sc_neurocore_engine/
recurrent.rs

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

sc_neurocore_engine/
recurrent.rs
