The way in which the brain solves classification tasks - a reactionary process that doesn't draw on deliberate reasoning ability - seems similar to how we recreate such abilities in ANN's. So in that sense it seems our AI systems are following the path laid out for us by our own brains.
However it seems to me (and I may be wrong) that the DNC tries to implement procedural, deliberate thinking in a way divergent from the method implemented by the human brain.
The brain is clearly capable of providing us with the ability to navigate through complex procedure (like interpreting subway maps), something beyond the abilities of modern AI which mainly excels at more reactionary classification. So the question becomes, why is Deep Mind diverging from the way the brain works in the case of the DNC?
Is it because we don't actually understand how the brain implements reasoning, or is it because in order to do so in the brain's own way it would exceed our technical capabilities? Or both?
> The way in which the brain solves classification tasks - a reactionary process that doesn't draw on deliberate reasoning ability
Does this mean that when I see an animal my brain reactivity just knows "cat" instead of going "well it a small thing with fur and whiskers, fits cat"?
> Does this mean that when I see an animal my brain reactivity just knows "cat" instead of going "well it a small thing with fur and whiskers, fits cat"?
Yeah – I mean, you would notice it if the process where conscious, right?
I'm mostly skeptical of brain/AI analogies. But there are a quite a few examples where the experiences and mechanisms are eerily similar. Deep dream was probably the first I saw–the similarity to hallucinations such as those induced by LSD is just too overwhelming to be dismissed.
Deep Dream is actually a good demonstration of pattern recognition, and as such it shows how similar the brain operates in this domain. Face recognition is the most advanced of these systems, which is why we have the tendency to see faces everywhere (clouds, :) etc. There are some nice visualisations of the intermediate layers of object recognition NNs, and they seem to operate similar to the brain–starting with edge detection, having certain nodes specialising on very specific tasks (parallel lines, two small circles on a horizontal axis etc), then aggregating these.
That's vision, and probably music: layers of patterns. I'd speculate that consciousness and language operate with a somewhat different paradigm that we haven't cracked yet. NNs have made some impressive progress with text and speech, but I don't see the same sort of analogy to anything the brain does. But as far as I know we actually know much less about how these systems work in the brain, anyway.
I'd say that the classification of a cat as a cat is a fundamentally different process from, say, solving a complex math problem.
The main difference seems to be that the former is a subconscious process and the latter is conscious, but it's more than that because to solve the complex problem we're relying on explicit selective recall from our long term memory then transferring that knowledge to our short term memory for conscious processing. I don't think this transfer of information from long term to short term memory occurs when looking at a cat.
The DNC has a similar structure as the neural turing machine. It has a controller network and a memory matrix with read and write heads. However, the DNC differs in it’s memory accessing mechanisms.
Firstly, the DNC has a memory allocation mechanism ensuring that the network can write to memory that is not being used. Secondly, it also contains free gates that allow the network to explicitly forget data stored in it’s memory. Thirdly, the DNC employs a temporal link matrix which tracks the order in which writes are made. The temporal link matrix allows the location based addressing mechanism to recover from a jump caused by a content based memory access.
IIRC, the difference is an additional read head. Also they removed location-based addressing for the read/write heads in favor of a combination of content-based adressing and a dynamic scheme to figure out empty locations.
A Neural Turing machine is a type of neural network with addressable memory - it learns rather than being explicity programmed.
The net also learns to boot, i.e. how to read/ write it's own memory from scratch.
Stochastic Gradient Descent learning relies on backpropagation of the error corrections to the individual weights to make a slight improvement - this requires a differentiable neural net.
This version innovates with content addressable memory. The authors demonstrate learning from big data sets and with reinforcement learning, (i.e. trial and error as an embodied agent).
I'll attempt to explain what I understand from the paper. This will be higher level, and abstracted because that's the best I can do as an undergraduate currently.
The DNC essentially gives the neural network external memory. This memory is controlled by read heads, and write heads. These heads are actually just part of the vector output of the neural network. Each of them has a key vector, which is used to find where in memory to read or write. This is done by finding the most similar vector in memory, using cosine distance. The neural network can then use the memory. It also has mechanisms to remember the order that the memory vectors were written in, to allow for sequential recall.
The network itself is trained by reinforcement learning generally. I'm not entirely sure what is done with the parts of the vectors that are not used at a given timestep, when only input and not output is needed.
Again, this is what I believe I was able to understand from the paper. Please let me know if I got anything wrong.
I can't wait to start playing with this. The DNC is a big part of the initial inspiration for how I started modeling and reprogramming my brain!
A key difference between one of my models for how our brain works and the DNC: there's a mechanism for loading/unloading neural networks into the DNC. Also, the inputs = signals from the brain + body + mind + environment (physical & mental) = the outputs (though some outputs require iteratively applying the inputs a good bit before manifesting).
I've developed a systematic framework for generating practices to change how my body, mind, and brain work. I haven't figured out a way to explain it in simple terms, but working toward it.
EDIT: Here are some basic thoughts...
Everything we experience flows through our perceptual sets (composites of relevant emotions, beliefs, behavioral strategies, and intentions), which we can intentionally craft over time. Each component in a perceptual set is processed with varying degrees of attention.
So to program your brain, the first step is to find a functional base line perceptual set to work from & learn how to direct/focus your attention on the different aspects of your perception.
To put it even more simply: go practice all the forms of Buddhism. All I seem to be doing at this point is systematically rediscovering stuff they've known for 2500 years and attempting to put a categorical structure onto it.
Sounds like you may have stumbled into the same ideas behind CBT more or less. It makes sense that they'd crop up in different places and times assuming they work well.
I'm certainly no expert on it, but if that's a decent assessment then you may be sitting on some very handy metaphors.
Correct! Some of my practices are specifically known things in CBT. Some of them are specifically known in physical therapy. And some are more related to mindfulness or tantra.
I find them incredibly handy & have been rapidly growing in weird ways since New Year's Eve. I realized I can leverage natural cognitive biases for pattern matching to start classifying each moment of my life as meeting some set of fundamental human needs of mine.
The skill of choosing one's beliefs (ie. shortening the time it takes for a belief to be internalized) is extremely useful and something we don't really teach in society. It's no wonder we're experiencing such divisive times, especially when we've developed all sorts of systems that fail to preserve human connection.
> The skill of choosing one's beliefs (ie. shortening the time it takes for a belief to be internalized) is extremely useful and something we don't really teach in society.
I don't follow... how is this extremely useful? What is an 'internalized' belief, as opposed to what? ...something you just want to believe but don't quite yet?
I think there's two skills here: the skill of choosing beliefs and the skill of internalizing them.
The usefulness of choosing beliefs stems from several applications:
1. Choosing to believe things in order to intentionally set cognitive biases. (eg. believing every moment meets some needs of mine combines with an understanding of fundamental human needs to prime my brain to spot how my needs are being met in the moment...particularly helpful for painful moments, which can meet my need to learn how to better meet my needs)
2. Choosing to believe the opposite of problematic thoughts. (discovered this one by choosing to stop believing an anxious thought that was plaguing me)
3. Choosing to believe I can safely abandon my judgments. (which accidentally changed my food, music, media, and people preferences to be WAY more inclusive)
4. Choosing to believe I can learn to identify as a woman in the future. (accidentally creating a second conscious voice in my head, ie. stumbling on the Buddhist practice of deity visualization)
Internalizing a belief is when the thought processes change based on the belief. Some beliefs are easier to internalize than others. I've got no idea what goes into determining the time required & am also uncertain if it truly is a skill we can learn.
I've heard such internalized beliefs described as "aliefs", indicating that they are, in a way, more fundamental. As an example, if you're walking in a dark forest at night, you might rationally believe that noone is hiding there waiting to murder you, but if you don't alieve that the forest is safe, that doesn't stop you from feeling unsettled.
If very intentionally programming our brains becomes a social norm in the future, which I believe it will, then I've got no problem with this characterization.
However it seems to me (and I may be wrong) that the DNC tries to implement procedural, deliberate thinking in a way divergent from the method implemented by the human brain.
The brain is clearly capable of providing us with the ability to navigate through complex procedure (like interpreting subway maps), something beyond the abilities of modern AI which mainly excels at more reactionary classification. So the question becomes, why is Deep Mind diverging from the way the brain works in the case of the DNC?
Is it because we don't actually understand how the brain implements reasoning, or is it because in order to do so in the brain's own way it would exceed our technical capabilities? Or both?