What would a model learn if it were trained on dreams instead of the internet?
Status: concept only. This repository is a single README describing a research idea. There is no code, dataset or model, and capturing dreams at the fidelity described below is beyond current neuroimaging.
Every night people dream: emotional, associative, illogical and often creative. DreamML proposes treating that as training data. Record brain activity during REM sleep, reconstruct the dream content, pool the reconstructions into a "dream lake", and train generative models on it. The bet is that a model trained on unfiltered, subconscious thought could produce different kinds of creative output and insight than one trained on polished text.
Partly, at a very coarse level. Researchers have used fMRI to decode broad categories of visual content from sleeping subjects (for example Horikawa et al., Science, 2013), and brain-to-image reconstruction from waking fMRI is an active research area. Reconstructing full dream narratives, emotions and abstract thoughts is well beyond that, and it would also need large-scale sleep-lab data collected with consent. DreamML sits on the far side of that gap.
- Dream capture. Record neural activity during REM sleep with fMRI, EEG or future non-invasive sensors.
- Dream reconstruction. Train decoders that map neural activity to images, emotions and abstract content.
- Dream data lake. Store reconstructions as a growing dataset of subconscious experience.
- DreamML model. Train generative models on the dream lake to produce ideas, art and analysis.
- Creative work: music, art and writing that follow dream logic
- Mental-health research: studying recurring dream patterns (as research, not diagnosis)
- Problem solving: using dream-like association as a source of unusual ideas
- Emotion research: analyzing the emotional tone of dreams at scale
- How much dream content can actually be decoded, and how would you validate a reconstruction against a subject's report?
- What consent and privacy model is acceptable for data this personal?
- Would a model trained on dreams be more creative, or just noisier?
- Neuroscientists: what is decodable today, and with what equipment?
- ML engineers: architectures for learning from noisy, sparse neural sequences.
- Hardware engineers: higher-resolution, sleep-friendly brain sensing.
- Psychologists: how dream content relates to emotional state.
- Everyone else: share the idea, argue with it, or build on it.
Open an issue or pull request to discuss. If you take this further, the author would enjoy hearing about it, and would not object to a firmware version named after him.