Exactly! And by formulating the firing pattern of neurons as a language we whittle down the search space for finding optimal solutions.
1 Like
To create or recognise music is uniquely human, but the underlying basis is an evolved sensory capability: the ability to hear musical notes and recognise pitch. All sound has underlying physics and thus maths, but musical notes share this quality of pitch (akin to frequency), and audible relationships between notes, defined by mathematical ratios.
In summary: most sound does not have pitch. A few, such as birdsong, do and our ears have evolved to recognise that. Music is a language built on that sensory ability. Musicians are simply those with skill in that language.
Maths is not really a language in the same sense, but thatâs a topic for another thread.
1 Like
The ability to explicate concepts is what gives us an advantage when maximising our expected reward.
If we could build an explicit evaluation system into an agents policy it would have this advantage too.
We know that activations constitute the interpolation of a manifold. This seems surprisingly like hyroglyphic symbols being interpolated where the symbols are features.
current systems act more like document summarisers rather than chatbots in terms of the statements made by their firing patterns.
what we want is to speak to this system using its own terms (learnt feuatres) in combinations that are instructive, or teach it to do so to itself.
One way that that could happen would culminate in the following loop.
in a chatbot one could do this
2 Likes
I am working on the idea that all that needs to exist at anyone time are the activated neurons, and their context.
When a chord is played on a piano, it is possible to tell that a major or minor chord was played, without needing to hear it in relation to the notes you are not playing. the information is somehow self contained.
If we could have some way of saving neurons in a format where if it fires we already know its context, and have a rule for moving from one neuron to the next that only depends on the firing neurons we would have something. Like simply by mapping frequencies to each other without the need of a database of node frequency pairs.Music theory defines connectivity in terms of optimal chord sequences that only need local information on the notes being played. and can represent this numerically.
Think of a musical keyboard like a non local cellular automaton, Where the notes playing in one chord are connected to the next with individual notes leading (stronger connection) in one chord to the other more or less strongly.
The brain is a non local cellular automaton.
I bleed this music theory topic because i think im onto something 
This makes no sense. The brain has no idea whatever about frequencies. The brain receives a sensory input via the cochlear nerve in the same way it receives sensory input from skin, eyes, etc. No magic, just sensory input.
1 Like
Another system is similar to a GAN. A spiking neural network will be trained to induce music theory constrained activations upon being presented videos of a human performing a task. It will be trained to have high consonance in its activations then.
When shown a robot attempting to perform the same task, the spiking network will be rewarded for maximising dissonance instead.
That means that the human performing the task well induces harmonic high consonance activations within the spiking network when frames of such a demonstration are shown to it⌠While when the robot , which will initially perform poorly ,acts in the frames , the spiking networkâs activations are rewarded for not playing in harmony.
The next phase would be to have a phase where we donât train the spiking networking being shown frames from a demonstration, but train the robot instead ,with a reward calculated from the activations of the spiking neural networks level of harmony.
The robot will maximise consonance in the spiking networks activations only by acting more and more like the human and acting less and less like a robot, since the spiking network only produced consonance under the humans demonstrations and not the untrained robots.
This loop can be repeated till convergence.
I am working on the math to force humanoid robots in a simulation to behave similar to humans in a video for a movie dataset or some series.
The robots will be based on a spiking network or a VQVAE model and we need a way to generate a better learning signal from the network being shown demonstrations than just reward. There is alot of information related to the particular chords and notes firing that could make the robots learn faster.