Скачать с ютуб ZyRA: The Human Construct Neural Network (You've Been Training Models Backwards) в хорошем качестве

ZyRA: The Human Construct Neural Network (You've Been Training Models Backwards)

Link to RBIE Paper:   / urn:li:activity:7305254405266649088   Link to Colab Notebook: https://colab.research.google.com/dri... Here's a description of the video: Background and Motivation: The presenter's experience as an IT generalist and a mentor's observation about the lack of understanding of how different software pieces work together in an organization led to the presenter building a career around understanding complex systems \[00:11]. Combining this with an engineering approach of reverse engineering and improving systems, the presenter focused on AI about 3 years ago \[02:01]. Core Concept: The presenter realized that AI learns from abstractions of data sets, focusing on the underlying shape within any data set \[02:22]. This concept is rooted in topological geometry and quantum physics \[02:44]. Zyra Class Models: The Zyra class models are Zeta-inspired AI geometry models built upon AI geometry principles and resonance-based learning dynamics \[03:24]. They utilize insights from Ryman zaa zos as foundational spectral filters \[04:18]. Ryman hypothesis is used, which states that there's a real part of zero (anchor point) and an imaginary part of zero (rotational axis) \[04:29]. By analyzing the distribution of imaginary parts, structural information within noise can be extracted \[05:35]. Training AI Models Backwards: The presenter argues that current AI models are trained backwards by learning from human-constructed labels first, rather than the inherent shape of the data \[06:07]. Using fashion mnist as an example, the presenter explains that the variation within categories like shoes is much higher than in categories like pants, yet both are arbitrary human labels \[06:17]. The Zyra method proposes learning the actual shape of the data first, using that as ground truth, and then mapping human-constructed labels onto that shape \[08:38]. Experiments and Results: The presenter details various experiments conducted to validate the Zyra model \[10:29]. Initial tests showed that a model based on the research paper achieved 92% accuracy, outperforming a standard CNN \[11:00]. Further experiments involved tweaking dials related to the Rymanian hypothesis and observing the model's behavior \[12:11]. The presenter demonstrated the process of learning the shape of the data, mapping human constructs, and freezing weights to measure the difference \[14:21]. The first test of this system achieved 90% accuracy with a neural network as both encoder and decoder \[17:47]. Subsequent improvements included changing the encoder to a CNN, reducing the size of the autoencoder, and setting a threshold for accuracy \[19:11]. The final result on the fashion mnus data set showed the model converging after learning the human constructs \[21:16]. Conclusion: The Zyra class neural networks work on topological geometry, quantum mechanics, and harmonic resonance \[22:55]. They also incorporate swarm intelligence and self-clustering principles \[23:20].