How can neural networks be used for machine learning?

With its ability to form multiple layers, each having independent weights, which feed forward (or reverse in recurrent ones) to the next layer, Neural Networks (NNs) give the ability to capture diverse pattern that will never be possible by standard regression and classification techniques. By adding layers and/or more nodes/neurons in the hidden layers one can tune hyper-parameters to get the desired accuracy required in the data set. However, as with any machine learning approach there are downsides including being prone to biases on the training data, while also being quite opaque on how a model is derived. While logistic regression model can serve ?the reference model, in most cases random forest will give better results than NNs while being very transparent with the derivation of the classification. In fact, even ?hyperplane based approaches like SVM should be tried before taking on NN. However, when analyzing/classifying images, NN in the Convolutional form are the way to go and every major approach in image recognition is a derivation LeCun’s original CNN. Hence, for image classification NNs are the goto solution while for other problems it remains as one of the options.

Neural networks can be used for machine learning in a number of ways. They can be used to improve the performance of a machine learning algorithm by increasing the accuracy of predictions, reducing the amount of data required for training, or increasing the speed of training. Additionally, neural networks can be used to create new features that can be used by a machine learning algorithm to improve its performance.

For neural networks to become more effective, I would draw on the recent comments from Andrew Ng that "data is food for AI." Ultimately, the effectiveness of neural networks, like any tool, depends upon how it is used - three specific ways to improve: 1. Understand the data - is the training data fully representative of "data in the wild" with the capacity to "generalize" in a predictable way? If there are meaningful features in "data in the wild" not reflected in the training data, then regardless of the model architecture, there will be instances where the inferences surprise us (in a bad way). 2. No model is perfect or "zero shot" for every use case - what is the implicit risk rating and "warranty scope" for inference outputs - based on the training data, what are the cases where there is high confidence of the right output, and where are the cases with higher probability of a poor outcome in inference. In other words, where is the model safe to use? 3. Model / Data / Problem fit - usually the simplest model that can capture the key features is the best one - (great data + simple model) usually outperforms (poor data + amazing model) ...

In my humble opinion, the more important question is "In what ways can Information Theory be used to improve both neural networks and machine learning?" Generally formulated (i.e., beyond Shannon's form), Information Theory provides both the proper method for making decisions in an optimal quantitative manner as well as the proper approach for measuring the performance of any chosen method (which is WHY one can claim that a "best" method exists). However, these "theoretical" results are necessarily built on statistical characterizations, while the majority (and the adjective vast is perhaps not overly strong) of neural network and machine learning descriptions never even broach the concept of probability representations - something is clearly amiss here. The principal downside of any information theoretic approach is that it fundamentally depends upon the specification of such densities for the random entities in the problem, and how to make such choices has not been well studied. However, an emergent (and presumably innovative) insight is that the information theoretic rules for quantitative decision making may be turned back upon themselves to decide which densities are better/worse choices for a given set of observations.

Neural Networks (and any variations of that, e.g., Deep CNN or ResNet) have a lot of theoretical and practical limitations. Our tech/algorithms (at ZAC) (Cognitive Explainable-AI (CXAI)) beat the Neural Networks, in many different aspects. (See our website.)