Beyond Neural Networks: How Artificial Dendrites Could Change AI

Artificial intelligence has made remarkable progress in recent years, but the fundamental building blocks behind today’s neural networks have changed surprisingly little. In fact, the mathematical model of the artificial neuron dates back to 1943, long before much of our modern understanding of neuroscience.

So, with AI increasingly focused on replicating human intelligence, could we be overlooking something important about how the human brain actually works?

That’s the question driving Perforated AI, a company developing artificial dendrites that it claims can make existing neural networks smaller, more accurate and more efficient.

We spoke with Rory Brenner, Founder and CEO of Perforated AI, and Ralph Crewe, Chief Communications Officer, to explore the science behind the technology, its reported results and what it could mean for the future of AI.

Taking Inspiration from the Brain

Rory’s interest in biological intelligence began during his undergraduate studies in computer science. Curious about why biological systems could outperform the AI models of the time, he pursued a PhD in neuroscience to investigate the computational mechanisms behind the brain.

One particular discovery caught his attention: dendrites.

In biological neurons, dendrites play an important role in receiving and processing signals. However, conventional artificial neurons use a considerably simplified mathematical representation of their biological counterparts.

Perforated AI’s approach introduces artificial dendrites into existing neural networks, allowing individual neurons to receive additional computational support.

Unlike conventional training methods, which primarily rely on gradient descent to minimise errors, Perforated AI’s dendrites use a correlation-based reinforcement learning mechanism to identify input patterns associated with high neuron errors.

The dendrites can then provide corrective signals when those patterns appear again.

Rather than requiring an entirely new neural network architecture, the technology is designed to integrate with existing models through a PyTorch plugin.

Smaller Models, Better Results?

While the underlying science is interesting, the practical implications are perhaps more significant.

Perforated AI reports several examples of its technology improving existing AI applications.

One comes from ThoroughAI, a robotics company developing autonomous forklifts and cleaning vehicles. According to Perforated AI, incorporating artificial dendrites helped produce a model that was 70% smaller, achieved a 2.5× faster frame rate and delivered slightly higher accuracy than the original.

For autonomous robots, inference speed has direct operational implications. A system that processes its surroundings more quickly may be able to operate at higher speeds while maintaining the response times needed for safe navigation.

Another example comes from Lasso Loop, which develops automated household recycling technology. Perforated AI reports that its approach improved classification accuracy from 95% to 99%, representing an 80% reduction in classification errors.

The company also participated in an Edge Impulse hackathon, integrating its technology into the platform’s classification block. Across 800 reported experiments, dendritic neural networks consistently demonstrated better accuracy at comparable parameter counts.

These are promising company-reported results, although the interview does not provide the complete benchmarking methodology needed to independently evaluate them.

What Does This Mean for Edge AI?

Perforated AI is currently focusing much of its attention on smaller neural networks, particularly those deployed on resource-constrained edge hardware.

For developers working with embedded AI, reducing parameter counts while maintaining accuracy can be particularly valuable. Smaller models can potentially reduce memory requirements, ease deployment constraints and improve inference performance.

However, there is a trade-off.

Training models with artificial dendrites takes longer per individual run. Rory explains that a model normally requiring one hour to train might take three hours with one dendrite and five hours with two.

Nevertheless, Perforated AI argues that the additional training time can be offset by reducing the number of optimisation experiments required to achieve a desired result.

In some of its ResNet experiments, the company reports reaching performance targets within approximately four experiments that otherwise required selecting the best result from 100 conventional experiments.

If these improvements translate consistently across applications, the implications could extend beyond model performance to the cost and time required for AI development.

Could Dendrites Become Standard in AI?

Although Perforated AI’s current focus is Edge AI, its ambitions extend considerably further.

The company has been expanding its research from relatively small neural networks to architectures containing tens or hundreds of millions of parameters, including work involving ResNet, YOLO and Segment Anything models.

Applying the technology to much larger language models remains an ambition rather than an established capability.

Ultimately, Perforated AI doesn’t want to become another AI model provider. Instead, its goal is to establish artificial dendrites as a standard infrastructure component that developers can incorporate into their existing machine learning workflows.

Rory’s five-year vision is ambitious: a future where building neural networks without dendrites seems as outdated as building modern AI without neural networks themselves.

Whether artificial dendrites can achieve that level of adoption remains to be seen. But Perforated AI is raising an interesting question about the direction of AI development.

Rather than simply making neural networks larger, could making their individual neurons more capable be another path forward?

Watch our full ipXchange interview with Rory Brenner and Ralph Crewe to learn more about the technology, its reported results and Perforated AI’s vision for the future of neural networks.