Skip to content

Differences with Keras

A Complete Guide

This document provides a comprehensive guide for translating Keras concepts into Synalinks. While Keras is designed for building traditional neural networks with tensor operations, Synalinks is a framework for creating neuro-symbolic programs that combine language models with structured reasoning.

Fundamental Paradigm Shift

Keras: Neural Network Framework

  • Purpose: Build and train deep neural networks using tensor operations
  • Core abstraction: Mathematical tensors flowing through differentiable layers
  • Training: Gradient-based optimization (backpropagation)
  • Computation: Matrix multiplications and activation functions
  • Use cases: Computer vision, time series, traditional ML tasks
  • Purpose: Build intelligent applications combining LMs with symbolic reasoning
  • Core abstraction: Structured JSON data flowing through modular programs
  • Training: Reinforcement learning with LM-based optimizers
  • Computation: Language model inference + symbolic operations
  • Use cases: AI agents, reasoning systems, structured generation, API orchestration

Quick Concept Mapping

Keras Concept Synalinks Equivalent Key Difference
Layer Module Processes JSON instead of tensors
Model Program DAG of modules with conditional logic
Tensor Data Model JSON object with schema validation
Tensor shape JSON schema Explicit structure definition
Weights/biases Trainable Variable JSON objects, not floating-point arrays
Loss function Reward function Maximize reward vs minimize loss
Backpropagation LM-based optimization No gradients; uses language model reasoning
model.compile() program.compile() Sets up LM optimizer instead of SGD/Adam
model.fit() program.fit() Reinforcement learning loop

Core Concepts Explained

Module (Layer Equivalent)

A Module is a self-contained computational unit that: - Receives: JSON Data Models conforming to input schemas - Processes: Via LM calls, symbolic operations, or hybrid approaches - Outputs: JSON Data Models conforming to output schemas

Key Module Properties:

  • Keras Layer: Receives tensors, performs matrix operations, outputs tensors
  • Synalinks Module: Receives JSON, performs LM/symbolic operations, outputs JSON with schema validation

Program (Model Equivalent)

A Program orchestrates multiple Modules into a directed acyclic graph (DAG):

  • Keras: Fixed computation graph of tensor operations
  • Synalinks: Dynamic graph with conditional branching based on LM decisions

In Keras, you build sequential or functional models with fixed layer connections. In Synalinks, you create programs that can include conditional branches, where different modules execute based on LM-driven decisions or data conditions.

Data Models & JSON Schemas

Instead of implicit tensor shapes, Synalinks uses explicit JSON schemas:

  • Keras: Input shape defined as dimensions (e.g., 784-dimensional vector)
  • Synalinks: Input structure defined as JSON schema with explicit fields, types, and validation rules

Why JSON? - Interpretability: Human-readable intermediate states - Validation: Built-in schema validation - Interoperability: Native compatibility with APIs and web services - Debugging: Easy to inspect and modify - Structured generation: Natural fit for LM constrained output

Training Paradigm Differences

Keras: Gradient Descent

  • Defines differentiable loss function (e.g., categorical crossentropy)
  • Computes gradients via automatic differentiation (backpropagation)
  • Updates weights using gradient information through optimizers like Adam or SGD
  • Requires continuous, differentiable operations throughout
  • Defines reward function (higher values indicate better performance)
  • No gradient computation—uses language models to reason about improvements
  • LM analyzes current performance and proposes better configurations
  • Updates trainable variables (JSON objects) based on LM suggestions
  • Can incorporate discrete decisions and non-differentiable operations

Key Training Differences:

Aspect Keras Synalinks
Objective Minimize loss Maximize reward
Optimization Gradient descent Reinforcement learning
Update mechanism Mathematical derivatives LM-generated improvements
Trainable params Float tensors Structured JSON objects

When to Use Each Framework

Use Keras when:

  • Working with numerical data (images, signals, time series)
  • Need fast, vectorized computations
  • Have large labeled datasets
  • Problem has clear differentiable objectives
  • Deploying to edge devices with limited resources
  • Building LM-powered applications
  • Need symbolic reasoning and logic
  • Working with structured/semi-structured text data
  • Require interpretable intermediate steps
  • Building API orchestration or agent systems
  • Need adaptive, context-aware processing
  • Want to combine multiple LMs and tools

Summary

While Keras excels at building traditional neural networks with tensor operations and gradient-based training, Synalinks is designed for a fundamentally different purpose: creating neuro-symbolic applications that combine the reasoning capabilities of language models with structured, validated data processing.

The shift from tensors to JSON, from gradients to LM-based optimization, and from fixed architectures to adaptive programs represents not just a technical change, but a paradigm shift in how we think about building intelligent systems. Synalinks is ideal when you need interpretability, flexibility, and the ability to integrate language understanding with symbolic reasoning—making it perfect for next-generation AI applications.