Neuro-Inspired Visual Pattern Recognition via Biological Reservoir Computing

Abstract: In this paper, we present a neuro-inspired approach to reservoir computing (RC) in which a network of in vitro cultured cortical neurons serves as the physical reservoir. Rather than relying on artificial recurrent models to approximate neural dynamics, our biological reservoir computing (BRC) system leverages the spontaneous and stimulus-evoked activity of living neural circuits as its computational substrate. A high-density multi-electrode array (HD-MEA) provides simultaneous stimulation and readout across hundreds of channels: input patterns are delivered through selected electrodes, while the remaining ones capture the resulting high-dimensional neural responses, yielding a biologically grounded feature representation. A linear readout layer (single-layer perceptron) is then trained to classify these reservoir states, enabling the living neural network to perform static visual pattern-recognition tasks within a computer-vision framework. We evaluate the system across a sequence of tasks of increasing difficulty, ranging from pointwise stimuli to oriented bars, clock-digit-like shapes, and handwritten digits from the MNIST dataset. Despite the inherent variability of biological neural responses-arising from noise, spontaneous activity, and inter-session differences-the system consistently generates high-dimensional representations that support accurate classification. These results demonstrate that in vitro cortical networks can function as effective reservoirs for static visual pattern recognition, opening new avenues for integrating living neural substrates into neuromorphic computing frameworks. More broadly, this work contributes to the effort to incorporate biological principles into machine learning and supports the goals of neuro-inspired vision by illustrating how living neural systems can inform the design of efficient and biologically grounded computational models.

• The paper demonstrates that cultured cortical neurons connected to a 4,096-electrode HD-MEA can act as a biological reservoir, enabling linear classification of visual patterns from evoked spike-count states.
• The paper reports 98% accuracy for point stimuli, 92% for oriented bars, about 71% for ten clock-digit patterns, and above-chance performance on a 200-image MNIST subset.
• The paper finds that short 5–10 ms readout windows preserve the most useful information, while cross-day accuracy falls sharply, highlighting biological drift as the main barrier to reliable deployment.

Overview

This paper presents a biologically grounded instantiation of reservoir computing (RC) in which the reservoir is not a simulated recurrent network but a physical network of in vitro cultured cortical neurons. The authors—spanning computer science, biophysics, and developmental biology laboratories in Italy—interface mouse embryonic stem cell (mESC)-derived cortical cultures with a high-density multi-electrode array (HD-MEA) of 4,096 electrodes, deliver spatially patterned electrical stimulation that mimics visual inputs, record the evoked spiking activity across the array, and train a single-layer perceptron (SLP) on the resulting 4096-dimensional reservoir states. The work extends two earlier preliminary studies by substantially broadening the experimental scope, most notably through an evaluation on MNIST handwritten digits, and by adding systematic ablations and a thorough biological characterization of the substrate (2602.05^^^^171^^^^).

The central claim is that living cortical networks, despite their noise, spontaneous activity, and day-to-day drift, generate sufficiently discriminative high-dimensional representations to support static visual pattern recognition via a purely linear readout. The evidence supporting this claim spans four tasks of increasing difficulty, and the paper is candid about where biological variability degrades performance.

Biological substrate and system architecture

The biological reservoir consists of entorhinal-like neurons derived from mESCs using the MiBi protocol (MAPK/ERK and BMP inhibition), which yields cultures whose gene expression closely matches the early postnatal entorhinal cortex and which exhibit high structural and functional connectivity, synchronized bursting, and self-organized criticality evidenced by scale-invariant neuronal avalanches—a regime often associated with optimal computation. Cultures are plated on HD-MEA chips around days 20–25 in vitro and stimulated experimentally after DIV40, i.e., at least three weeks post-plating.

Input encoding maps each pixel of an $H \times W$ input grid onto a corresponding electrode pair operated in bipolar mode; pixel intensity modulates stimulation probability or amplitude. Stimulation uses monophasic or biphasic rectangular pulses with randomized stimulus ordering and a fixed 10 s inter-stimulus interval, chosen to allow return to baseline and to avoid sequence effects. Spike detection uses a real-time double-threshold algorithm at a 20 kHz sampling rate; postprocessing removes stimulation artifacts via an amplitude threshold ($V_{thr} = 500\ \mu V$) plus a normalized-area criterion ($w_{thr} = 25$) that filters out unusually wide events. Each stimulus evokes a 4096-dimensional feature vector of per-electrode spike counts within a post-stimulus window $W$, with the stimulated region optionally excluded to test distal propagation. The readout is trained with SGD under cross-entropy for 1,000 epochs and evaluated with 5-fold cross-validation; no validation set or early stopping is used, so all data contribute to both training and evaluation.

A notable methodological point: the artificial reservoir (AR) baseline—4,096 rate-based units with 10% sparsity and spectral radius 0.9, driven by the same patterns with empirically measured noise—is explicitly framed as an approximate upper bound rather than a peer competitor, since it operates under idealized, fully optimized conditions. This framing matters for interpreting every comparison below.

Results across task complexity

Pointwise stimuli. Four adjacent-electrode pairs stimulated at $10\ \mu A$, $20\ \mu s$ yield 98% ± 2% average accuracy, outperforming the AR baseline (82% ± 6%). The authors attribute the AR's weakness here to their noise-injection scheme perturbing isolated pointwise signals—an artifact of the baseline design specific to this scenario rather than evidence of biological superiority.

Oriented bars. Four bar orientations (0°, 45°, 90°, 135°) sharing a common center produce overlapping activation zones, preventing trivial spatial separability. The BRC achieves 92% ± 6%, slightly below the AR's 98% ± 3%—the expected ordering given the upper-bound framing.

Clock-digit-like patterns. Ten digit configurations mapped as seven-segment-display-like electrode subsets, delivered with reduced biphasic amplitude ($4\ \mu A$/pair), yield approximately ^^^^1^^^^1% average accuracy over nine sessions across three biological replicates—comparable to the AR and far above a spatially shuffled random-guess baseline (12%). Digits "1" and "^^^^1^^^^" are consistently easiest, plausibly because they activate fewer electrodes with limited overlap; the authors flag this explanation as speculative.

MNIST. A 200-image subset is downsampled, stretched to fit electrode pairs, and encoded probabilistically (darker pixels stimulate with higher probability). With biphasic stimulation at PRESERVED_PLACEHOLDER_^^^^1^^^^ and $W = 10$ ms, the BRC reaches accuracy far above chance (~9%) and comparable to the AR for several digits. Digit "9" performs worst in all conditions, attributed to class imbalance in the sampled subset. This is the strongest result in the paper: it demonstrates that a living neuronal culture can encode structure in real-world, non-synthetic visual inputs well enough for linear decoding.

Ablation findings

Three ablations carry most of the analytical weight. First, the post-stimulus window $W$: accuracy peaks at 5 ms for clock digits and 10 ms for MNIST, then declines monotonically toward 50 ms. The interpretation is biologically grounded—the optimal window matches roughly one (or one to two, for MNIST) synaptic transmissions, capturing direct postsynaptic activation before recurrent dynamics, spontaneous activity, and noise dominate. That the optimum doubles for MNIST is explained by the larger number of stimulated electrodes producing stronger, more broadly propagated responses. Second, day-wise analysis shows stable performance within each day (e.g., 69–^^^^1^^^^3% for clock digits across Days 1–3). Third—and this is the most important negative finding—cross-day generalization degrades sharply: training on Day 1 and testing on later days drops clock-digit accuracy to 46 ± 15% (Day 2) and 3^^^^1^^^^ ± 19% (Day 3); for MNIST, to 32 ± 3% and 28 ± 5%. Both remain above chance, indicating partial preservation of stimulus-evoked structure, but the decline reflects genuine reconfiguration of network dynamics driven by development and intrinsic plasticity. t-SNE projections corroborate this picture qualitatively: biological reservoir states form broader, more overlapping class structures than the compact, well-separated clusters of the AR, indicating partial loss of discriminative information that nonetheless remains recoverable by the linear readout.

Limitations and open questions

The paper is explicit about several constraints. The cross-day degradation means the current framework effectively requires per-session retraining of the readout; no mechanism exists yet to stabilize or compensate for drift in the biological substrate. All experiments use three biological replicates of a single culture type (MiBi-derived), so generalization across culture preparations, ages, and species is untested. The MNIST evaluation uses only 200 images, leaving scalability of both the experimental protocol and the representational capacity open. The AR baseline injects noise in a particular way that disadvantages it on pointwise stimuli, so the claim that BRC can outperform AR should be read narrowly. Finally, the explanation for why certain digit classes (notably "^^^^1^^^^" and "1") are more robust remains conjectural, and the energy-efficiency argument for BRC is asserted rather than quantified—no power measurements or comparisons against neuromorphic hardware are reported.

Conclusion

This paper establishes, through a carefully staged experimental progression, that in vitro cortical networks interfaced with HD-MEAs can function as physical reservoirs for static visual pattern recognition, achieving ~^^^^1^^^^1% on ten-class synthetic digit patterns and clearly above-chance classification of handwritten MNIST digits with a linear readout alone. The most consequential open problem raised by the results is temporal stability: stimulus-evoked representations drift substantially across days, and whether plasticity-inducing stimulation protocols or adaptive readout schemes can restore long-term robustness remains unresolved.