Hacker News Reader

Comment by sabalaba | Hacker News Reader

This doesn't look like a neural network that's solving any actual problems beyond visualizing firing patterns. Most ANNs don't have any concept of time built into the model. When you're in feed-forward mode, the underlying computation is simply a bunch of dot products.

However, you might be able to make some cool visualizations if your ANN was a Spiking Neural Network.

[1] https://en.wikipedia.org/wiki/Spiking_neural_network

View on HN

Well it's thresholded so you could visualize the "activated" neurons in a manner such as this.

Would it be useful though? Maybe. Probably not as the fact that a neuron fires isn't nearly as useful as why it fired (what it represents).

But I like that you mention time. It always seemed odd that NN's today completely ignore that aspect. Sure whether a neuron fires is binary. But the accumulation of spikes is not binary and highly temporal. yet we completely ignore this aspect.

I'm confused by you guys saying NN's today completely ignore time? If I have a network of Hodgkin-Huxley neurons and I integrate the equations, I'm integrating over time. I'm not sure how that is ignoring it? Disclaimer: it has been about 4 years since I've written any code to integrate H-H neurons, so I might be forgetting something really obvious. :) But I can't imagine EVERYONE is ignoring time when they simulate networks of neurons.

They are talking about NNs used in machine learning, which are atemporal, unlike H-H or other models of human cognition.

In a machine learning neural network there is no integration; they are basically just nonlinear data transformations that can (usually...) be trained.

Reply on news.ycombinator.com