Volitional activation of remote place representations with a hippocampal brain-machine interface

Abstract

The hippocampus is critical for recollecting and imagining experiences. This is believed to involve voluntarily drawing from hippocampal memory representations of people, events, and places, including maplike representations of familiar environments. However, whether representations in such "cognitive maps" can be volitionally accessed is unknown. We developed a brain-machine interface to test whether rats can do so by controlling their hippocampal activity in a flexible, goal-directed, and model-based manner. We found that rats can efficiently navigate or direct objects to arbitrary goal locations within a virtual reality arena solely by activating and sustaining appropriate hippocampal representations of remote places. This provides insight into the mechanisms underlying episodic memory recall, mental simulation and planning, and imagination and opens up possibilities for high-level neural prosthetics that use hippocampal representations.

Fig. 1.. Hippocampal map-based brain-machine interface (BMI) in a virtual reality (VR) system.

(A) Steps for performing the two different BMI experiments in this study. Rats first physically ran to a series of goals (“Running task”) while their hippocampal neural activity and (virtual) location in a square arena were recorded. This data was used to train a decoder to take neural activity as input and output the animal’s current location in the Running task. In BMI task 1 (“Jumper”), animals needed to generate neural activity that would be decoded as locations they wanted to move to so that they could reach each goal (to obtain reward). In BMI task 2 (“Jedi”), animals were fixed at the center of the virtual arena (but could rotate) and needed to generate activity corresponding to locations where they wanted an external object to move to so that the object reached the goal, then they needed to sustain that activity to maintain the object there (to maximize reward). (B) Schematic of VR system (left). Animal was free to rotate its body in the horizontal plane. In the Running task, animal’s location in the virtual arena environment was updated based on treadmill movement. Simultaneously recorded spiking from a population of hippocampal CA1 units expressed place fields—the basis of the cognitive map of the environment (right). Decoder was then trained using binned spiking activity and location data. (C) In both BMI tasks, treadmill no longer updated VR. Instead, the animal or object location was controlled solely by real-time hippocampal activity. A neural signal processor rapidly assigned activity to individual units, whose spike counts were fed into the decoder. VR projection was updated based on locations output by the decoder. In the “Jumper” (“Jedi”) task, the animal’s (object’s) virtual location was moved toward the most recent decoded locations.

Fig. 2.. Rats can navigate to goals by controlling their hippocampal activity.

In both Running and Jumper BMI tasks, animals were rewarded when they reached each goal. (A) Animal trajectories in virtual arena for consecutive Running task trials. Trial duration (time to reach goal) in seconds shown. (B) Example Running task trial. From top: trajectory, firing rate (z-scored) of individual units (units were ordered by time of peak activity), treadmill speed, and LFP from one recording channel and corresponding wavelet spectrogram during trial. (C) Accuracy of trained decoder of animal’s current location for held-out Running task data. Actual and decoded trajectories during example trial (top left) and across several trials (for X and Y coordinates separately, bottom left). Median decoding error (distance between actual and decoded locations) with range and quartiles (bottom right). (D) Example Jumper BMI trial with similar trajectory as Running trial in (B). From top: trajectory generated by the animal controlling its hippocampal activity and the decoder output (animal is teleported toward decoded location; each gray circle represents the decoded location at the time the animal is at the corresponding point in the trajectory connected by the dark line, sampled here every 1 s), firing rate of individual units (using same order of units as in (B)), treadmill speed, LFP, and spectrogram. (E) Example Jumper BMI trial in which animal did not move the treadmill. Trajectory as in (D) (left). Right, from top: unit activity, treadmill speed, LFP, and spectrogram. See fig. S10 for all 10 non-movement trials. (F) BMI-generated trajectories for consecutive Jumper trials. (G) Mean Jumper trial duration (vertical line) is significantly lower than distribution of expected mean duration for simulated trials if goals were in random locations. (H) Polar distribution of angle between direction of movement and direction to goal during Running and Jumper tasks. Zero corresponds to animal movement directly toward the goal center.

Fig. 3.. Rats can move objects to remote goal locations and maintain them there by controlling their hippocampal activity.

In the Jedi BMI task, trials did not end when the external, controlled object first reached the goal; instead, animals were rewarded as long as the object was in the goal region (white circle), for up to 3 min per trial. The animal was always fixed at center of virtual arena, but could rotate its body and generally turned toward each goal. (A) Distribution of real-time decoded locations (output every 100 ms) generated by the animal controlling its hippocampal activity across 8 consecutive Jedi BMI trials for rats 1–3. Panels show decoded locations during each trial (up to 3 min, fig. S11). Periods when animal’s body rotated >12°/s were excluded. See text and methods for details. The external, controlled object (which was visible for rats 1–2, invisible for rat 3) was moved toward the decoded location (fig. S11 shows that the distribution of object locations was essentially the same as the distribution of decoded locations). (B) A 40-second-long period during an example trial during which animal did not move the treadmill. From top: Summed activity across all units with population burst events (PBEs) identified, treadmill speed, distance of decoded location from goal (0 means inside goal region), and close-ups of two 5-second periods (left: as animal moves object to goal; right: as animal maintains object at goal; points in arena represent sequence of decoded locations) with spike trains of units, LFP, and spectrogram. See fig. S12 for additional example periods. (C) Mean distance of decoded location from goal across all trials (vertical line) is significantly lower than mean distance expected for randomized goal locations. (D) Treadmill speed distribution during periods shown in (A) showing animal was generally still during task performance.

Fig. 4.. Volitionally generated non-local activity is similar to the activity when the animal is at the corresponding locations and is associated with theta-band power in the LFP.

(A)-(E) The population vector (PV) of ongoing spiking activity was compared to the average place field activity (rPV) at a given location during the Running task. (A) Schematic of comparison. (B) Mean correlation of instantaneous (500 ms window) PV during Running or Jumper task with rPV for the current location (in Running task), current decoded location (in Jumper task), or random location in Running (randRun) or Jumper (randJumper) task. (C) Same as (B) but for Jedi task. For Jedi, only periods when decoded location was near (within 5 cm of) goal were included (also for (E)). (D)-(E) Correlation of PV with rPV relative to baseline random value as a function of time integration window for determining the PV. (F)-(G) Evaluation of decoder performance when ground truth activity for each location, i.e., the rPV, was input into decoder. (F) Schematic of evaluation procedure. (G) Comparison of our DNN decoder to Bayesian decoder for different levels of added noise, with example traces using a specific level of noise (top). (H) Distribution of decoded location (left) during Jedi task segment with no treadmill movement (right). Right, from top: Summed activity across all units with population burst events (PBEs) identified, treadmill speed, distance of decoded location (excluding data during PBEs) from goal, and close-up LFP with spectrogram. (I) Power spectral density of z-scored (for pooling across animals) LFP during Jedi task for periods of treadmill movement and all long segments (≥8 s) without treadmill movement. See text and methods for details. Here and elsewhere all CIs are 95% CIs.