A spiking YOLO detector cuts sonar-object energy by 3 to 6 times

What the work claims

The authors claim that forward-looking sonar imagery is structurally well matched to spiking neural-network computation. Because sonar returns are sparse, high-contrast, and edge-dominated, a fully spiking detector can fire only for informative pixels and use accumulate-only operations instead of multiply-accumulate operations. They benchmark SpikeYOLO against YOLOv8m, RT-DETR-L, and other baselines on three datasets and report competitive detection accuracy at much lower theoretical energy.¹

How it works

Forward-looking sonar produces a 2D acoustic image of objects ahead of an underwater vehicle. The images lack the colour and fine texture of optical photographs; instead they show bright object returns, dark acoustic shadows, and speckle noise from wave interference. The authors argue that this sparse, high-contrast structure is a natural fit for spike coding, where neurons remain silent unless their input exceeds a threshold.

SpikeYOLO is a fully spiking object detector trained end-to-end with surrogate gradients. It follows the YOLOv8 backbone-neck-head layout but replaces the C2f modules with two custom spiking blocks. Early layers use inverted separable convolutions; later layers use re-parameterisation convolutions for fewer parameters. The neurons are integer-valued leaky-integrate-fire units, which output integers during training to reduce quantisation error but behave like soft-reset LIF neurons at inference.

The energy model follows Horowitz's 45 nm CMOS estimates: 4.6 pJ per multiply-accumulate operation and 0.9 pJ per accumulate-only operation. Because spikes are binary, a spiking synaptic operation is an accumulate rather than a multiply-accumulate. Energy scales linearly with the number of simulation timesteps T. The authors use T equals 2, which they find to be Pareto-optimal across accuracy and energy on all three datasets.

The evaluation uses three public datasets. UATD contains 9200 images of ten shape classes collected with a Tritech Gemini multibeam sonar. Marine-Debris-FLS contains 1868 watertank images of ten debris classes. WHFLS contains 3752 open-water images of victim, boat, and plane classes collected with a BlueView M900 sonar. Mean average precision is averaged over three independent random seeds.

Where a skeptic should push

The most load-bearing assumption is that the theoretical energy numbers translate to real hardware savings. The paper reports 97 mJ versus 322 mJ per inference for YOLOv8m on UATD, 73 mJ versus 322 mJ on Marine-Debris-FLS, and 52 mJ versus 322 mJ on WHFLS. These are calculated from operation counts and the Horowitz model, not measured on a deployed neuromorphic chip. The gap could shrink once data movement, on-chip memory, and conversion overhead are included.

Second, the accuracy advantage is not uniform. On the strict mean-average-precision metric that averages across intersection-over-union thresholds from 0.5 to 0.95, SpikeYOLO trails YOLOv8m by 0.046 on UATD, 0.020 on Marine-Debris-FLS, and 0.011 on WHFLS. The authors interpret this as a box-localisation gap rather than a detection gap, but it is still a gap. The largest dataset, UATD, shows the biggest accuracy sacrifice.

Third, the mean spike activations are high enough to weaken the sparsity story. Measured spike activations per neuron per timestep are 0.880 on UATD, 0.668 on Marine-Debris-FLS, and 0.476 on WHFLS, where the maximum integer output is 4. At 88 percent of the maximum on UATD, the network is not dramatically sparser than a low-precision conventional network, and most of the energy saving comes from replacing multiply-accumulate with accumulate operations.

Fourth, the datasets are still constrained. UATD and Marine-Debris-FLS contain controlled or semi-controlled scenes. WHFLS adds real ocean clutter, but all three are static image datasets. A real AUV must also handle motion blur, varying range, and changing seabed texture, none of which are tested here.

OI implication: spike coding for structured sensors

For organoid intelligence and biological computing, the paper is instructive rather than competitive. It shows that the largest energy wins from spiking computation come when the input domain is already sparse and event-like. Forward-looking sonar, with its bright returns and dark shadows, is one such domain. Neural tissue, with its sparse action potentials and rich temporal dynamics, is another. The question is whether organoid readouts can be encoded in a similarly sparse form.

The opportunity is to borrow the design philosophy. If an organoid-computing system is paired with sparse sensors, whether acoustic, tactile, or event-based visual, the same accumulate-only logic could apply. A spike-based readout from a multi-electrode array already has binary structure; the challenge is to route and weight those spikes efficiently rather than converting them back to dense floating-point vectors at every step.

The threat is that silicon SNNs are now good enough to exploit the same sparse structure without the life-support burden. If the energy argument for organoid intelligence rests mainly on spike sparsity, this paper shows that conventional spiking hardware can capture most of that benefit on a well-chosen task. The biological substrate must therefore justify itself on properties that silicon SNNs lack, such as developmental plasticity, intrinsic homeostasis, or the ability to learn from chemical and electrical signals simultaneously.

The honest conclusion is that SpikeYOLO is a useful reference point for how to build efficient perception around sparse, structured inputs. It does not replace organoid intelligence, but it raises the bar. Future organoid systems should either outperform it on some dimension or explicitly target sensor modalities and learning problems where living tissue is genuinely advantaged.

The bottom line

The paper demonstrates that a fully spiking YOLO detector can process forward-looking sonar imagery at 3.3 to 6.2 times lower theoretical energy than YOLOv8m, with a small but real drop in strict mean average precision and improved robustness to speckle noise. The energy estimates are theoretical, the networks are not deployed on real AUVs, and the largest dataset shows the largest accuracy gap. For organoid intelligence, the value of the work is to show where spiking computation wins: on sparse, structured inputs. Biological computing must now show that living tissue can exploit that same sparsity while adding capabilities that silicon cannot replicate.