How do I recognize a dog?

Question

I know, neurons take in data from the eyes and pass the data on through synapses to subsequent neurons and synaptic weights from many neurons pass into the subsequent neuron and the neuron adds up all synaptic weights in the axon hillock and if it is strong enough, it decides to send it to the next neuron through the axon terminal.

But if the neurons just keep adding up numbers and sending those numbers to the next neuron, how does the final neuron understand that the image is of a dog? The last neuron just gets a big number, the sum of all the previous neurons' values. It's just a number then, isn't it?

I've learnt about Hebbian learning and stuff, but because the same synapse can allow data about cats, dogs, rats, any other thing to pass through it, how does it realize which data is about a dog to strengthen the connection? It's just a number isn't it, the sum of all those neurons before it?

I am pretty new to this field, and I haven't studied this in school or anything, but I am a programmer and I thought of a pretty cool idea about AI to simulate the brain, so I've been buggin ChatGPT to explain how the brain works. I've got a pretty good idea right now, but I'm stuck on this point and I can't seem to make sense of it at all. Please help me out, thanks in advance.

PS: I don't even know what tags to apply here, sorry 😅

Answer 1

For the sake of viewing the brain from an analog computer typology, I will try not comment on the visual-perceptual and cognitive processes at play in our perception of recognizing a dog as a dog, which are of great interpretational significance by the way. However, while there is no easy answer to your question in its pure neurological form, Gestalt's top-down, bottom-up perceptual theory may be of most interest to you. According to Gestalt's theory, neuronal processing of stimuli does not occur concurrently with preconceived ideas of what we are being exposed to, neither do we take in every single characteristic of a dog. Rather we take in the most salient identifying features that have the most 'weight' which allows us to interpret what we are seeing - and discriminate between a dog and a cat at the neuronal level. From this view, all stimuli is data-driven based on the most salient features of incoming stimuli. Interestingly, from this data-driven approach we see that our brains make sense of little bits of incoming information and then fill in the rest based on our interpretation of the data (which is where higher-order cognitive, perceptual, and experiential processes come into play. Gestalt's top-down bottom-up theory is often illustrated by an abstract black and white painting of a dog (which is essentially a smattering of indistinct ink blots). At first glance, we see a dog running through the woods, however, when we take a closer look, huge chunks of the picture are missing and our brains merely fill in the missing data points because we initially received the portions of the picture that featured salient dog data. I am aware that this does not get down to the basic neuronal processes of ionic weight and action potentials; however, Gestalt's theory is a great place to start when examining perception from a data-driven framework.

Answer 2

To understand how we recognize a dog, we need to understand how neural networks work -- because image recognition is what neural networks do. So let's see how they do it.

Imagine you want to teach (to train) a neural network to tell whether there is a dog in a given picture, captcha style. Here is how you might go about it:

1. You show the network a picture, and let it guess if there is a dog in it.
2. Then you give it the correct answer, so it can learn from this experience.
3. Repeat the same the above steps a few million times.

At first, the network has no idea how a dog looks, so it will be guessing at random. But as it learns, it will start forming an singular idea of how-a-dog-looks-on-a-picture. Internally, this idea -- this statistical model -- is a collection of visual patterns that positively correlate with the presence of a dog. With each iteration, the neural network would refine this model, adding new patterns, or adjusting the weights of the existing ones. A few million pictures later, it would form a pretty good idea of how dog looks, and it would know it too (that is it also tracks the level of confidence in its ability to know the dog when it sees one).

And this is exactly what the neural networks in your subconscious mind do -- the learn ideas¹ from experience, and they apply those ideas to guess what you are looking at (and how you should feel about it). The whole process -- the learning and the application of ideas -- happens automatically entirely under the radar of your conscious awareness. Only the bottom line is communicated up to your conscious mind -- and that's why it feels like a mystery or magic. You don't know how you know that you are looking at a dog, you just do.

¹ English philosopher John Locke would refer to those as "simple" ideas -- as opposed to the "complex" ideas of our conscious mind.