
In Chapter 11 of AI Crash Course (by Hadelin de Ponteves), I did the work. I downloaded, inspected, and ran the examples, which are based on Google's Deep Mind project. The idea is to use an AI to control server temperature, and compare this with an "internal" (no AI) temperature manager.
What you would do, is to train the model (first), and it would produce a model.h5 file, that would then be used when you ran the actual model through testing.
The problem, though, is that on my rather powerful Mac Pro laptop, the training would never run. I would return HOURS later, only to see [ killed ] on the command prompt. The OS apparently was running out of resources (memory probably).
So I started tinkering with the code.
First, I reduced the number of epochs (from 25 to 10).
#number_epochs = 25
number_epochs = 10
Which looked like it helped, but ultimately didn't work.
Then, I reduced the number of times the training loops would run. When I looked at the original code, the number of iterations was enormous.
# STARTING THE LOOP OVER ALL THE TIMESTEPS (1 Timestep = 1 Minute) IN ONE EPOCH
while ((not game_over) and timestep <= 5 * 30 * 24 * 60):
This is 216,000 loop iterations in the inner loop, and of course this needs to be considered from the context of the outer loop (25, or, adjusted down to 10 as I did). So 216,000 * 25 = 5 million, 400 thousand. If we reduce to 10 the number of Epochs, we are dealing with 2 million, 600 thousand.
I don't know how much memory (Heap) is used over that many iterations but on a consumer machine, you are probably going to tax it pretty hard (remember it has to run the OS and whatever tasks happen to be running on it).
I was FINALLY able to get this to run by reducing the number of Epochs to 10, and reducing the steps to 5 * 30 * 24 (3600). And even with this drastic reduction, you could see the benefits the AI performed over the non-AI temperature control mechanism.
I read about Q Learning, and was feeling somewhat proud of myself for sticking my toe into the water.
Then I read about Deep Q Learning - in this same book - and it was as if someone took an ice bath and dumped it over my head. I went into the tunnel - advancing through chapters 9,10 and 11, only to come out the other end confused ("what did I just read?").
The coding examples were interesting enough that I kept pushing forward, but a lot of what is in the code is masked by the fact that the math and formulas were hidden away in libraries like Keras. So while I thought the examples were cool and I had a grasp of the problems they were attempting to solve, I still came out at the end with confusion and question marks in my head.
Q Learning vs Deep Q Learning
In Chapter 9, which covers Deep Q-Learning, things start to get very complex very fast. So what is the difference between Q Learning (introduced in Chapter 7-8), and Deep Q Learning?
With regards to the approach to solving problems, the book gets into a good discussion - worth mentioning here - about the difference between ArgMax and SoftMax.
Argmax vs Softmax
In Q Learning, the 'name of the game' was to find (and use) the highest Q Value. This is referred to as "Exploitive" and is known as the ArgMax method of Reinforcement Learning.
In Deep Q Learning, probability distributions across several
variables are being continually updated during the training of the model. You have a set of (input) variables, with specific weights, but as you take random samples and compare the predicated value to the actual value, the weights are updated according to the new realities (results). This process is referred to as Explorative (in nature) and is named the SoftMax method.
Chapter 9 starts you off simple(r). With a Real Estate example of predicting home prices. Seems sensible enough, since we can all think of input variables that help drive the price of a home. The focus here is on trying to show the process, which is broken down into the following steps:
From here, the book advances into Deep Learning Theory. The idea borrows from the human brain, which is connected by Synapses that send signals. This is the fundamental concept behind Deep Q Learning because it starts with a certain number of "Layers". There are a minimum of three layers that are as follows:
The book mentions Activation Functions that take weighted input values, and return an output value. The book mentions three of these, which sound intimidating, but if you are familiar with Electronics and/or Trigonometry, these names actually make some sense:
Now from here (Chapter 9), the book goes into Chapter 10 - Self Driving Car - which an implementation of Chapter 9 - and quite fun to do. Then it dives into Chapter 11, which uses the example taken from Google's DeepMind project that optimizes server temperature with a simulation.
Chapter 11 in particular really drives home the process by showing how you can optimize or minimize costs.
Seeing is believing, and when you see this code run and start to view the results, I have to admit it is pretty darn cool.
It also takes a LONG time to run. I had to shorten the epochs from 100 to 25 to keep the job from getting killed (I am not sure what exactly was killing it). Running for 100 epochs was taking my laptop HOURS to finish (2-3 hours). But at the end of each Epoch, almost always the energy savings from the AI was superior to the energy savings of not using the AI (which in this case is modeled by the server's mainboard temperature controller).
There's so much more to discuss. But I think I have hit the highlights here.
The Crash Course in AI presents a fun project for the purpose of developing a familiarization with the principles of Q Learning.
It presents the situation of Robots in a Maze (Warehouse), with the idea that the AI will learn the optimal path through a maze.
To do this, the following procedure is followed:
Zabbix 7 introduced some better support for SLAs. It also had better support for VMware. VMware, of course now owned by BroadSoft, has prio...
