7 Machine Learning Breakthroughs

7 Machine Learning Breakthroughs

Introduction

The Ryskamp Learning Machine represents a quantum leap in the world of machine learning. It breaks from the traditions of the past and uses a completely new paradigm for the core machine learning algorithm. This algorithm focuses on logic over pure mathematical solutions and specific information processing (currently associated with traditional programming) combined with categorization and pattern recognition (currently associated with machine learning) into a single algorithm and engine. Additionally, the RLM saves every decision it ever makes, making debugging simple. Neural network “black box” problems are now a thing of the past.

1. The RLM is a General Purpose System.

The RLM can solve a wide range of problems. Many of these problems are not solvable using traditional machine learning techniques. Some are solvable using tradition learning techniques, however, the RLM is able to solve them in minutes rather than weeks. We recently completed a proof of concept with Walmart related to the best use of space on their store shelves. The outcome demonstrates that in an extremely complex environment the RLM can solve a problem in 3 minutes and 18 seconds that takes TensorFlow hours or even days to solve. This has been so compelling to Walmart that the executives at Walmart have agreed to be a Beta customer for the company that purchases our intellectual property.

2. Categorization, Pattern Recognition and Specific Memory are Native.

Machine learning is known to be good at pattern recognition and categorization; however, it is not good at specific memory problems. We like using our maze demonstration to illustrate this concept. A maze is not a pattern – it requires specific memory to be accessed to remember where it has been. The RLM, having pattern recognition and specific memory, can solve a maze once and then perfectly solve it every time. The RLM engine, with no changes, can go on to find patterns in a complex categorization problem. The RLM does not use a hybrid method to achieve this result. All of this functionality is native to the RLM system. Its native design allows for both problems to be solved, as well as any problem, which requires a combination of patterns and specific edge case handlings. In a medical example, the difference can be illustrated in diagnosing the difference between stomach problems in general and appendicitis in particular. Combining categorization with specific knowledge comes naturally to humans, while machine learning systems have struggled with this problem for years.

3. Substantially increased accuracy.

The RLM’s native learning abilities are not only faster, they are also substantially more accurate. Our testing has revealed that we achieve consistently better scores than TensorFlow and other machine learning systems. These test results and proof for all our other claims is available in the form of open source sample applications on our GitHub page.

4. Exponentially Faster Learning Times.

The RLM uses a completely new paradigm for machine learning that breaks from the limitations of traditional machine learning algorithms. This paradigm is exponentially faster. There are many reasons for this; however, the core reason is that this paradigm requires calculation of only one neuron per cycle, whereas tradition machine learning engines require calculations for each neuron in the network for every cycle. These neurons often number in the millions or more.

5. Uses a Fraction of CPU/GPU Hardware.

The test mentioned with Walmart in #4 was conducted on a 40 core machine. The RLM is so scalable that it gets very similar times on a 4 core laptop. TensorFlow, on the other hand, requires as many CPUs and GPUs as you can give it. In addition to being more scalable, the RLM can simply solve much more complex problems in much less time when given tremendous hardware resources.

6. Natively Tracks Every Event for Easy Diagnostics.

The RLM is not subject to the “AI Black Box” problem that plagues traditional machine learning. In fact, the opposite is true for the RLM. In our open source library, you will find example applications that allow you to point and click your way through how the machine learning engine makes decisions. For example, in the Walmart proof-of-concept, users can click on an item and learn why it was selected to be on the shelf and see a history of learning over time. This is described in detail in the following paper: https://github.com/useaible/RyskampLearningMachine/blob/master/Documentation/RLM%20and%20Explainability.pdf

7. Configuration is simple.

According to Wired “…Microsoft Research vice president Peter Lee said that the cost of a top AI researcher had eclipsed the cost of a top quarterback prospect in the National Football League.” The RLM reverses this trend. Although data scientists created the engine, the engineers that implement it each day have a simple Computer Science degree. We encourage you to see the simplicity at useAIble.com/code.

  • 7 Machine Learning Breakthroughs

    Introduction

    The Ryskamp Learning Machine represents a quantum leap in the world of machine learning. It breaks from the traditions of the past and uses a completely new paradigm for the core machine learning algorithm. This algorithm focuses on logic over pure mathematical solutions and specific information processing (currently associated with traditional programming) combined with categorization and pattern recognition (currently associated with machine learning) into a single algorithm and engine. Additionally, the RLM saves every decision it ever makes, making debugging simple. Neural network “black box” problems are now a thing of the past.

  • 1. Exponentially faster calculation.

    The Ryskamp Learning Machine uses a completely new algorithm for machine learning that breaks from the limitations of traditional machine learning algorithms. Hard to believe? Yes it is, which is why we invite you to download the source and try it yourself. Or you can test our speed against other engines on some of the applications on our challenge site.

  • 2. Categorization, pattern recognition and specific memory in the same algorithm.

    We invite you to try out our maze applications in the code on GitHub or on our challenge site. Notice how the Ryskamp Learning Machine only needs to complete the maze a single time to master it? Now compare the same algorithm running the lunar lander application. The maze demonstrates the use of specific memory (that is, remembering something after seeing it only once) while the lunar lander demonstrates that this can be combined with broader categorizations (i.e. using approximation for speeds and altitudes that vary much more than locations on a maze) in the same algorithm. These traits are critical in real-world environments where some specific situations require specific actions and other things can be dealt with in broader categories.

  • 3. Substantially increased accuracy

    In every challenge we have run, our engine has converged on a more accurate result in less time on every challenge against every other engine we tried. This quantum leap is hard to believe. So we open-sourced the code for evaluation and other non-commercial purposes. We invite you to try it on your own problem or challenge.

  • 4. Solves a wider variety of problem sets than traditional machine learning.

    The Ryskamp Learning Machine uses a different core algorithm. This algorithm’s inherent differences open it up to solving more types of problems. Traditional machine learning algorithms are good at pattern recognition, classification, and other things. Most engineers will use a combination of non-machine learning solutions enhanced by machine learning to solve complex problems that require solutions outside of the traditional “machine learning box”. The Ryskamp Learning Machine does not fit in this box. You can solve many types of problems with nothing more than the Ryskamp Learning Machine using its default settings. We encourage you to try it or contact us for more detailed information.

  • 5. Uses a fraction of CPU/GPU hardware.

    Unlike many traditional machine learning methods, the Ryskamp Learning Machine does not use heavy amounts of mathematics or statistics. Depending on the problem being solved, our use of the CPU or GPU is often less than 10% of the usage required by traditional machine learning to converge upon the same solution. There are two core reasons for this difference. First, we use more logic and less math in our core algorithm. This means less calculations are required. Second, our design is architected around today’s hardware. It uses a much more balanced approach to hardware allocation. This is simply a ramification of the algorithm being designed in a different age of hardware. Compare today’s hardware to the hardware available at the times when most advances of the neural network occurred. These networks still use many core concepts dating back to the Perceptron of the 1950’s.

  • 6. Natively tracks every event for easy diagnostics.

    The Ryskamp Learning Machine natively tracks every single decision it ever makes. This is part of what enables the specific memory discussed above. This is also a great diagnostic tool. Although traditional machine learning can be supplemented with logs or other tools, the Ryskamp Learning Machine natively tracks every action it ever takes. This allows for a completely new level of diagnostics. The “black box” of some traditional methods that cannot really explain “why” they made a decision are a thing of the past. With the Ryskamp Learning Machine simply use the SessionCaseHistory API to access every decision the machine has ever made. Note this could be limited if you choose to store less information for disk space conservation.

  • 7. Configuration is simple.

    In a traditional neural network, you must often configure many settings in order to match the problem you are solving. Settings like algorithms, number of neurons, layers of neurons, activation functions, inputs, outputs, and many other settings must be carefully considered to match the design of the network to the problem being solved. The Ryskamp Learning Machine is simple. There is only one algorithm and there are no activation functions. As with any network, users are required to set up inputs and outputs, and define the number of sessions to be run. A few other optional parameters are available but not required. We invite you to view a sample application in our source code to see how easy this process really is.

Business Technical

Contact Us

Machine Learning Challenges