sc_neurocore_engine/
recurrent.rs

1// SPDX-License-Identifier: AGPL-3.0-or-later
2// Commercial license available
3// © Concepts 1996–2026 Miroslav Šotek. All rights reserved.
4// © Code 2020–2026 Miroslav Šotek. All rights reserved.
5// ORCID: 0009-0009-3560-0851
6// Contact: www.anulum.li | protoscience@anulum.li
7// SC-NeuroCore — SC recurrent/reservoir layer (echo state network)
8
9//! SC recurrent/reservoir layer (echo state network).
10//!
11//! Jaeger, GMD Report 148, 2001.
12
13use rand::{RngExt, SeedableRng};
14use rand_chacha::ChaCha8Rng;
15
16/// Recurrent reservoir layer with state feedback.
17#[derive(Clone, Debug)]
18pub struct RecurrentLayer {
19    pub n_inputs: usize,
20    pub n_neurons: usize,
21    /// Input weights, flat [n_neurons, n_inputs].
22    pub w_in: Vec<f64>,
23    /// Recurrent weights, flat [n_neurons, n_neurons].
24    pub w_rec: Vec<f64>,
25    /// Current state [n_neurons].
26    pub state: Vec<f64>,
27}
28
29impl RecurrentLayer {
30    pub fn new(n_inputs: usize, n_neurons: usize, seed: u64) -> Self {
31        let mut rng = ChaCha8Rng::seed_from_u64(seed);
32        let w_in: Vec<f64> = (0..n_neurons * n_inputs)
33            .map(|_| rng.random::<f64>() * 0.5)
34            .collect();
35        let w_rec: Vec<f64> = (0..n_neurons * n_neurons)
36            .map(|_| rng.random::<f64>() * 0.2)
37            .collect();
38        Self {
39            n_inputs,
40            n_neurons,
41            w_in,
42            w_rec,
43            state: vec![0.0; n_neurons],
44        }
45    }
46
47    /// Process one time step. Returns reference to updated state.
48    #[allow(clippy::needless_range_loop)]
49    pub fn step(&mut self, input: &[f64]) -> &[f64] {
50        assert_eq!(input.len(), self.n_inputs);
51        let mut new_state = vec![0.0; self.n_neurons];
52        for i in 0..self.n_neurons {
53            let mut val = 0.0;
54            for j in 0..self.n_inputs {
55                val += self.w_in[i * self.n_inputs + j] * input[j];
56            }
57            for j in 0..self.n_neurons {
58                val += self.w_rec[i * self.n_neurons + j] * self.state[j];
59            }
60            new_state[i] = val.clamp(0.0, 1.0);
61        }
62        self.state = new_state;
63        &self.state
64    }
65
66    pub fn reset(&mut self) {
67        self.state.fill(0.0);
68    }
69}
70
71#[cfg(test)]
72mod tests {
73    use super::*;
74
75    #[test]
76    fn state_changes_after_step() {
77        let mut layer = RecurrentLayer::new(3, 5, 42);
78        let input = vec![0.5, 0.3, 0.8];
79        let state = layer.step(&input).to_vec();
80        assert!(state.iter().any(|&s| s > 0.0));
81    }
82
83    #[test]
84    fn reset_clears_state() {
85        let mut layer = RecurrentLayer::new(3, 5, 42);
86        layer.step(&[0.5, 0.3, 0.8]);
87        layer.reset();
88        assert!(layer.state.iter().all(|&s| s == 0.0));
89    }
90
91    #[test]
92    fn state_bounded() {
93        let mut layer = RecurrentLayer::new(2, 4, 99);
94        for _ in 0..100 {
95            layer.step(&[1.0, 1.0]);
96        }
97        assert!(layer.state.iter().all(|&s| (0.0..=1.0).contains(&s)));
98    }
99
100    #[test]
101    fn output_shape() {
102        let mut layer = RecurrentLayer::new(4, 8, 0);
103        let out = layer.step(&[0.1, 0.2, 0.3, 0.4]);
104        assert_eq!(out.len(), 8);
105    }
106}
sc_neurocore_engine/recurrent.rs

sc_neurocore_engine/
recurrent.rs
