Review of:
Associational memory, as its name implies, is a type of memory that allows one to fuse multiple events in memory. If your boss constantly yells at you in his office, you might begin to form some bad memories of being in that office. While the phenomenon of associative memory is a familiar experience, the neural basis for it isn’t well understood. A prominent theory, which was formed in the mid-20th century but only tested recently is that neurons encode associations by wiring together. In the boss’s office example, the sight of the boss’s office might activate one subset of neurons, and those neurons would then “fill in” activation of neurons that code for fear or memories of yelling (obviously this is a gross oversimplification - I’m only using it to demonstrate the principle).
Pattern completion has been observed in the hippocampus but not in primary sensory cortices, which were originally thought of as passive transformers of sensory stimuli rather than modules capable of learning in adulthood. It is now clear that this isn’t entirely true - adult learning happens in primary sensory cortex following changes to statistics of stimuli (e.g. adult plasticity).
The central question of the paper by Xu et al was to test if a neural ensemble in primary visual cortex (V1) could in principle recapitulate its response to a stimulus when only a part of that stimulus was actually shown. In this paper, the stimulus was sequential, consisting of a small dot of light moving across the receptive fields of neurons that the researchers were recording from.
In the experimental setup, rats were restrained (either while awake or anesthetized) and made to watch a small dot move across their retinas 100 times. Meanwhile, electrodes placed in V1 were recording the activity of neurons arranged linearly across V1. Crucially, the location of the moving dot on the screen was chosen so that it activated the recorded neurons sequentially. To test whether the ensemble of neurons in V1 “learned” anything from the 100 trials of the moving dot, the researchers compared the activity of the ensemble in response to stimulation by a stationary dot (i.e. the test dot was flashed in the same location as where the moving dot would start during learning, but the test dot was stationary) before and after the 100 learning trials.
If the recorded ensemble had in some way learned the sequential stimulation by the moving dot, then it would be expected that flashing the dot at the starting point of the sequence would elicit spontaneous “filling in” of the entire sequence - the rest of the cells in the ensemble would fire one after another as if the dot were moving across their receptive fields, as during the training. To examine this statistically, the authors compared the correlations among the firing times of all pairs of recorded cells, and plotted those correlations as a function of time between firing and distance between the cells’ receptive fields. Plotting this for the training condition, in which the dot was actually moving across the receptive fields and activating all cells sequentially, shows high correlations stretched out as a line, or sequence: this represents the ground truth (or control condition) because the moving dot is expected to activate the cells sequentially. The punchline is that the stationary dot flashed after 100-trial training elicited a similar sequential activation (as evidenced by a line of high correlations among cell pairs), whereas the stationary dot flashed before the training did not.
Some caveats:
1. The average correlation among firing patterns of pairs of cells in response to the moving dot was ~0.9 - really high, as expected. The average correlation among pairs in response to the stationary dot prior to training was just ~0.2 - very low, also as expected. But after training, the correlation only improved to ~0.3 - disappointingly low. Statistically, the effect is real, but it would be interesting to know whether such low correlations would be enough for the animal to perceptually “hallucinate” the rest of the sequence (this is beyond the scope of this paper, which only examined changes during passive viewing).
2. The animals in this study viewed the stimuli passively and were either awake or anesthetized. While the improvements in correlations among the firing of cell pairs were similar in both conditions, the spike raster plots (which summarize the times at which the neurons fired in response to stimulation) are much more convincing for the anesthetized case. This suggests that some property of the anesthetic (perhaps massive disinhibition) enhanced the observed effect. Why would “learning” be better in anesthetized rats than awake rats?
3. While the extent to which correlations among cells improved did depend on the number of “learning” trials and the speed at which the dot moved across the screen, the learning effect all but disappeared six minutes after the 100 trials of learning.
4. The authors observed that while the rats were restrained, they would periodically enter periods of mellow inactivity (I like to think of this as a “zombie” state - when you’re not really asleep, but definitely not fully awake) characterized by low-frequency high-amplitude oscillatory neural activity. In contrast, awake states featured whisker movements and high-frequency low-amplitude activity. When they separated the learning trials by the animal’s brain state (i.e. zombie vs. awake), they found that learning the moving dot stimulus during the zombie state but not the awake state elicited improvements sequential “recall” in the recorded neural ensemble. This result in essence nullifies the earlier claim that the neural ensemble can learn the sequence while the animals are awake. The authors offer no explanation for this strange result except that they wondered whether eye movement could be the culprit behind the lack of learning during the awake case (they tracked eye movements only in this experiment and concluded that because the deviations in eye position were smaller than the size of the stimulus, eye movements could not explain the lack of learning in the awake rats).
5. Neurons in V1 respond best to bars of light (or edges) rather than the circular spots used here. The authors do not explain why they chose the stimulus to be circular and do not show any raw data of the neurons’ responses.
Despite these shortcomings, the paper demonstrates a cool effect of repeated stimulation on responses in an ensemble of neurons. The effect is clearly not naturalistic or long-lasting, but still constitutes a type of memory While it would be more interesting to see if such an effect is present when the animals are actually engaged in the task (i.e. if they get a reward for viewing the training stimulus but not others), the paper should be judged by how it answered the questions that it posed.