a short description of the research topics that I'm interested in/working on
"most arguments are just two neural networks trained on different datasets yelling about their programming."
Some topics to start the conversation!
- VLMs as models for human visual processing
- Higher-dimensional representations of how continuous perceptual inputs are transformed into discrete semantic states (idea from the paper Visual Language Hypothesis!), and this as an explanation for multimodal processing in humans and models
- RAGs, parametric memory, engrams and memory as a reconstruction mechanism: to what extent do we understand human memory storage and selectivity? Is it even applicable under engineering principles? What's the future of AI memory beyond RAG? What about multimodal memory?
- Selective attention and saliency as an efficient way for designing small, local models: human-like attention architecture in vision for faster computing, memory storage and smaller model size
- Neuromorphic engineering and hardware designs for efficient computing: can we feed models with rice, curry and hotdogs level of energy? Model size VS brain size?
- Do we even have a circuitry level understanding of emotion pathways in the human brain? Does AI already possess a degree of theory of mind understanding? What ontologically is theory of mind, and how to quantify/determine (is it quantifiable?) an explanation?
- Organoid intelligence and utilizing biological computing systems: is it possible/plausible/computationally efficient/ethical? (works by Cortical Labs!)
- Consciousness? How do we even approach consciousness?
- Are bioelectronic systems and electrical synapses the ultimate form of BCIs and HCI?
- What about music? Models can't even read scores/interpret audio data/generate any form of output-even extremely simple ones-without noticeably hallucinating. Is it simply because of not having enough data? what's the effectiveness of scaling law in multimodal models?
And so on and so on.