Chapter 09

Why this matters

What you have just walked through is, at heart, one capability. In a sleeping rat, a closed-loop system can identify which memory is being rehearsed in a given hundred-millisecond window and selectively interrupt only the rehearsals of that memory while leaving everything else alone. From electrode pickup to laser pulse, the system completes a full cycle in roughly one millisecond. That capability did not exist a few years ago. It exists now.

Animation · one millisecond, end to end

A single replay event, processed through every stage of the closed-loop system. Tetrodes pick up the spike (Ch. 2), the HSE detector flags it within ~10 ms (Ch. 5), the Bayesian decoder classifies its room within roughly one more millisecond (Ch. 6), and the laser TTL fires onto the ArchT-expressing population (Ch. 7). Median spike-to-TTL latency on the actual hardware: 1.04 ms.

Three audiences should care about this for slightly different reasons.

For neurotech

Closed-loop content-specific neural manipulation is the design pattern that next-generation brain-machine interfaces will be built around. The decoder, the latency budget, the marked-point-process trick, the dynamic-threshold detector — each piece transfers. A BMI that reads intent in real time and writes feedback at the single-event scale is the same architectural problem as the one solved here, with rats swapped for humans and optogenetics swapped for whatever the actuator turns out to be (electrical micro-stimulation, focused ultrasound, magnetothermal). The feasibility question for that whole class of system is partly an existence proof. This is one such existence proof.

For memory therapeutics

A long line of pharmacology has tried to perturb memory consolidation with drugs — benzodiazepines, propranolol, NMDA antagonists, sleep-stage modulators. The intervention is almost always systemic and almost always blunt. Closed-loop content-specific disruption is a precision tool that targets one neural pattern during one phase of consolidation, leaving other patterns and other phases unaffected. PTSD reconsolidation, age-related memory decline, addiction-related cue memory: each of these is a candidate domain where pattern-specific intervention during natural sleep would be qualitatively different from the pharmacological state of the art.

For pharma and biotech research

The same infrastructure that triggers a laser can trigger any other intervention — a drug delivery pump, a cooling element, a transcranial pulse, a feedback signal to another system. Real-time content-specific neural decoding gives drug development programs a way to correlate pharmacokinetics with the actual neural patterns they are trying to influence. For nootropic and consolidation-focused programs, that closes a measurement gap that has held the field back for decades.

The system, as software

The closed-loop pipeline that ran all of this lives, open-source, on GitHub. It is a custom C++ codebase running on a real-time Linux kernel, written to be deployable on commodity research hardware (a workstation, an Axona acquisition rig, optic fibres, a 532-nm laser). It has been used in published work (this paper) and is structured for re-use in other labs that need closed-loop spike-level intervention.

Open source
lfp_online github.com/igridchyn/lfp_online

About this work

The PhD this site walks through was completed in the Csicsvari lab at IST Austria. The paper is Gridchyn et al., Neuron, 2020 — “Assembly-specific disruption of hippocampal replay leads to selective memory deficit.” Since the paper came out I have moved further into AI and applied machine learning, with a focus on biotech and pharma; ongoing writing on that lives at the AI in Pharma & Biotech newsletter.

For collaboration, hiring, or technical questions about the closed-loop system, the cleanest path is the GitHub repo or the contact form on the main site.

Eternal Sunshine of the Spotless Mind imagined memory as something you could erase. The work here suggests something less dramatic and more useful: memory access is something you can tune, in real time, with sub-millisecond precision — and that is the more interesting capability.

Thank you for reading.

§