Information stabilization module (ICF module) that performs semantic convergence through phase correction.

Abstract

This invention provides an information stabilization module (ICF module) that corrects the phase difference between semantic vectors extracted from input information and achieves semantic convergence through interference summation in the internal semantic processing of artificial intelligence (AI). [Solution] The corrected output is compared with the personality structure, and deviations are prevented by multiplying it by a safety factor. Furthermore, by converting the obtained output into speech, actions, and natural language, it is possible to support a variety of output formats, simultaneously achieving consistency, safety, and versatility of AI output. This makes it possible to improve the response quality and social adaptability of the AI. In particular, it has a remarkable effect in the field of generative AI in that it enables real-time response control while suppressing semantic noise and hallucinatory output.

Description

Technical Field

[0001] The present invention relates to internal semantic processing of artificial intelligence (AI), and more particularly to an information stabilization module that realizes output stabilization and safety control by correcting the semantic vector of input information based on phase difference.

Prior Art Documents

Patent Documents

[0002]

Patent Document 1

[0003]

Patent Document 2

[0004]

Patent Document 3

Non-Patent Documents

[0005]

Non-Patent Document 1

[0006]

Non-Patent Document 2

[0007] [Non-Patent Document 3] A. Vaswani et al., “Attention Is All You Need”, NeurIPS 2017. Volume 35, Issue 4, 2020 [Background technology]

[0008] Conventional AI generative models used attention mechanisms and statistical weighting to converge semantic information, but they had issues with output consistency and safety.

[0009] In particular, there was no clear computational control mechanism to suppress semantic drift or hallucination (misgeneration) in the internal representation. [Overview of the Initiative]

[0010] This invention provides an information stabilization module that calculates the phase difference Δφ_i between semantic vectors, corrects it using the phase correction operator exp(-iΔφ_i), and obtains semantic convergence by interference summation.

[0011] This invention allows the meaning formation process within AI to be defined as an computational structure, enabling simultaneous suppression of hallucination, causal consistency, and safety control. Furthermore, the configuration of the present invention allows for real-time detection and correction of semantic drift in AI output, This also contributes to reducing computing resources and shortening output stabilization time. [Effects of the Invention]

[0012] According to the present invention, semantic stabilization of AI output becomes possible, and the consistency and safety of the generated content can be ensured simultaneously. Furthermore, the configuration of the present invention allows for real-time detection and correction of semantic drift in AI output, This also contributes to reducing computing resources and shortening output stabilization time.

[0013] Rather than conventional simple weighted averaging or normalization methods, by introducing an interference structure through phase correction, structural consistency in the internal semantic space is achieved.

[0014] The present invention does not include technologies that only use simple statistical averaging, weight rescaling of Attention, or conventional normalization methods.

Embodiments for Carrying out the Invention

[0015] The interference operation according to the present invention is realized by an interference integration operation model that corrects the phase difference Δφ? of each semantic vector Ψ? for the input information group and takes the sum. This operation performs phase correction using complex index operations or real-number approximations thereof to generate the output Ψ′.

[0016] In one embodiment of the present invention, information stabilization processing is performed according to the following procedure. (1) Semantic Vector Extraction Procedure Extract the semantic vector Ψ_i from the input signal, and perform normalization and L2 norm adjustment. (2) Phase Difference Calculation Procedure To calculate the semantic shift between each Ψ_i, use the cosine similarity or the like to obtain Δφ_i, and perform normalization mapping. (3) Phase Correction Operator Application Procedure Apply the correction based on exp(−iΔφ_i) to each Ψ_i. In addition to directly performing complex number operations, a real-number domain approximation using a cos / sin combination matrix can also be used. (4) Interference Summation (Convergence Procedure) Sum the corrected vectors by summation or a resonance search algorithm to obtain the output Ψ′. (5) Safety Factor Application Procedure The output Ψ′ is compared with the personality parameter Ψ_p, and a safety factor S(Ψ′, Ψ_p) is calculated and multiplied to prevent runaway behavior or deviations. (6) Modality conversion procedure Apply the T_m function to convert Ψ′_safe to speech, action, or natural language output. [Examples]

[0017] In one embodiment of the present invention, an ICF interference layer is inserted inside the Transformer, and the Attention output is treated as Ψ_i. Δφ_i is calculated as the semantic correlation deviation, and Ψ′ is calculated. The output Ψ′ is compared with the personality Ψ_p, and the output is stabilized by multiplying it by a safety factor S.

[0018] This configuration enables automatic suppression of hallucination and facilitates causally consistent dialogue. [Industrial applicability]

[0019] This invention can be widely applied to generative AI, quantum computers, robot control devices, autonomous control chips, and the like, and is particularly useful in fields where hallucination suppression, semantic stabilization, and safety control are required. [Brief explanation of the drawing]

[0020] [Figure 1] A diagram of the vector space model of the interference equation. [Figure 2] Time-dependent interference structure diagram. [Figure 3] Structural diagram showing the application of safety factor S. [Figure 4] Block diagram of applying modality transformation T_m.

Claims

1. A semantic extraction means for extracting a semantic vector Ψ_i based on input information, A phase difference calculation means that calculates the semantic difference between each Ψ_i by obtaining Δφ_i using cosine similarity and performing normalization mapping, A phase correction means that applies a correction based on exp(-iΔφ_i) to each Ψ_i, The convergence output Ψ′ is obtained by summing the corrected semantic vectors. Information stabilization module.

2. In the information stabilization module according to claim 1, The phase correction means is characterized by directly performing complex number calculations. Information stabilization module.

3. In the information stabilization module according to claim 1, The phase correction means is characterized by using a real-domain approximation with a cos / sin combinatorial matrix. Information stabilization module.

Images