Biologically-inspired AI
I’ve been making progress with Million Minds AI. I added a fourth level to the hierarchy, expecting another long debugging cycle. Instead, it just worked. The system uses the same brain-inspired column machinery at every level. The first level learns paths through character evidence. The next levels observe completed objects from below and learn larger [ ]
What does python” mean? It depends on context of course.. I showed Million Minds AI a small collection of writing about snakes and reptiles. Then I showed it another collection about code and software. I never explained that python can mean two different things. There were no animal or programming labels during learning. When I [ ]
Your brain can recognize a song from its opening bars. It does not conclude that the song is over. Million Minds AI was making the equivalent mistake. The hierarchy could recognize a familiar pattern, but recognition and completion had become entangled. When one layer recognized something, it could prematurely tell the next layer that the [ ]
I have been building a small step-through demo so I can watch the system read text: the sensor moves, Layer 1 settles, Layer 2 reacts, and I can replay what happened. That demo exposed a timing issue in the Layer 1 -> Layer 2 handoff 😱 Layer 1 needs time to settle on what it [ ]
Ambiguity is not always failure. That was the main lesson from the latest round of work. The recent focus was Layer 1 and early Layer 2 behavior over text. Layer 1 now has a better substrate for text object representations. One concrete mechanism we added is a text-form SDR generator. Previously, learned text objects could [ ]
Layer 1 is starting to work like an object-recognition layer over text. The system scans text through small sensor patches, then cortical-column-like modules build sparse maps of what they observe: feature@location - movement - feature@location - movement - Recent progress: Pose stays monotonic instead of resetting at boundaries. Scanning now has saccade/fixation timing, so [ ]
Recent work has been focused on a deceptively simple question: When many independent columns observe related evidence, how should they agree? Columns do not all share one global representation. Each column builds its own private model from its own vantage point. Agreement should emerge through learned correlations and lateral support, not because every column is [ ]
I ve been making progress.. TextSensor is built and acts like a simple eye over text. It has multiple patches, moves with a saccade/fixation rhythm, and emits sparse feature SDRs from local character windows. Layer-1 columns now receive those SDRs, maintain monotonic pose, learn temporal transitions with a graph-based memory, and settle into private object SDRs. [ ]
When this system is up and running, because it can learn continuously (unlike current AI), it will be just like Johnny 5 from Short Circuit movie. More input!
Finally we have a sensor, a thalamus, a column and a voting module. It is all neuroscience legit although we did keep some things simple at the moment. The sensor is retina-like, with several patches, partially overlapping. Each patch takes in raw input (character based text), looks for features and generates an SDR. Thalamus is [ ]