Log-Antilog Neural Circuit for Interpretable Differentiable Learning

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Traditional neural networks face challenges in balancing model complexity with interpretability, stability of differentiability, and overfitting, particularly in deep architectures, which lack transparency and require immense computational resources.

Innovation Solution

A circuit-based neural network architecture with two or three layers, utilizing multiply-add units and logarithmic/antilog operations to compute interpretable and differentiable functions, incorporating storage and control units for weight updating and gradient computation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If deep neural network architectures with large number of layers are used, then computational power and learning capacity are improved, but device complexity and computational resource consumption increase significantly

Engineering Contradiction:

Improvelearning capacityVSAvoidnetwork architecture complexity

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent segments the neural network into functional blocks: first level processing units with multiply-add and logarithmic units, second level processing units with multiply-add and antilog units. This modular segmentation allows achieving high learning capacity with reduced complexity by organizing computations into manageable functional segments rather than deep hierarchical layers.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent changes the computational parameters by using logarithmic and antilogarithmic functions instead of traditional activation functions. This parameter transformation enables the network to achieve expressive power comparable to deep networks while using shallower architectures, as the log/antilog operations provide strong non-linearity and representational capacity in fewer layers.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If traditional neural networks are used, then computational power is sufficient for complex mappings, but interpretability is lost due to black-box nature

Engineering Contradiction:

Improvecomputational powerVSAvoidinterpretability

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The patent introduces intermediate representations through the first and second level processing units that provide interpretable explanations. The multiply-add units compute weighted sums that can be inspected, and the logarithmic/antilog units provide transformed representations that maintain differentiability while offering structural interpretability. These intermediary layers serve as mediators between input and output, enabling understanding of computational steps.

Inventive Principle:

Principle #24Intermediary (Mediator)

3Adaptability or versatility

If deep neural networks are used, then learning capacity is improved, but differentiability becomes numerically unstable in very deep models

Engineering Contradiction:

Improvelearning capacityVSAvoiddifferentiability stability

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent inverts the traditional deep network approach by using shallow networks with logarithmic and antilogarithmic operations. Instead of pushing depth to increase capacity, the patent uses the mathematical properties of log/antilog functions to achieve capacity in fewer layers, thereby maintaining numerical stability in differentiability while preserving learning capacity through functional composition rather than depth accumulation.

Inventive Principle:

Principle #13The other way round (Inversion)

4Adaptability or versatility

If standard neural networks with large number of parameters are used, then learning capacity is improved, but overfitting occurs on limited or noisy datasets

Engineering Contradiction:

Improvelearning capacityVSAvoidoverfitting

Core Design Contradiction:

Adaptability or versatilityVSObject-generated harmful factors

Solution Approach 1:

The patent applies partial action by using a reduced number of parameters organized in specific functional blocks. Instead of employing massive parameter counts to capture complexity, the patent uses a smaller set of parameters in the first and second level processing units, leveraging the logarithmic/antilogarithmic transformations to achieve sufficient representational capacity without excessive parameters that would cause overfitting on limited data.

Inventive Principle:

Principle #16Partial or excessive action

Data Source

PatentUS12625677B2Circuit for efficiently performing operations on input data to compute an interpretable and differentiable function

Publication Date: 2026.05.12 SPARSEMIND TECHNOLOGY LABS PTE LTD

AI summary

A circuit for efficiently performing operations on input data to compute an interpretable and differentiable function, comprising a first level processing unit that obtains one or more first level inputs and a second level processing unit. The first level processing unit comprises (i) a first level MA unit that is configured to compute a first level weighted sum of the one or more first level inputs, and (ii) a logarithmic unit that is configured to compute a first level output. The second level processing unit obtains one or more second level inputs. The second level processing unit comprising (i) a second level MA unit that is configured to compute a second level weighted sum of the one or more second level inputs, and adding the computed second level product, and (ii) an antilog unit that is configured to compute a second level output.