For context: https://dolphin-nesy.github.io/
I understand that there are differences between the two systems, but I’m not exactly sure what the differences are.
I’d love to know if anyone has done any experiments comparing the two systems or if the general consensus is that Monty & Dolphin address completely different problem domains.
Monty is neither connectionist, nor symbolic, nor neurosymbolic. In its current state, it’s a machine vision / object recognition software, in a class of its own.
In my very blunt opinion, Dolphin is yet another backprop slop framework. Apples to oranges!
Although, on second read, it’s actually kinda interesting. They introduce layers with reprogrammable weights that can be controlled externally by Python to do handcrafted symbolic reasoning during training. It’s not a novel technique, but doing the symbolic stuff via Python definitely is.
LLMs have agentic tool use where they run generated Python code during chain of thought to handle things like that, well Dolphin just skips code generation and injects some computing capability directly at the layer level.
The similarity I see with Monty here, is that in the heterarchy scafolding, you’ll be able to have multiple learning modules chained together, and you could technically introduce custom ones somewhere along the chain to do some symbolic computing; a “cortico-symbolic” heterarchy if you will. In the future when Monty is more advanced, this line of thinking could natively allow it to do some types of operations more efficiently at the silicon level, instead of at the cortical level.
Imagine if your brain had a Neuralink chip, and you could directly communicate with its CPU using its native instruction set to do some number crunching, and the result would directly pop up in your mind. Dolphin kinda does that via custom operations, implemented in Python rather than a low-level instruction set, but the principle shares the same spirit.