@jhawkins introduces an idea that learning can occur across multiple cortical columns even when individual columns have only experienced a subset of the sensory inputs, or none at all, and the discussion explores a speculative role for the hippocampus in enabling this through repetition and episodic replay.
Summary Video
Main Video
00:00 Introduction 00:33 Multiple Columns Working Together in the Cortex 01:43 Things We Don’t Understand Yet 14:26 Question: What Does the Circuitry Look Like for the Hippocampus? 31:46 The Hippocampus is Always Learning 32:23 The Cortex Always Tries to Learn Compositional Models of the World 34:10 The Hippocampus is the Staging Area to Extend Models 35:56 The Hippocampus Learns Fast 41:25 Attention is Important 01:02:12 Episodic Memory Replay During Sleep 01:15:11 Behavioral Models 01:16:11 Further Discussion 01:21:04 Sparse Models and Data Transfer 01:37:04 Learning is Restricted to One Column at a Time 02:02:42 Does Replay Contain Movement Data?
Open issue
Jeff raised several open issues, including this: how can hippocampal replay help train cortical columns broadly, when anatomically the hippocampus does not connect directly to all cortical areas?
A proposal
A speculative possibility involves the thalamus. The hippocampus projects (directly or indirectly) to specific thalamic nuclei (e.g. midline nuclei and anterior nuclei via the Papez pathway). These nuclei participate in extensive cortico-thalamo-cortical loops. One possibility is that hippocampal replay engages thalamic circuits that broadcast gating and synchronization signals, rather than detailed content, thereby enabling widespread cortical plasticity during replay.
I expect that even if this proposal is correct, it can’t be the full story. But what does everyone think, is the proposed HC➡️Thal➡️cortex a feasible part of the answer?
Jeff said it’s unclear who specifies what is attended to, ie. who selects a region of space (or conceptual space)
A possible answer
A speculative possibility involves the amygdala, not in a fear-centric sense, but as a salience detector. The amygdala is known to respond to novelty, surprise, and both to big and small “Aha!” moments—all situations where attention is involved.
It has strong connections to the hippocampus, and also converges with hippocampal inputs via entorhinal and parahippocampal regions.
One possible role, then, is that the amygdala biases which cortical signals are admitted into hippocampal processing, effectively helping define the region of interest when something unexpected or salient occurs.
If my speculation ie feasible (What do you think?), the amygdala would be a contributor to attention initiation, but it’s not likely to be the sole controller.
It seems like the complete stored model of the world is accessible to every column when they need to use it.
—
I worked in a nursing home and there was a sailor who was on a battleship that got hit in the the South Pacific in WWII, who couldn’t form new memories. He thought he was in a navy hospital in the 40s, even though it was the 80s. They had to ration his cigarettes or he would keep lighting a new one until the whole pack was in the ashtray. He was very difficult to deal with because as a sailor, he needed orders, and the other aides treated him like an old man. Taking him as he was, I started barking orders and we got along great, even though he didn’t know who I was because I behaved like he expected.
Because brains are built on the cheap, they use the same part of the brain to imagine things that they do to sense and process input. It can play back episodic memories of experience using the structure that recorded them. The brain tags that activity with ‘I did that’ and if that fails we perceive the sounds and images as coming from the outside.
We need to sleep for biological reasons, and the major one is cleaning house and taking out the garbage. Whatever else is going neurologically can be dealt with in Monty as it goes, because it doesn’t need to wait for the opportunity to take care of it during housekeeping.
Open issue
