echohive - code, art, learn, do

more coming very soon

Become a Member Consulting

Blog & Research & Learnings

App Source Code & Exclusive Videos 500+ more at Patreon

neural cellular automata

TRY THE RANDOMIZE: for some insane emergent behavious. each cell has a vector state [r, g, b, (4 hidden)]. we have a convolution just evolve randomly, no training yet. This is inspired by this paper: https://distill.pub/2020/growing-ca. 1. Each cell has a state vector of 7 channels (RGB + 4 hidden channels). 2. The convolutional neural network processes the entire grid simultaneously: - First conv layer (3×3 kernel) processes local neighborhoods - Second conv layer (1×1 kernel) mixes features spatially. 3. The network output is blended with the current state using an update rate (0.5). 4. Patterns emerge from random weight initialization - no training required! 5. Each reset randomizes the network weights, ensuring different patterns from the same seed

statistical distribution of flocks of particles as they move

as particles move around randomly we plot their distribution at the bottom as they move, each bar represents how many particles are present at that vertical section ---CODE IS ATTACHED BELOW THE BUTTON---

for paid members

MoE Arithmetic Network: One Router, Two Expert Networks (Add & Sub), trains all 3 together

Built a Mixture of Experts (MoE) neural network that learns addition and subtraction by routing problems to specialized experts. ---CODE ATTACHED BELOW--- How it works: A router network looks at each problem (e.g., "5 + 3" or "20 - 7"). It decides which expert should handle it. Expert Add specializes in addition. Expert Sub specializes in subtraction. The router learns to send addition problems to Expert Add and subtraction problems to Expert Sub. The cool part: All three networks train together. The router learns routing, and each expert learns its operation. Over time, the router gets better at choosing the right expert. Visualization: Real-time training dashboard with: Live network architecture visualization, Training metrics (loss, accuracy, expert usage), Interactive testing after training.

for paid members

Mixture of experts mouse with 2 networks that learns to find cheese and avoid predators with Hebbian learning

Mixture of experts mouse learns to find cheese and avoid predator with 2 networks which learn via Hebbian rule. 2 ring attractor networks separately detect cheese and predator direction and control their own motor neurons and learn separately. There is a 3rd set of motor neurons that get weighted sum of both networks and actually control the mouse. With the slider we can adjust the weight of influence each network has. Will experiment more with multiple networks like these and share with you my experiments and findings as there can be more sophisticated gating mechanisms between the 2 networks. --- SOURCE CODE ATTACHED BELOW---

for paid members

Mechanical neural network learns by dropping pebbles, gravity powered.

A gravity-powered, physics-based visualization of a neural network that learns through mechanical adjustments. Watch pebbles drop through 3 layers of tilting levers to learn simple addition! includes a detailed readme and "how it works" page...

Spiking neuron flow

This is mostly fun but also educational(bio inspired) spiking neuron activations interactive animations. --- CODE IS ATTACHED BELOW THE POST--- - Action Potentials: Neurons spike with realistic membrane potential dynamics - Refractory Period: Neurons cannot fire again until the refractory period passes - Fatigue Mechanism: Neurons become less responsive with excessive firing and recover when resting - Propagation Delay: Spikes travel along connections with distance-based delays - Transmission Failure: Probabilistic synapse failure for realistic behavior - Hebbian Plasticity: Connections strengthen with correlated activity

L-system pattern generator

An L-system (Lindenmayer System) is a way to build complex patterns from simple rules. Biologist Aristid Lindenmayer came up with it in 1968 to model how plants grow. It works like this: start with a simple string, then apply rules that replace parts of it. Each step makes it longer and more complex. If you interpret that final string as drawing commands, you get patterns like trees, fractals, and other shapes. It's like a recipe that keeps rewriting itself, simple instructions that produce complex results. ---CODE ATTACHED BELOW---

for paid members

Neuroevolution training of Spiking neural network

50 neurons arranged in a 5×10 grid. Each neuron fires brief electrical "spikes," and only the last two columns represent digits 0–9 (one neuron per digit). When we want the network to add two numbers, we feed the first number's spikes into column 0 and the second number's spikes into column 1. Code is attached below the post.

Neural network learns to play the piano with Hebbian rule

This project demonstrates a feed-forward neural network that learns musical patterns through Hebbian learning, a simple, biologically plausible learning rule based on the principle "neurons that fire together, wire together." The network receives musical notes as input, processes them through 72 interconnected neurons, and outputs musical notes in response, creating a feedback loop where the network learns from its own behavior. ---CODE IS ATTACHED BELOW AS ZIP---

for paid members

Branching Probability Network webapp using Markov chains - UPDATE WITH VERSION 2

A webapp for creating branching probability networks using Markov chains. It has all the customization one can imagine and an Auto mode where endless many patterns just emerge. Includes a learn page and a readme where you can learn more about markov chains.